HK1183550B

HK1183550B - Method for e-commerce site navigation and system thereof

Info

Publication number: HK1183550B
Application number: HK13110909.4A
Authority: HK
Inventors: 曾安祥; 潘春香
Original assignee: 阿里巴巴集团控股有限公司
Filing date: 2013-09-25
Publication date: 2017-12-08

Description

Electronic commerce website navigation method and system

Technical Field

The present application relates to website navigation technologies, and in particular, to a method and a system for navigating an e-commerce website.

Background

At present, shopping modes of e-commerce websites are mainly divided into three categories: category browsing, advertising operations, and searching. The category is a classification of the commodity, and is divided into a foreground and a background, the foreground is used for UI (user interface) display, the background is used for commodity management, and the mapping relationship between the foreground and the background is described by a rule. Currently, the mainstream category system is represented by a tree structure, each parent category has a plurality of sub-categories, but each sub-category has only one parent category, so that the range of representation of the categories from top to bottom is smaller and smaller.

The category browsing mode is realized by website operation, firstly, the first-level categories are combined, then the combinations are serially displayed according to the attention degree of a user, when the user wants to buy the commodities under a certain category, the user enters the category and clicks the sub-categories layer by layer to screen the commodities. This category browsing approach requires the user to be familiar with the category hierarchy to find the desired item. The advertisement operation refers to the propaganda of the single item or the shop through the advertisement space, and the user can directly purchase or enter the shop for purchase by clicking the advertisement link. In the search mode, the user inputs keywords according to the purchase intention to inquire so as to obtain a recommended category list and a recommended commodity list, and the method does not require the user to know a category system, is more convenient to use and becomes one of mainstream purchase modes.

In such a mainstream search mode, in order to reduce the search time and the number of clicks of a user during shopping in an e-commerce website, an intelligent navigation technique has been developed.

Early electronic commerce web sites used category commodity quantity navigation. The category commodity quantity navigation means that after a user inputs a keyword, the recommended category sequencing is determined by the quantity of related commodities under the categories and displayed layer by layer. Under the navigation mode of category commodity quantity by using text matching, along with the sharp increase of commodity quantity and commodity categories, the number of the obtained categories is greatly increased when a user specifies a keyword for query, the relevance between a query word and the categories cannot be reflected by text matching, and the user cannot judge which categories should be obtained for more fine screening. For example, the first category recommended when searching for a certain model of mobile phone is "3C digital accessories market", because the number of goods under accessories category far exceeds that of mobile phone, but this category obviously does not meet the search intention of the user.

Aiming at the problems, one solution is to score the relevance of categories according to historical category clicking behaviors, the categories are dynamically displayed according to the scores, the relevance is gradually reduced from left to right, and meanwhile, categories with poor relevance are folded and hidden by the aid of categories, and the navigation mode is called category clicking navigation. However, this navigation method still does not open the frame displayed from the top-level category, and the user needs to click many times to select a finer category for filtering. Moreover, the category click navigation is obtained according to the analysis of the historical behavior data of the user, the calculation of the part adopts the top-down traversal for the category processing mode, and the display mode is also the layer-by-layer expansion, so that the method is not suitable for a complex category system. For example, for a rich, complex and deep-level category system, the progressive presentation may result in a long search path.

In summary, the relevance of the search result given by the current navigation methods and the search intention of the user is poor, and the user cannot quickly find the desired product, which not only increases the search time of the user, but also increases the access burden of the category-related server.

Disclosure of Invention

The application provides an e-commerce website navigation method and system, which are used for solving the problems that search results given by various existing navigation methods are poor in correlation with search intentions of users, and the users cannot quickly find desired commodities, so that the burden of accessing a server is reduced.

In order to solve the above problem, the present application discloses an e-commerce website navigation method, including:

performing statistical analysis on user behaviors on website log data to obtain reference data required by a navigation recommendation model corresponding to a query word, wherein the reference data comprises category click distribution data and/or category purchase distribution data and/or category correlation distribution data corresponding to the query word;

selecting a navigation recommendation model to be used by the query word according to the reference data corresponding to the query word;

inputting reference data corresponding to the query word into the selected navigation recommendation model, and calculating recommendation result data corresponding to the query word, wherein category and/or attribute data meeting conditions are used as recommendation result data corresponding to the query word according to a bottom-up recommendation mode;

generating a recommended word list containing a mapping relation between the query words and recommended result data;

and when receiving a user query word input online, querying the recommended word list, and outputting recommended result data corresponding to the user query word.

Preferably, the using the category and/or attribute data meeting the condition as recommendation result data corresponding to the query term according to a bottom-up recommendation manner includes: replacing the next-level category meeting the conditions with the previous-level category according to a bottom-up recommendation mode to obtain category and/or attribute data, and using the category and/or attribute data as recommendation result data corresponding to the query word; the next-level category meeting the conditions is the next-level category with the click ratio exceeding a preset threshold, and the click ratio is the ratio of the click quantity of the next-level category to the total click quantity of the current category.

Preferably, the category click distribution data corresponding to the query term is obtained by: and according to the website log data, counting the distribution of the categories clicked by the query word and the click quantity of each category to obtain category click distribution data corresponding to the query word.

Preferably, the method further comprises: accumulating the click rate of each category clicked by the query word to obtain the total click rate of each query word; and filtering out the query words with the total click rate less than a preset value.

Preferably, the method further comprises: and counting the click rate of each query word clicked by each user according to the website log data, and weakening the click rate of the query word clicked by the user if the click rate of a certain query word clicked by a certain user exceeds a preset value.

Preferably, the category purchase distribution data corresponding to the query term is obtained by: according to the website log data, the user identification corresponding to the query word and the clicked commodity identification are associated with the user identification and the purchased commodity identification through the user identification, and category purchase distribution data which is clicked to a certain commodity through a certain query word and generates purchase is obtained.

Preferably, the category relevance distribution data corresponding to the query term is obtained by: and counting the number of commodities related to the query word in each category clicked by the query word according to the website log data to obtain category correlation distribution data corresponding to the query word.

Preferably, the navigation recommendation model includes a parent-child category recommendation model, the inputting of the reference data corresponding to the query term into the parent-child category recommendation model, and the calculating of the recommendation result data corresponding to the query term include: summarizing reference data corresponding to the query words to a first class; respectively taking the two primary categories which are clicked most as father categories, and taking the rest primary categories as sub-categories under the same father category; and checking whether each sub-category of the parent categories can be replaced by the next level category meeting the conditions according to bottom-up recommendation, and finally taking the category and/or the attribute data meeting the conditions as recommendation result data.

