CN108804577A - A kind of predictor method of information label interest-degree - Google Patents

Abstract

The invention discloses a method for estimating the degree of interest in information tags, which includes: creating and maintaining a candidate information base containing tags; obtaining user attribute information tag interest degree vectors according to user demographic information; and obtaining information tags of multiple users within a preset time period Preprocess the historical behavior data to obtain a trained deep learning model; obtain the current user's historical behavior data and perform preprocessing to obtain the current user's user behavior information label interest degree vector; according to the current user's user attribute information label interest degree The vector and the user behavior information tag interest degree vector are calculated to obtain the user-information tag interest degree vector, and finally determine several information tags that the user is most interested in. The invention solves the cold start problem of user interest degree estimation, avoids the problem of low information quality that often occurs when directly selecting information from the Internet, and reduces the calculation amount of user interest degree estimation problem, and is applicable to each sample Scenes with multiple tags.

Description

A Method for Predicting the Interest Degree of Information Tags

technical field

The invention relates to a method for estimating interest degree of information tags, which belongs to the technical field of computing.

Background technique

With the rapid development of the Internet, the amount of information on the Internet is huge and growing explosively, but the quality of the information is uneven. If the user interest degree estimation operation is directly performed on all the obtained information, it is likely that the information with poor content quality Pushing it to users will affect the user experience, and the operation of estimating user interest for all information will increase the calculation amount of the algorithm and cause a waste of computing resources.

Although the information browsed by users is different, the information tags corresponding to the information can often be divided into several major categories, and the user's interest in a certain information tag lasts much longer than that of a certain information. For example, after a user browses an information tagged with "Finance", the user will not read the same information again, but the user is still interested in other information with the "Finance" tag. Therefore, it is of great significance to find the information tags that users are interested in by estimating the user's interest in information tags, which is of great significance to the research and application of personalized information push.

A common problem in current user interest estimation methods used in practical applications is the cold-start problem, that is, how to estimate user interest when the user has not browsed information.

In the prior art, there is a method based on a recurrent neural network for estimation. This method sequentially inputs the information tags corresponding to the information browsed by the user into the recurrent neural network to train and predict the information tags that the user is interested in. This method can take advantage of the time series features in the user's historical behavior, so the effect is better when the training samples are sufficient and the recurrent neural network parameters are adjusted properly. But this method has the following defects:

1 It is impossible to estimate user interest when the user has not browsed any information.

2 Unable to use the user's demographic information, such as gender, age, region, etc.

In the prior art, there is a method based on TF-IDF (term frequency-inverse text frequency index) to obtain the keywords of each information, and to obtain the user's opinion of each keyword by statistically analyzing the keywords in the information currently browsed by the user. interest.

The TF-IDF method is a statistical method. For a piece of information, the frequency of a word appearing in the information reflects the importance of the word. The more times a word appears in the information, the more the word appears. The more important the word is in the news, the importance of the word decreases as the frequency of occurrence of the word in the overall news increases. That is, if a word appears frequently in the current information and rarely appears in other information, it is considered that the word can well represent the information, and at this time the word is the keyword of the current information. By statistically analyzing the keywords in the information browsed by users, the user's interest in each keyword can be obtained, which can be used for subsequent personalized information push based on information keywords, but this method has the following defects:

1 Count the number of times the word appears in the current information and the number of times it appears in all the information for each word in each information, which requires a large amount of calculation.

2 The distribution of keywords obtained through statistics is too wide, and the content represented by each keyword may be specific to a small field, which is not conducive to controlling the estimated range of information that users are interested in. For example, using TF-IDF to obtain the keyword corresponding to a certain information browsed by the user is "Lin Daiyu". If the user's interest degree is estimated based on this keyword, the subsequent information push to the user is likely to be too focused on including " "Lin Daiyu" information, but it is difficult to well expand to "Dream of Red Mansions" or "Chinese Literature", which affects the interest rate estimation and information push effect.

3 Even if the keywords contained in a piece of information have a high score in user interest estimation, they may not be able to arouse user interest due to information quality problems.