Preferably, the navigation recommendation model includes a divergence recommendation model, the inputting of the reference data corresponding to the query term into the divergence recommendation model, and the calculating of the recommendation result data corresponding to the query term include: summarizing reference data corresponding to the query words to a first class; sorting the primary categories according to the click rate from high to low, and selecting a plurality of primary categories which are sorted in the front; and checking whether each selected primary category can be replaced by the next qualified category according to bottom-up recommendation, and finally taking the qualified categories and/or attribute data as recommendation result data.

Preferably, the navigation recommendation model includes a category search recommendation model, if category information is hidden in the query term, the reference data corresponding to the query term is input into the category search recommendation model, and calculating recommendation result data corresponding to the query term includes: searching corresponding categories according to category information hidden in the query vocabulary; summarizing reference data corresponding to the query words to a next-level category of the categories; sorting the next-level categories according to the click rate from high to low, and selecting a plurality of next-level categories which are sorted in the front; and checking whether each next-level category can be replaced by a qualified next-level category according to bottom-up recommendation, and finally taking the qualified categories and/or attribute data as recommendation result data.

Preferably, checking for each category according to a bottom-up recommendation whether it can be replaced by a qualified next level category comprises: determining the current category from top to bottom according to the category hierarchy; if the current category is the leaf category, returning recommendation result data as the current category; otherwise, summarizing the reference data corresponding to the query words to the next level category of the current category; selecting a next-level category with a click ratio exceeding a preset threshold, wherein the click ratio is a ratio of the click quantity of the next-level category to the total click quantity of the current category; calculating the number of the categories expected to be recommended in the next level according to the number of the categories expected to be recommended currently; if the number of the selected next-level categories is larger than the number of the categories expected to be recommended by the next level, canceling the recommendation to the next level, and returning recommendation result data as the current categories; and if the number of the selected next-level categories is less than or equal to the number of the categories expected to be recommended by the next level and the next-level categories are not leaf categories, replacing the current categories with the next-level categories, determining each category of the next-level categories as the current category, and continuously recommending the next-level categories of the current categories according to the steps.

Preferably, if the number of the selected next-level categories is smaller than the number of the categories expected to be recommended by the next level and the next-level categories are leaf categories, replacing the current categories with the next-level categories, and selecting the number of the categories with a difference according to the category purchase distribution data in the reference data for supplement, so as to reach the number of the categories expected to be recommended by the next level.

Preferably, the navigation recommendation model includes a category attribute mixed recommendation model, and if the description information of the commodity is hidden in the query term, the reference data corresponding to the query term is input into the category attribute mixed recommendation model, and the recommendation result data corresponding to the query term is calculated, including: summarizing reference data corresponding to the query words to leaf categories, wherein attributes appear in the leaf categories; sorting the attributes from high to low according to the click rate and/or the entropy difference, and selecting a plurality of attributes which are sorted in the front; for each selected attribute, sorting the attribute values according to the click rate from high to low, and selecting a plurality of attribute values which are sorted in front; and taking the selected attributes and the attribute values corresponding to the attributes as recommendation result data corresponding to the query words.

Preferably, the method further comprises: judging whether each selected attribute has a sub-attribute, and if so, replacing the attribute with the sub-attribute; and taking the replaced sub-attributes and the attribute values thereof as recommendation result data corresponding to the query words.

Preferably, the navigation recommendation model comprises a direct category recommendation model, the inputting of the reference data corresponding to the query term into the direct category recommendation model, and the calculating of the recommendation result data corresponding to the query term comprises: summarizing reference data corresponding to the query words to a first class; selecting a first class with a click ratio exceeding a preset threshold, wherein the click ratio is the ratio of the click quantity of the first class to the total click quantity; and checking whether the quantity of the selected related commodities of each primary category is greater than a preset quantity, if so, determining the primary category as a direct category, and using the direct category as recommendation result data corresponding to the query words.

The present application further provides an electronic commerce website navigation system, including:

the data analysis module is used for carrying out statistical analysis on user behaviors on the website log data to obtain reference data required by a navigation recommendation model corresponding to the query word, wherein the reference data comprises category click distribution data, category purchase distribution data and/or category correlation distribution data corresponding to the query word;

the model prediction module is used for selecting a navigation recommendation model to be used by the query term according to the reference data corresponding to the query term;

the category attribute recommendation module is used for inputting the reference data corresponding to the query word into the selected navigation recommendation model and calculating recommendation result data corresponding to the query word, wherein the category and/or attribute data meeting the conditions are used as the recommendation result data corresponding to the query word according to a bottom-up recommendation mode;

the recommendation word list generating module is used for generating a recommendation word list containing the mapping relation between the query words and the recommendation result data;

and the online query module is used for querying the recommended word list when receiving a user query word input online and outputting recommended result data corresponding to the user query word.

Compared with the prior art, the method has the following advantages:

firstly, the intelligent navigation provided by the application draws essences of category click navigation and category commodity quantity navigation, comprehensively considers historical factors such as click and purchase corresponding to key words, commodity quantity information related to query words and the like, and provides categories or attributes most related to search intentions.

The biggest difference from the category clicking navigation is that the intelligent navigation of the application adopts a bottom-up recommendation mode, recommendation result data corresponding to query words are taken as the recommendation result data corresponding to the query words according to the bottom-up recommendation mode category and/or attribute data, and if the clicking or purchasing occupation ratio of a certain category reaches a certain threshold value, the upward backtracking process is stopped, so that the mode of displaying the category from top to bottom is eliminated, and the commodity category desired by a user can be positioned more quickly.

The biggest difference with category commodity quantity navigation is that the intelligent navigation extracts the commodity quantity related to the query word under the category instead of the commodity quantity in the text matching meaning, and the commodity quantity information is only used as reference data under a specific condition (such as when processing a low-frequency query word).

Secondly, the intelligent navigation of the application introduces an attribute recommendation function and an attribute preselection function, and if the query words of the user imply the description information of the commodity, the navigation provides the attribute preselection function, so that the target commodity is accurately locked.

Of course, it is not necessary for any product to achieve all of the above-described advantages at the same time for practicing the present application.

Drawings

FIG. 1 is a flowchart illustrating a method for navigating an e-commerce website according to an embodiment of the present application;

FIG. 2 is a flow chart of an automatic deployment recommendation method according to an embodiment of the present application;

FIG. 3 is a schematic diagram of automatically deploying recommendations in an embodiment of the present application;

FIG. 4 is a schematic diagram of attribute recommendation in an embodiment of the present application;

FIG. 5 is a diagram illustrating sub-attribute recommendation in an embodiment of the present application;

FIG. 6 is a block diagram of an electronic commerce site navigation system according to an embodiment of the present application;

FIG. 7 is a diagram of an application system architecture for intelligent navigation in an e-commerce web site.

Detailed Description

In order to make the aforementioned objects, features and advantages of the present application more comprehensible, the present application is described in further detail with reference to the accompanying drawings and the detailed description.

The application provides an intelligent navigation method, which is particularly suitable for navigation of an e-commerce website, the navigation method draws essence of category click navigation and category commodity quantity navigation, comprehensively considers historical factors such as click and purchase corresponding to keywords, and simultaneously adds commodity quantity information related to query words to provide categories or attributes most related to search intentions.

The categories refer to the classification of commodities, and are divided into a foreground and a background, wherein the foreground is used for UI (user interface) display, the background is used for commodity management, and the mapping relationship between the foreground and the background is described by rules. Currently, the mainstream category system is represented by a tree structure, each parent category has a plurality of sub-categories, but each sub-category has only one parent category, so that the range of representation of the categories from top to bottom is smaller and smaller. The attributes refer to description of commodities, such as the material and price of the T-shirt, and can be used as attributes of the T-shirt, each attribute at least has 2 or more values, and for example, the attribute value of the material can be pure cotton and wool.

The following describes the implementation process of the method of the present application in detail by way of examples.

Referring to fig. 1, a flowchart of a method for navigating an e-commerce website according to an embodiment of the present application is shown.

Step 101, performing statistical analysis on user behaviors on website log data to obtain reference data required by a navigation recommendation model corresponding to a query word, wherein the reference data comprises category click distribution data and/or category purchase distribution data and/or category correlation distribution data corresponding to the query word;

the log refers to a record of a server on user query and click events, and is usually stored in a distributed file system due to huge data volume. The statistical analysis of the user behaviors refers to the statistical analysis of behaviors of clicking, purchasing and the like of the user.

How to obtain category click distribution data, category purchase distribution data and category correlation distribution data corresponding to the query term through data statistics is described in detail below:

1) the category click distribution data corresponding to the query term may be obtained by:

and according to the website log data, counting the distribution of the categories clicked by the query word and the click quantity of each category to obtain category click distribution data corresponding to the query word.

For example, for the query word "glove", the category distribution clicked by the user after inputting the word "glove" is counted according to the log, the category distribution includes the distribution of the breadth and the depth, wherein the primary category clicked by the breadth distribution such as inputting "glove" is "outdoor/mountain-climbing/camping/traveling" and "clothing accessory/belt/scarf hat", the depth distribution such as continuing to click on the secondary category, the tertiary category, and so on under the primary category, and so on, until the leaf category. And counting the click quantity of each category clicked by the user according to the log. Through the statistical analysis, the category click distribution data including the click category distribution and the click amount of each category can be finally obtained.

Optionally, in order to remove data noise and improve the accuracy of data statistics, the method may further include the following steps:

accumulating the click rate of each category clicked by the query word to obtain the total click rate of each query word; and filtering out the query words with the total click rate less than a preset value.

For example, the click rates of all the primary categories, secondary categories, leaf categories clicked by the query term are added to obtain the total click rate. In this way, query terms with a small total click number can be filtered out, and query terms with a large total click number are reserved for subsequent steps.

Optionally, in order to prevent user cheating, the following steps may be further included:

and counting the click rate of each query word clicked by each user according to the website log data, and weakening the click rate of the query word clicked by the user if the click rate of a certain query word clicked by a certain user exceeds a preset value.

For example, if the click rate of a user a for clicking a query word X exceeds a certain threshold, the user is considered to have the malicious click intention or the special preference of the user a, and the click tendency of the query word X by most users cannot be reflected, or even cheating behavior may exist, and at this time, the click rate of the user may be appropriately attenuated. One commonly used method of attenuation is: the click rate of the user for clicking the query word is limited to a preset upper limit value, and certainly, other weakening methods can be provided in practical application, and are not listed one by one.

2) The category purchase distribution data corresponding to the query term may be obtained by:

according to the website log data, the user identification corresponding to the query word and the clicked commodity identification are associated with the user identification and the purchased commodity identification through the user identification, and category purchase distribution data which is clicked to a certain commodity through a certain query word and generates purchase is obtained.

Optionally, the query term obtained in 1) above may be directly used as a query term to be counted, and then the volume of deals guided by the query terms may be counted. Specifically, according to the association between the user id of the query term and the clicked commodity id, and the commodity id purchased by the user id, a certain query term is found and searched, a certain commodity is clicked, and the purchased data is generated. The search for the query term is considered to result in the purchase, and the query term is considered to have a purchase relationship with the category in which the item is located.

3) The category relevance distribution data corresponding to the query term may be obtained by:

and counting the number of commodities related to the query word in each category clicked by the query word according to the website log data to obtain category correlation distribution data corresponding to the query word.

Alternatively, the query term obtained in 1) above may also be directly used as the query term to be counted. For example, the number of commodities related to the query word in each category under all the primary category, the secondary category, the.

It should be noted that the number of items related to the query term is different from the number of items in the text matching sense. For example, when searching for a certain type of mobile phone, if matching according to the text, since the number of goods under the accessory category far exceeds the mobile phone category, the first recommended category is "3C digital accessory market". However, if recommendation is performed according to the category correlation distribution data described in the embodiment of the present application, since many accessories under the accessory category are used for digital products, not for mobile phones, the number of products related to mobile phones under the accessory category is smaller than that under the mobile phone category, and therefore, the mobile phone category is preferentially recommended.

It should be noted that the reference data may include any one or more of the above category click distribution data, category purchase distribution data, and category correlation distribution data, and need not be included at the same time. In addition, when any two or more kinds are included, the values obtained by weighting in a proportional manner may be used as reference data, and a specific example may be as shown in fig. 2.

Further, optionally, based on the processing of 1), 2), 3), any one or more of the following steps may be further included:

4) and counting navigation click data of the user when searching for the query word.

Specifically, the navigation click data refers to categories clicked according to navigation after the user inputs the query word, and commodities under the categories clicked. The navigation click data may also be used as reference data.

5) When the same meaning category exists in the category system, the data of the same category needs to be merged after the data are counted.

For example, the "notebook computer desk" belongs to the category of "furniture" and "computer accessories", and the two categories, namely "furniture" and "computer accessories", are synonymous and need to be combined.

6) And carrying out preliminary data noise removal processing on the click data, and filtering the categories with less click quantity.

The click rate of each category is counted, then the categories with small click rate are filtered, and the categories with large click rate are reserved for subsequent processing.

7) Data for stale categories are filtered out.

After the website system adjusts the category distribution each time, some categories may be deleted or merged into other categories, and such categories are the failure categories.