4. When the user has not browsed any information, it is impossible to estimate the user interest.

5. All the information that a user has browsed has the same status in estimating user interest, and it cannot reflect the difference in estimating user interest at the current moment for information in different browsing sequences. However, information that has been browsed recently has a greater impact on the current interest degree estimation.

6 Unable to use the user's demographic information, such as gender, age, region, etc.

In the prior art, there is a method for estimating user interest based on a gradient boosting decision tree method. The gradient boosting decision tree (GBDT) is a machine learning method that makes joint decisions by iterating multiple regression trees. The gradient boosting decision tree is composed of multiple regression trees, and each regression tree is fitted to obtain the current regression tree by learning the results and residuals of all previous regression trees. Residual here refers to the actual value minus the predicted value. The results of all regression trees are added together as the final result of the gradient boosted decision tree. The method can simultaneously utilize the demographic information of the user and the information label information corresponding to the information browsed by the user. But this method has the following defects:

1 The gradient boosting decision tree is essentially suitable for regression problems, or completes binary classification problems by setting thresholds. For the problem of estimating the user's interest in information tags, the information tag library contains a large number of tags, and each information contains more than one tag. The gradient boosting decision tree can only get the user's interest in a certain The estimated value of the interest degree of the label. If you want to obtain the user's interest degree for each information label, you need to use the gradient boosting decision tree method for each information label to estimate the interest degree. The calculation amount is when solving the binary classification problem. The calculation amount of the gradient boosting decision tree is m times (m is the total number of labels in the information label library), and the calculation amount is relatively large.

2. All the information that the user has browsed has the same status in estimating the user interest degree, which cannot reflect the difference in estimating the user interest degree at the current moment for the information in different browsing order. However, information that has been browsed recently has a greater impact on the current interest degree estimation.

Contents of the invention

In view of the above-mentioned defects, the present invention provides a method for estimating the interest degree of information tags. By establishing a user attribute-information tag interest degree vector, the cold start problem of user interest degree estimation is solved, and a candidate information base containing tags is established to avoid It solves the problem of low information quality that often occurs when directly selecting information from the Internet, reduces the amount of calculations for user interest estimation, and is suitable for scenarios where each sample contains multiple tags.

In order to achieve the above object, the present invention is specifically realized through the following technical solutions:

The present invention provides a method for estimating interest in information tags, the method comprising:

Create and maintain a tagged candidate repository;

According to the demographic information of the user, the user attribute-information tag interest degree vector is obtained;

Obtain and preprocess the historical behavior data of multiple users within a preset time period, input the deep learning model for training to obtain the trained deep learning model;

Obtain and preprocess the historical behavior data of the current user, and use the trained deep learning model to calculate the current user's user behavior-information tag interest degree vector;

Calculate the user-information tag interest degree vector according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector, and finally determine several information tags that the user is most interested in.

Further, the steps of creating and maintaining a tagged candidate information base include:

Select one or more tags that best match the content of the information from the preset information tag library as the tag of the information, and add the tagged information to the candidate information library containing tags; for each information in the candidate information library , according to the information tag corresponding to the information, each information is represented by an m-dimensional information vector, m is the total number of tags in the default information tag library; when the information contains a tag T _j , the j-th dimension of the m-dimensional information vector The value is 1, otherwise the value of the jth dimension is 0;

Regularly maintain the tagged candidate information base, add new information, and remove out-of-date information.

Further, the user demographic information includes, but is not limited to: one or more of the obtainable gender, age and/or geographical information to classify users into several groups.

Further, said obtaining the user attribute-information label interest degree vector according to the user demographic information includes:

The user attribute-information tag interest degree vector H _ij of the i-th group G _i to the j-th information tag T _j is:

The value of H _ij is between [0, 1].

Further, the acquisition of historical behavior data of multiple users within a preset time period and preprocessing, inputting the deep learning model for training to obtain a trained deep learning model, including:

Obtain the information vector corresponding to each information browsed in the historical behavior data of multiple users within the preset time period, and input the information vector into the recurrent neural network model in the order of browsing information time sequence to train the recurrent neural network model, and obtain A trained deep learning model.