102, selecting a navigation recommendation model to be used by a query word according to reference data corresponding to the query word;

in the embodiment of the application, a plurality of navigation recommendation models are provided, and each navigation recommendation model is suitable for the query terms with different types of reference data. According to the distribution characteristics of the reference data, the navigation recommendation model to be used can be predicted. Examples are as follows:

1. firstly, summarizing and weighting the category click distribution data (click data for short), the category purchase distribution data (transaction data for short) and the category correlation distribution data (commodity quantity data for short) of each category according to a proportion, and normalizing. Wherein each type of data has a certain proportion of weight. The normalization is used for the convenience of subsequent calculations.

Alternatively, for a high frequency query term, the weight of the click data may be increased, and for a low frequency query term, the weight of the number of goods may be increased. For example, if the total number of clicks of a certain query word is less than 100, the weight of the number of commodities is increased, wherein the weight is 100/total number of clicks.

2. And backtracking the data obtained in the step 1 to the first-level category according to the category hierarchy.

And (3) obtaining data which is subjected to summary weighting and normalization corresponding to each category through the processing in the step 1, tracing the data of the lower-level categories to the upper-level categories layer by layer according to the category hierarchy, and finally tracing to the first-level categories to obtain statistical data corresponding to the first-level categories, wherein the statistical data is the data which is subjected to summary weighting and normalization processing.

3. According to the statistical data of the first-level category, selecting a navigation recommendation model to be used, wherein the specific selection method comprises the following steps:

if the clicks of the primary categories are concentrated in the category of 'books', calling a recommendation model of the category of 'books';

if 2 primary categories which are clicked most are concentrated in the categories of 'men' and 'women' then calling a parent-child category recommendation model;

if 2 primary categories with the maximum clicks are concentrated on categories of 'men' shoes 'and' women 'shoes', the 'shoes' recommendation model is called;

if the click ratio of the most clicked first-level category is less than 0.2, calling a divergent recommendation model;

if the click ratio of the most clicked first-level category is more than 0.98, calling a direct category recommendation model;

if the category information is hidden in the query word, calling a category search recommendation model;

if the description information of the commodities is hidden in the query words, calling a category attribute mixed arrangement recommendation model;

......；

and if the conditions are not met, calling a tiled category recommendation model.

The click ratio of the first-level class object refers to a ratio of the click quantity of a certain first-level class object to the total click quantity, namely the click ratio of the first-level class object.

Of course, the above description is only given by way of example, and in practical applications, another navigation recommendation model may be selected according to other situations.

103, inputting reference data corresponding to the query word into the selected navigation recommendation model, and calculating recommendation result data corresponding to the query word, wherein category and/or attribute data meeting the conditions are used as recommendation result data corresponding to the query word according to a bottom-up recommendation mode;

the recommendation result data may be category data (as shown in fig. 3), attribute data (as shown in fig. 4), or data in which categories and attributes are mixed (an example is not shown).

The listed multiple navigation recommendation models are basically based on an automatic expansion recommendation mode, and the core idea of the recommendation mode is as follows: according to the category hierarchy, a bottom-up recommendation mode is adopted, if the click or purchase ratio of a certain category reaches a certain threshold value, the category is directly upgraded, or the category is replaced by the previous category. Through the recommendation method, clicking or purchasing more low-level categories can directly cross the levels and promote the low-level categories to high-level display, so that the traditional mode of displaying the categories from top to bottom is eliminated, and the user is helped to quickly locate the commodity categories desired by the user.

It should be noted that, in the specific implementation process, the category and/or attribute data meeting the conditions are used as recommendation result data corresponding to the query term according to a bottom-up recommendation manner, and there may be at least two implementation manners:

one way is that, according to a bottom-up recommendation manner (i.e., the automatic expansion recommendation manner described in fig. 2), category and/or attribute data obtained after replacing a previous level category with a next level category meeting the condition is used as recommendation result data corresponding to the query term; the next-level category meeting the conditions is the next-level category with the click ratio exceeding a preset threshold, and the click ratio is the ratio of the click quantity of the next-level category to the total click quantity of the current category.

In another mode, according to a bottom-up recommendation mode (i.e., the automatic expansion recommendation mode described in fig. 2), the category and/or attribute data that meets the condition are directly hierarchically ranked and used as recommendation result data corresponding to the query term.

104, generating a recommended word list containing the mapping relation between the query words and recommended result data;

and 105, when receiving a user query word input online, querying the recommended word list, and outputting recommendation result data corresponding to the user query word.

Based on the above, the automatic expansion recommendation method will be described in detail below.

Referring to fig. 2, it is a flowchart of an automatic deployment recommendation method according to an embodiment of the present application.

Inputting various reference data, such as category click distribution data (click data for short), category purchase distribution data (transaction data for short), category correlation distribution data (commodity quantity data for short) and the number of categories expected to be recommended by the first current category, and executing the following steps:

step 201, determining a current category from top to bottom according to a category hierarchy;

for example, suppose that the category distribution corresponding to the query word "clothing" includes a category "women's clothing" and a category "men's clothing", and each includes a plurality of sub-categories under the category "women's clothing" and "men's clothing", such as a sub-category "T-shirt", "POLO shirt", "long-sleeved shirt", etc. under the category "men's clothing". From top to bottom in the category hierarchy, a number of sub-categories may first be determined to be the current category, respectively.

The method for determining the current category by each navigation recommendation model is different, and will be described in detail later when the navigation recommendation model is introduced.

Step 202, if the current category is a leaf category, returning recommendation result data as the current category;

the leaf category is the lowest category in the category hierarchy.

Step 203, otherwise, summarizing the reference data corresponding to the query word to the next level category of the current category;

the category click distribution data (click data for short), the category purchase distribution data (transaction data for short) and the category correlation distribution data (commodity quantity data for short) of each category are summarized and weighted in proportion and normalized. And then, tracing the data of the lower-level category to the next-level category of the current category layer by layer according to the category hierarchy.

For example, assume that when the current category is the category "T-shirt" under the category "men's clothing", the reference data under the category "T-shirt" are all collected to the next-level category "long-sleeved T-shirt", "short-sleeved T-shirt", "seven-cent/five-quarter-sleeved T-shirt", "sleeveless T-shirt", and the like of the "T-shirt".

Step 204, selecting a next-level category with a click ratio exceeding a preset threshold, wherein the click ratio is a ratio of the click quantity of the next-level category to the total click quantity of the current category;

for example, assuming that the click ratio of each of the "long-sleeved T-shirt" and "short-sleeved T-shirt" categories exceeds a preset threshold value in the next-level category of "men's clothing" - > "T-shirt", the two categories are selected. Similarly, among the next-level categories of "men's clothing" - > "POLO shirts", the next-level categories in which the hit ratios all exceed the preset threshold include "long-sleeved POLO shirts" and "short-sleeved POLO shirts". Among the next-level categories of "men's clothing" - > "long-sleeved blouse", the next-level categories in which the click occupation ratios all exceed the preset threshold include "collar long-sleeved blouse" and "collar-free long-sleeved blouse".

Then, for the next hierarchical category selected, reference data for each category may also be filtered, for example, reference data under two categories of "long-sleeved T-shirt" and "short-sleeved T-shirt" are filtered, and subsequent calculations for the category "T-shirt" will use these filtered data instead of querying all reference data corresponding to the words "clothing" or "men's clothing".

Step 205, calculating the number of the categories expected to be recommended in the next level according to the number of the categories expected to be recommended currently;

the number X of categories desired to be recommended currently is set and input before step 201, and accordingly, the number Y of categories desired to be recommended at the next level of the current category can be calculated according to the following formula:

click ratio of each category +1

For example, suppose X takes a value of 6, the click duty ratio for the category "men's clothing" - > "T-shirt" is 0.5, the click duty ratio for the category "men's clothing" - > "POLO shirt" is 0.4, and the click duty ratio for the category "men's clothing" - > "long-sleeved shirt" is 0.1; then the number Y of categories desired to be recommended for the next level of the category "T-shirts" is 4, the number Y of categories desired to be recommended for the next level of the category "POLO shirts" is 3(3.4 rounded), and the number Y of categories desired to be recommended for the next level of the category "long-sleeved shirts" is 1(1.6 rounded).

Step 206, if the number of the selected next-level categories is larger than the number of the categories expected to be recommended by the next level, canceling the recommendation to the next level, and returning recommendation result data as the current categories;

as described above, if the number 2 of next-level categories selected from the "men's clothing" - > "long-sleeved shirt" is greater than the number 1 of categories expected to be recommended in the next level, the recommendation to the next level of the "long-sleeved shirt" is cancelled, and data of which recommendation result data is the "long-sleeved shirt" category is returned.

Step 207, if the number of the selected next-level categories is less than or equal to the number of the categories expected to be recommended by the next level and the next-level categories are not leaf categories, replacing the current categories with the next-level categories, determining each category of the next-level categories as the current category, and continuing to recommend the next-level categories of the current categories according to the steps;

as described above, the number 2 of the next-level categories selected by the man's clothing "- >" T-shirt "is smaller than the number 4 of the categories expected to be recommended by the next level, and the categories" long-sleeved T-shirt "and" short-sleeved T-shirt "of the next level are not leaf categories, the categories" long-sleeved T-shirt "and" short-sleeved T-shirt "are replaced by the categories" T-shirt ", and the categories" long-sleeved T-shirt "and" short-sleeved T-shirt "are determined as the current categories, respectively, and recursive calculation is continued according to the above steps 202 to 208.

Similarly, if the number 2 of next-level categories selected by the "men's clothing" - > "POLO shirts" is smaller than the number 3 of next-level categories expected to be recommended, and the "long-sleeve POLO shirts" and the "short-sleeve POLO shirts" are not leaf categories, the "long-sleeve POLO shirts" and the "short-sleeve POLO shirts" are substituted for the "POLO shirts", and the "long-sleeve POLO shirts" and the "short-sleeve POLO shirts" are determined as the current categories respectively, and the recursive calculation is continued according to the above steps 202 to 208.

In the subsequent recursive calculation, the long-sleeved T-shirt, the short-sleeved T-shirt, the long-sleeved POLO shirt and the short-sleeved POLO shirt may be replaced by the category of the next level. Therefore, the categories which are located at the lower layer of the category distribution and have high weighted scores can be directly promoted to the categories at the upper layer through layer-by-layer screening and replacement, so that a user can quickly find the categories, and the categories can be found at the lowest layer without clicking layer by layer from top to bottom.

And 208, if the number of the selected next-level categories is smaller than the number of the categories expected to be recommended by the next level and the next-level categories are leaf categories, replacing the current categories with the next-level categories, and purchasing distribution data according to the categories in the reference data to complement the category data.

Assuming that the next hierarchy categories "long-sleeved POLO shirts" and "short-sleeved POLO shirts" are all the leaf categories, the categories "long-sleeved POLO shirts" and "short-sleeved POLO shirts" are also replaced with the categories "POLO shirts", but the recursive computation is not performed any more. At this time, the number of recommended categories is expected to be 3, but the number of categories is 2 at present, and category data may be complemented by category purchase distribution data (transaction data for short), and for example, another category with the largest transaction volume may be selected according to the transaction situation to fill up under the category of "men clothing".

Finally, after the first calculation processing of steps 201 to 208, the result shown in fig. 3 can be obtained. It is assumed that the long-sleeved POLO blouse and the short-sleeved POLO blouse are not in the order of leaves. If the calculation of "long-sleeved T-shirt", "short-sleeved T-shirt", "long-sleeved POLO shirt" and "short-sleeved POLO shirt" is continued according to the steps in fig. 2, the display result in fig. 3 may be changed.

Based on the automatic development recommendation method, a specific recommendation method of each navigation recommendation model is described in detail below.

1) Recommendation model for father and son categories

The recommendation method comprises the following steps:

step A1, summarizing reference data corresponding to the query words into a first class;

as mentioned above, the reference data includes category click distribution data (abbreviated as click data), category purchase distribution data (abbreviated as transaction data), category correlation distribution data (abbreviated as commodity quantity data), and the like, and the summarizing includes weighting and normalization processing in proportion.

Step B1, respectively taking the two primary categories which are clicked most as father categories;

specifically, if there are two primary categories that have the most clicks, that is, the click volumes of the two primary categories are the same, the two primary categories are selected as parent categories. And if one primary category is clicked most and one primary category is clicked second most, selecting the two primary categories as parent categories.

However, if there is only one primary category clicked most and there are two primary categories clicked second most, then the parent category may be determined as follows:

click on category second most > click on category maximum 0.05;

alternatively, clicking on the second plurality of categories may expand and clicking on the second plurality of categories clicks > 0.1 clicks on the most numerous categories.

And if any condition is met, recommending the first and second categories with the most clicks as parent categories.

Step C1, filtering out reference data under the father category respectively, and recommending in an automatic expansion recommending mode;

and finally, taking the qualified category and/or attribute data as recommendation result data.

For example, if the parent category is "men's clothing" and "women's clothing", the recommendation is made for each of the sub-categories of "men's clothing" and "women's clothing" in the auto-expansion recommendation manner shown in fig. 2. At this time, each sub-category is initially set as the current category.

Step D1, taking the remaining first-level categories as sub-categories of the same father category, and recommending in an automatic expansion recommendation mode;

this parent category may be defined as the "other" category as shown in FIG. 3. For each sub-category of the 'other' categories, using the automatic development recommendation mode shown in fig. 2, according to the bottom-up recommendation, checking whether the sub-category can be replaced by the next category meeting the conditions, and finally obtaining the recommendation result data with the new category hierarchy.