Further, the acquisition of the historical behavior data of the current user and preprocessing, using the trained deep learning model to calculate the current user's user behavior-information label interest degree vector, including:

Obtain the information vector corresponding to each information browsed in the current user's historical behavior data, and arrange them in chronological order;

Input the information vectors corresponding to each information in the current user's historical behavior data into the trained deep learning model in chronological order. After all the information vectors corresponding to each information in the historical behavior data are input, the The m-dimensional prediction vector obtained by the trained deep learning model is the current user's user behavior-information tag interest degree vector.

Further, the user-information tag interest degree vector is calculated according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector, and finally determines several information tags that the user is most interested in, including:

According to the calculated user attribute-information tag interest degree vector and user behavior-information tag interest degree vector of the current user, the m-dimensional interest degree vector of the current user to the information tag can be calculated, and the calculation formula is as follows:

V (user, information label) = (1-w) * V (user attribute, information label) + w*V (user behavior, information label)

Among them, V (user, information label) is the m-dimensional interest degree vector of the current user to the information label; V (user behavior, information label) is the user behavior-information label interest degree vector; V (user attribute, information label) is the current The user's user attribute - information label interest degree vector; w represents the weight of V (user behavior, information label) in calculating the current user's interest degree vector for information label, and the value of w should always be in [0,1] within range.

Further, the calculation formula of w is as follows:

w=tanh(a*number of information browsed by the current user within the preset time period T)

Among them, tanh is the hyperbolic tangent function, and a is a constant greater than 0.

A method for estimating the interest degree of information tags provided by the present invention innovatively proposes a method of combining user attributes-information tag interest degree vectors and user behavior-information tag interest degree vectors to obtain user-information tag interest degree vectors, so that When the user has not browsed the information, use the demographic information of the user to find the interest of the information label in the group to avoid the cold start problem, and when the information that the user browses gradually increases, the weight of the user behavior-information label interest degree vector Gradually increasing, because the user behavior-information tag interest degree vector is calculated by the deep learning model based on the recurrent neural network, using the time series features in the user's historical behavior data, so when the training samples are sufficient and the recurrent neural network parameters are adjusted appropriately In some cases, the effect is better than that of commonly used non-deep learning models.

User attribute-information tag interest degree vector can make up for the cold start problem when the user has not browsed the information. Simply using the deep learning model based on the recurrent neural network to estimate the interest degree of the information tag cannot be estimated when the user has not browsed the information. interest.

The present invention establishes and maintains a tagged candidate information base, avoids the problem of low information quality that often occurs when information is directly selected from the Internet, and reduces the amount of calculation for the user interest degree estimation problem due to the screening of information . Adding tags to the filtered information can be used for subsequent operations of estimating the user's interest in the information tags.

The present invention makes full use of the demographic information of the user, the historical behavior data of the user and the time sequence information of the information in the historical behavior data. By establishing a user attribute-information tag interest degree vector, the cold start problem of user interest degree estimation can be solved. By establishing a user behavior-information tag interest degree vector, the advantages of the deep learning model are used.

Establishing a candidate information base with tags avoids the problem of low information quality that often occurs when directly selecting information from the Internet, and reduces the amount of calculation for user interest estimation due to the screening of information.

The information in the candidate information base contains one or more information tags. Many other interest estimation methods only support one information tag per sample. In this method, the user behavior-information tag interest degree vector calculation is based on recursion The deep learning model of the neural network is calculated, and the input and output of the deep learning model can directly use vectors containing multi-label information, which is suitable for scenarios where each sample contains multiple labels.

The tanh hyperbolic tangent function used when calculating the user-information tag interest degree vector can convert the number [0, +∞) into a number between [0, 1), which just meets the user attribute-information tag interest in this method. The degree vector is combined with the user behavior-information tag interest degree vector, and the estimated value of each dimension of the calculated user-information tag interest degree vector is between [0,1] to reflect the user interest degree need.

Description of drawings

FIG. 1 is a flow chart of Embodiment 1 of a method for estimating the interestingness of information tags provided by the present invention.