At step E1, if the "other" category contains less than the desired number of sub-categories, the category data is complemented with category purchase distribution data (transaction data for short).

2) Divergence recommendation model

The recommendation method comprises the following steps:

step A2, summarizing reference data corresponding to the query words into a first class;

B2, sorting the primary categories according to the click rate from high to low, and selecting a plurality of primary categories which are sorted in the front;

for example, a maximum of 16 primary categories are selected.

Step C2, for each selected first-level category, according to bottom-up recommendation, checking whether the selected first-level category can be replaced by a next-level category meeting the conditions, and finally using the category and/or attribute data meeting the conditions as recommendation result data;

d2, taking the remaining first class as the lower class of other class, and recommending in an automatic expansion recommending mode;

at step E2, if the "other" category contains less than the desired number of sub-categories, the category data is complemented with category purchase distribution data (transaction data for short).

3) Direct category recommendation model

The direct category recommendation model is that when the query of a user has a definite corresponding relation with a foreground category, the user enters the category to search, and a search result list only returns commodities under the category. The recommendation method comprises the following steps:

step A3, summarizing reference data corresponding to the query words into a first class;

Step B3, selecting a first class with a click ratio exceeding a preset threshold value, wherein the click ratio is the ratio of the click quantity of the first class to the total click quantity;

for example, a primary category with a click ratio greater than > 0.98 is selected as the through recommendation category.

And step C3, checking whether the quantity of the selected related commodities of each primary category is greater than the preset quantity, if so, determining the primary category as a direct category, and using the direct category as recommendation result data corresponding to the query words.

For example, the preset number is set to 50,.

Step D3, if the category reached is a leaf category, then the attribute under that category also needs to be recommended.

The recommendation about the attribute can be seen in the category attribute shuffling recommendation model in 5) below.

4) Search recommendation model with categories

If the category information is hidden in the query word, the category information can be directly positioned to the category, and then the category recommendation is carried out under the category by adopting an automatic expansion recommendation mode.

The recommendation method of the search recommendation model with the categories comprises the following steps:

step A4, searching corresponding category according to the category information hidden in the query vocabulary;

for example, if the query word "men's T-shirt" includes the category "women's clothing", reference data under the category "men's clothing" is looked up.

Step B4, summarizing the reference data corresponding to the query words to the next level category of the categories;

for example, the reference data is weighted and normalized in scale totals, going back to the next hierarchical category "T-shirt", "POLO shirt", "long-sleeved shirt", etc., of the "men's clothing" category.

C4, sorting the next-level categories according to the click rate from high to low, and selecting a plurality of next-level categories which are sorted in the front;

for example, three categories "T-shirt", "POLO shirt" and "long-sleeved shirt" are selected.

And D4, checking whether each next-level category can be replaced by a qualified next-level category according to the bottom-up recommendation, and finally taking the qualified categories and/or attribute data as recommendation result data.

For example, automatic unfolding recommendations for "T-shirts", "POLO shirts" and "long-sleeved shirts" are made in accordance with the method of fig. 2.

5) Category attribute mixed-ranking recommendation model

The embodiment of the application further provides an attribute recommendation function on the basis of providing category recommendation. The attribute recommendation is direct attribute recommendation, when the query of a user and a certain value of the attribute have a definite corresponding relation, the condition is taken to carry out searching, and the searching result list only returns the commodity with the attribute value.

The recommendation method comprises the following steps:

step A5, summarizing the reference data corresponding to the query words to leaf categories, wherein attributes appear in the leaf categories;

step B5, sorting the attributes according to the click rate and/or the entropy difference from high to low, and selecting a plurality of attributes which are sorted in the front;

for example, the top 5 attribute recommendations are selected.

Wherein, the entropy difference refers to the information gain of selecting a certain attribute.

Step C5, for each selected attribute, sorting the attribute values according to the click rate from high to low, and selecting a plurality of attribute values sorted in front;

since each attribute has multiple attribute values, it is also necessary to perform rank screening on the attribute values. For example, if the attribute is a value type attribute, the 6 values with the highest click are selected and sorted by value.

Step D5, text matching is carried out on the attribute value and the query word, and if the attribute value and the query word are completely matched or the synonym is matched, attribute preselection is carried out;

step E5, using the attributes subjected to attribute preselection and the corresponding attribute values thereof as recommendation result data corresponding to the query word;

step F5, for the attribute to be pre-selected, judging whether there is sub-attribute, if there is sub-attribute, the sub-attribute is displayed.

And if yes, replacing the attributes with the sub-attributes, and taking the replaced sub-attributes and the attribute values thereof as recommendation result data corresponding to the query words.

For example, referring to fig. 4, the result is obtained under the category of one-piece dress corresponding to the query word "one-piece dress", where "brand", "sleeve length", "material", "dress length" and "pattern" refer to attributes, and the information on the right side of each attribute is an attribute value, for example, the attribute value of the attribute "pattern" includes "solid color", "flower color" and "stripe".

For another example, referring to fig. 5, after the corresponding query word "nokia" gives a brand property through recommendation, since the corresponding brand property also has a sub-property, the sub-property and its property value are directly exhibited.

6) Recommendation model for book category and recommendation model for shoe category

The recommendation of the explicit categories is similar to the recommendation of the band categories, and after the explicit categories are directly positioned to a certain category, the category recommendation is carried out under the category by adopting an automatic unfolding recommendation mode. But the difference lies in that: the recommendation of the belt category is that category information is hidden in a query word, but the recommendation of the book or the shoe does not have the category information hidden in the query word, but the recommendation model of the book or the shoe can be judged to be called according to the centralized click of the primary category.

Of course, in addition to these two recommendation models, there may be "cell phone" or other commodity specific category recommendation models.

7) Tiled category recommendation model

The tiling recommendation is a method similar to the divergence recommendation, but the number of categories of the tiling recommendation is relatively small, generally 8 primary categories are recommended at most, while the number of categories of the divergence recommendation is large, generally 16 primary categories can be recommended at most. And for each primary category of the tiled recommendation, performing category recommendation by adopting an automatic expansion recommendation mode.

The above lists only some typical navigation recommendation models, but the scope of protection of the present application is not limited to the above lists.

In summary, the intelligent navigation provided by the embodiment of the application can return category or attribute navigation related to the search intention of the user according to the query word of the user, and the user can quickly find the needed goods without clicking many times, so that the search time of the user is saved, and the access burden of a related server is also reduced.

The biggest difference from the category clicking navigation is that the intelligent navigation of the application adopts a bottom-up recommendation mode, recommendation result data corresponding to the query words is obtained after replacing a next-level category meeting conditions with a previous-level category according to bottom-up recommendation, and if the clicking or purchasing percentage of a certain category reaches a certain threshold value, the upward backtracking process is stopped, so that the mode of displaying the categories from top to bottom is eliminated, and the commodity category desired by the user can be positioned more quickly.

In addition, the intelligent navigation of the application introduces an attribute recommendation function and an attribute preselection function, and if the description information of the commodity is hidden in the query words of the user, the navigation provides the attribute preselection function, so that the target commodity is accurately locked.

Also, when there is less historical click data that can be referenced, the richness of the recommended categories may be reduced if these click data are utilized purely. Aiming at the problem, the intelligent navigation of the application also refers to various historical data such as transaction data, the number of related commodities, the click navigation condition of a user and the like, and enriches the recommendation categories.

It should be noted that the foregoing method embodiments are described as a series of acts or combinations for simplicity in explanation, but it should be understood by those skilled in the art that the present application is not limited by the order of acts described, as some steps may occur in other orders or concurrently depending on the application. Further, those skilled in the art will also appreciate that the embodiments described in the specification are presently preferred and that no particular act is required to implement the invention.

Based on the description of the above method embodiment, the present application further provides a corresponding navigation system embodiment to implement the content described in the above method embodiment.

Fig. 6 is a structural diagram of an electronic commerce website navigation system according to an embodiment of the present application.

The navigation system may include the following modules:

the data analysis module 10 is configured to perform statistical analysis on user behaviors on website log data to obtain reference data required by a navigation recommendation model corresponding to a query term, where the reference data includes category click distribution data, and/or category purchase distribution data, and/or category correlation distribution data corresponding to the query term;

the model prediction module 20 is configured to select a navigation recommendation model to be used by a query term according to reference data corresponding to the query term;

the category attribute recommendation module 30 is configured to input reference data corresponding to the query term into the selected navigation recommendation model, and calculate recommendation result data corresponding to the query term, where the category and/or attribute data meeting the conditions are used as recommendation result data corresponding to the query term according to a bottom-up recommendation manner;

a recommended word list generating module 40, configured to generate a recommended word list including a mapping relationship between the query word and the recommendation result data;

and the online query module 50 is configured to query the recommended word list when receiving a user query word input online, and output recommended result data corresponding to the user query word.

The category attribute recommendation module 30 replaces the category and/or attribute data obtained after the upper-level category is replaced by the lower-level category meeting the condition according to a bottom-up recommendation mode, and the category and/or attribute data is used as recommendation result data corresponding to the query word; the next-level category meeting the conditions is the next-level category with the click ratio exceeding a preset threshold, and the click ratio is the ratio of the click quantity of the next-level category to the total click quantity of the current category.

The navigation recommendation model can comprise a father-son category recommendation model, a divergent recommendation model, a category search recommendation model, a category attribute mixed recommendation model, a direct category recommendation model, a tiled category recommendation model, a book category recommendation model, a shoe category recommendation model and the like, and various types of recommendation models.

Specifically, the data analysis module 10 may perform statistical analysis by using the method of step 101 in fig. 1, and obtain various reference data corresponding to the query word, which is not described in detail herein.

The model prediction module 20 may select the navigation recommendation model using the method of step 102 in fig. 1, which will not be described in detail herein.

Further, the category property recommendation module 30 may include an automatic expansion recommendation sub-module, and the automatic expansion recommendation sub-module may perform automatic expansion recommendation calculation by using the method shown in fig. 2, which is not described in detail herein.

For the above listed navigation recommendation models, the model prediction module 20 may calculate recommendation result data corresponding to the query term by using different recommendation methods, which may be specifically described in the above descriptions 1) to 7) of seven models.

For the embodiment of the navigation system, since it is basically similar to the embodiment of the method, the description is simple, and the relevant points can be referred to the partial description of the embodiment of the relevant method.

In addition, based on the above description of the method embodiment, the present application also provides another application system for intelligent navigation in an e-commerce website, as shown in fig. 7.

Referring to FIG. 7, a diagram of an application system architecture for intelligent navigation in an e-commerce web site is shown.

The application system mainly comprises three modules of category attribute recommendation, offline data processing and online real-time query.

The core of the category attribute recommendation module is as follows: the behavior data (such as search logs, click logs and purchase transaction logs) of the user are analyzed by using a distributed computing function provided by a cloud computing platform to obtain reference data required by a recommendation system, and intelligent navigation recommendation data such as categories or attributes are recommended by using a corresponding model according to the type of the query word.

After the offline data processing module obtains the intelligent navigation recommendation data, processing logics such as consistency check, slope leveling adjustment, category renaming, failure check, commodity quantity check and the like are added to adjust the recommendation effect, and a final recommendation word list is formed by combining the manual editing effect of part of ultrahigh frequency query words.

The online real-time Query module finally reads the category attribute recommendation data into a QP server (a server with the full name of Query player for rewriting Query words and adding auxiliary category attribute information) by means of a strong apache framework, and the QP provides real-time online Query service for the front-end server.

The main working process of the application system is as follows:

1. the log collection server converts the user behavior data into recognizable records and writes the records into an HDFS storage system at intervals;

2. due to the fact that the log data volume is huge, the whole recommended computing process depends on a cloud computing platform. Firstly, analyzing an original log by using a cloud computing platform to obtain reference data such as category click distribution, category purchase distribution and category correlation distribution (namely category commodity quantity information) related to query words; calling a corresponding navigation recommendation model according to the reference data type of the query word to calculate recommendation of category attributes;

3. performing a series of processing on output data of the cloud computing platform offline, including consistency conversion, slope leveling, category renaming, commodity quantity inspection, introduction of manual editing data and the like, to form recommendation data organized in a certain format;

4. compiling the final recommended data into a binary file, loading the binary file to a QP server, and providing real-time query service under an apache framework;

5. and after the user inputs the keyword in the search box, clicking a search button, sending a request to the QP server after the front-end server receives the request, and returning data including the intelligent navigation recommendation information by the QP server.

The embodiments in the present specification are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.

Further, the word "and/or" above means that the relation "and" or "is included herein, wherein: if the scheme A and the scheme B are in an 'and' relationship, the method indicates that the scheme A and the scheme B can be simultaneously included in a certain embodiment; if the scheme a and the scheme B are in an or relationship, this means that in some embodiment, the scheme a may be included separately, or the scheme B may be included separately.

The method and the system for navigating the e-commerce website provided by the application are introduced in detail, a specific example is applied in the text to explain the principle and the implementation mode of the application, and the description of the embodiment is only used for helping to understand the method and the core idea of the application; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.