Detailed ways

The technical solution of the present invention is described in detail below, it should be pointed out that the technical solution of the present invention is not limited to the implementation manner described in the examples, those skilled in the art refer to and learn from the content of the technical solution of the present invention, on the basis of the present invention The improvement and design carried out above shall belong to the protection scope of the present invention.

Embodiment one

Embodiment 1 of the present invention provides a method for estimating interest in information tags, the method includes steps S110-S150:

Step S110, creating and maintaining a tagged candidate information base.

In order to ensure the quality of the information pushed to users and keep the amount of algorithm calculation within a reasonable range, it is necessary to create a candidate information library with tags, filter the information on the network, select high-quality information and select from the preset information according to the content of the information. One or more tags that best match the content of the news are selected from the information tag library as the tags of the information, and the tagged information is added to the tagged candidate information library for subsequent operations.

The tag numbers in the default information tag library are T ₁ , T ₂ , . . . , T _m , where m is the total number of tags in the default information tag library. For reference, the preset information tag library during specific implementation can be [finance, sports, military, entertainment, life, education, health, technology, culture, travel, others], or set according to the actual situation, each information tag They need to be relatively independent, and the division of information labels should not be too detailed, so as to avoid affecting the accuracy of subsequent operation results due to too little information corresponding to each information label.

For each information in the candidate information base, each information is represented by an m-dimensional information vector according to the information tag corresponding to the information, where m is the total number of tags in the preset information tag library. When the information contains a tag T _j , the value of the jth dimension of the m-dimensional information vector is 1, otherwise the value of the jth dimension is 0. For example, some information contains tags T ₁ , T ₂ , and T ₅ , and the m-dimensional information vector corresponding to the information is [1, 1, 0, 0, 1, 0, 0, . . . , 0, 0].

Since the information has certain timeliness, it is necessary to regularly maintain the tagged candidate information base, add new information, and remove out-of-timeliness information.

Step S120. Obtain the user attribute-information tag interest degree vector according to the demographic information of the user.

User demographic information includes gender, age, region and other available information, so as to classify users into several groups. In order to avoid large errors due to the insufficient number of samples in each group due to too fine division of groups, the age of users can be divided into several gradients, for example: 20 years old and below, 21-30 years old, 31-40 years old, 41-50 years old Age, 51-60 years, 60 years and above. For the user's geographical information, when the number of samples is sufficient, it can be divided according to provinces. When the number of samples is small, the data of several provinces can be merged. For example, Heilongjiang, Jilin, and Liaoning can be merged into "Northeast Region".

According to user demographic information, users are divided into several groups, for example [male, 31-40 years old, Beijing] is a group, [female, 21-30 years old, Shanghai] is a group, and the groups are numbered G ₁ , G ₂ , . . . , G _n . The interest degree H _ij of the i-th group G _i to the j-th information tag T _j is:

The value of H _ij is between [0, 1], and the larger the value of H _ij is, the greater the interest of the i-th group G _i to the j-th information tag T _j is. For each group, an m-dimensional vector can be obtained, which is called the user attribute-information tag interest degree vector of the group, represented by V (user attribute, information tag). The value of the jth dimension in the user attribute-information tag interest degree vector V (user attribute, information tag) is the group's interest degree to the jth information tag T _j . For example, the user attribute-information tag interest degree vector of the i-th group is [H _i1 , H _i2 , . . . , H _im ], and all users in this group share the same user attribute-information tag interest degree vector.

In this method, user demographic information is used for group division. The demographic information used includes but is not limited to the user's gender, age, region, etc., and the method of group division can be determined according to specific circumstances. The division provided in this method The samples are for reference only.

Step S130, acquiring and preprocessing the historical behavior data of multiple users within a preset time period, inputting the deep learning model for training to obtain a trained deep learning model.