Claims

1. An electronic commerce website navigation method is characterized by comprising the following steps:

selecting a navigation recommendation model to be used by the query word from a plurality of navigation recommendation models according to reference data corresponding to the query word; wherein each of the plurality of navigation recommendation models is applicable to query terms having different types of reference data;

inputting reference data corresponding to the query word into the selected navigation recommendation model, and calculating recommendation result data corresponding to the query word, wherein the category and/or attribute data obtained after replacing the upper-level category with the lower-level category meeting the conditions in a bottom-up recommendation mode are used as the recommendation result data corresponding to the query word; the next-level category meeting the condition is the next-level category with the click ratio exceeding a preset threshold value, the previous-level category is the current category corresponding to the query word, and the click ratio is the ratio of the click rate of the next-level category to the total click rate of the current category;

2. The method of claim 1, wherein the category click distribution data corresponding to the query term is obtained by:

3. The method of claim 2, further comprising:

accumulating the click rate of each category clicked by the query word to obtain the total click rate of each query word;

and filtering out the query words with the total click rate less than a preset value.

4. The method of claim 3, further comprising:

5. The method according to any one of claims 1 to 4, wherein the category purchase distribution data corresponding to the query term is obtained by:

6. The method according to any one of claims 1 to 4, wherein the category relevance distribution data corresponding to the query term is obtained by:

7. The method of claim 1, wherein the navigation recommendation model comprises a parent-child category recommendation model, and the inputting the reference data corresponding to the query word into the parent-child category recommendation model and calculating the recommendation result data corresponding to the query word comprises:

summarizing reference data corresponding to the query words to a first class;

respectively taking the two primary categories which are clicked most as father categories, and taking the rest primary categories as sub-categories under the same father category;

and checking whether each sub-category of the parent categories can be replaced by the next level category meeting the conditions according to bottom-up recommendation, and finally taking the category and/or the attribute data meeting the conditions as recommendation result data.

8. The method of claim 1, wherein the navigation recommendation model comprises a divergence recommendation model, and the inputting the reference data corresponding to the query word into the divergence recommendation model and calculating the recommendation result data corresponding to the query word comprises:

summarizing reference data corresponding to the query words to a first class;

sorting the primary categories according to the click rate from high to low, and selecting a plurality of primary categories which are sorted in the front;

and checking whether each selected primary category can be replaced by the next qualified category according to bottom-up recommendation, and finally taking the qualified categories and/or attribute data as recommendation result data.

9. The method of claim 1, wherein the navigation recommendation model comprises a category search recommendation model, if category information is implicit in the query term, reference data corresponding to the query term is input into the category search recommendation model, and calculating recommendation result data corresponding to the query term comprises:

searching corresponding categories according to category information hidden in the query vocabulary;

summarizing reference data corresponding to the query words to a next-level category of the categories;

sorting the next-level categories according to the click rate from high to low, and selecting a plurality of next-level categories which are sorted in the front;

and checking whether each next-level category can be replaced by a qualified next-level category according to bottom-up recommendation, and finally taking the qualified categories and/or attribute data as recommendation result data.

10. A method according to any one of claims 7 to 9 wherein checking for each category according to a bottom-up recommendation whether it can be replaced by a next level category that is eligible comprises:

determining the current category from top to bottom according to the category hierarchy;

if the current category is the leaf category, returning recommendation result data as the current category;

otherwise, summarizing the reference data corresponding to the query words to the next level category of the current category;

selecting a next-level category with a click ratio exceeding a preset threshold, wherein the click ratio is a ratio of the click quantity of the next-level category to the total click quantity of the current category;

calculating the number of the categories expected to be recommended in the next level according to the number of the categories expected to be recommended currently;

if the number of the selected next-level categories is larger than the number of the categories expected to be recommended by the next level, canceling the recommendation to the next level, and returning recommendation result data as the current categories;

and if the number of the selected next-level categories is less than or equal to the number of the categories expected to be recommended by the next level and the next-level categories are not leaf categories, replacing the current categories with the next-level categories, determining each category of the next-level categories as the current category, and continuously recommending the next-level categories of the current categories according to the steps.

11. The method of claim 10, wherein:

and if the number of the selected next-level categories is smaller than the number of the categories expected to be recommended by the next level and the next-level categories are leaf categories, replacing the current categories with the next-level categories, and selecting the number of the categories with the difference according to the category purchase distribution data in the reference data for supplement to reach the number of the categories expected to be recommended by the next level.

12. The method according to claim 1, wherein the navigation recommendation model comprises a category attribute mixed recommendation model, and if the description information of the commodity is hidden in the query term, the reference data corresponding to the query term is input into the category attribute mixed recommendation model, and the recommendation result data corresponding to the query term is calculated, including:

summarizing reference data corresponding to the query words to leaf categories, wherein attributes appear in the leaf categories;

sorting the attributes from high to low according to the click rate and/or the entropy difference, and selecting a plurality of attributes which are sorted in the front;

for each selected attribute, sorting the attribute values according to the click rate from high to low, and selecting a plurality of attribute values which are sorted in front;

and taking the selected attributes and the attribute values corresponding to the attributes as recommendation result data corresponding to the query words.

13. The method of claim 12, further comprising:

judging whether each selected attribute has a sub-attribute, and if so, replacing the attribute with the sub-attribute;

and taking the replaced sub-attributes and the attribute values thereof as recommendation result data corresponding to the query words.

14. The method according to claim 1 or 12, wherein the navigation recommendation model comprises a direct category recommendation model, the inputting of the reference data corresponding to the query term into the direct category recommendation model, and the calculating of the recommendation result data corresponding to the query term comprises:

summarizing reference data corresponding to the query words to a first class;

selecting a first class with a click ratio exceeding a preset threshold, wherein the click ratio is the ratio of the click quantity of the first class to the total click quantity;

and checking whether the quantity of the selected related commodities of each primary category is greater than a preset quantity, if so, determining the primary category as a direct category, and using the direct category as recommendation result data corresponding to the query words.

15. An electronic commerce website navigation system, comprising:

the model prediction module is used for selecting a navigation recommendation model to be used by the query word from a plurality of navigation recommendation models according to the reference data corresponding to the query word; wherein each of the plurality of navigation recommendation models is applicable to query terms having different types of reference data;

the category attribute recommendation module is used for inputting reference data corresponding to the query word into the selected navigation recommendation model and calculating recommendation result data corresponding to the query word, wherein the category and/or attribute data obtained after the next-level category meeting the conditions is replaced by the previous-level category according to a bottom-up recommendation mode are used as the recommendation result data corresponding to the query word; the next-level category meeting the condition is the next-level category with the click ratio exceeding a preset threshold value, the previous-level category is the current category corresponding to the query word, and the click ratio is the ratio of the click rate of the next-level category to the total click rate of the current category;