Obtain and preprocess the historical behavior data of multiple users within a preset time period. The preset time period T and the number of extracted users can be set according to the actual application situation. For example, if the time period T is set to three months and the number of extracted users is N _user , the random selection within three months will be extracted from all current data Historical behavior data of N _users . To protect user privacy and facilitate data processing, the N _user users are numbered as 1, 2, 3, . . . , N _user . For each user, obtain the information vectors corresponding to each information browsed in the historical behavior data, and arrange them in chronological order, and the information vectors corresponding to the information browsed first are arranged in front of the information vectors corresponding to the information browsed later.

The deep learning model adopted in this method is a recursive neural network model, and the recurrent neural network model can be an RNN model and its improved models, such as LSTM. For each parameter in the deep learning model, first randomly initialize, and then according to the user number in the preprocessed historical behavior data, for each user, input the m-dimensional information vector corresponding to the browsed information according to the order in which the user browsed the information The recurrent neural network model is trained in the recurrent neural network model.

For the recurrent neural network model, for each user, the kth input is the information vector corresponding to the kth information browsed by the user in the preprocessed behavior data. At this time, the model obtains a predicted output vector, which is The output vector is compared with the information vector corresponding to the k+1th information browsed by the user in the preprocessed behavior data, the deviation of the recursive neural network is calculated, and the parameters of the neural network model are continuously corrected according to the deviation. When the information vectors corresponding to all the information browsed by a user in the preprocessed historical behavior data are sequentially input into the deep learning model based on the recurrent neural network for training, the next user browsed in the preprocessed historical behavior data The information vectors corresponding to the information are input into the recursive neural network model being trained in chronological order, and the training continues until the information vectors corresponding to the information browsed by N _users in the preprocessed behavior data are all input into the deep learning model for training. . At this point, the trained deep learning model is obtained. Since the recurrent neural network model has been widely used in the field of deep learning, the specific construction of the recurrent neural network model will not be repeated in this method.

The trained deep learning model can predict the user's interest in information tags by inputting the preprocessed historical behavior data of a user into the model in order of time when browsing information. The specific details of the trained deep learning model The output is an m-dimensional prediction vector. The value of the j-th dimension in the m-dimensional prediction vector represents the user’s interest in the information tag T _j predicted by the trained deep learning model. The higher the value of the j-th dimension, the higher the user’s interest in The more likely the information tag T _j is interested.

Step S140: Acquire and preprocess the historical behavior data of the current user, and use the trained deep learning model to calculate the current user's user behavior-information tag interest degree vector.

When using the trained deep learning model to estimate the interest degree of information tags for the current user, the information vector corresponding to each information browsed in its historical behavior data is obtained, arranged in chronological order, and the information vector corresponding to the information browsed first Rank in front of the information vector corresponding to the information browsed later.

Input the information vectors corresponding to each information in the current user's historical behavior data into the trained deep learning model in chronological order. After all the information vectors corresponding to each information in the historical behavior data are input, the The m-dimensional prediction vector obtained by the deep learning model that has been trained is the current user's user behavior-information tag interest degree vector, represented by V (user behavior, information tag).

In this method, the deep learning model for calculating the user behavior-information tag interest degree vector includes a recurrent neural network and its improved model. The improved recursive neural network model may be the number of neurons of the network model, the number of layers of the network model, the addition of threshold functions, etc. If the recurrent neural network model after the improved structure does not directly affect the calculation of user behavior-information label interest vector way, it can be regarded as one of the realization methods of this method.

Step S150: Calculate the user-information tag interest degree vector according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector, and finally determine several information tags that the user is most interested in.

According to the calculated user attributes of the current user-information tag interest degree vector V (user attribute, information tag) and user behavior-information tag interest degree vector V (user behavior, information tag), the current user's interest in information tags can be calculated The m-dimensional interest degree vector of is denoted by V(user, information tag). Calculated as follows:

V (user, information label) = (1-w) * V (user attribute, information label) + w*V (user behavior, information label)

The w in the above formula represents the weight of V (user behavior, information tag) in calculating the current user's interest degree vector for the information tag. For new users, there is no historical behavior data, so w needs to satisfy that w is 0 when the current user has no historical behavior data, and as the user's historical behavior data gradually increases, V (user behavior, information label) can be more accurately Reflects user interest, so the weight of V (user behavior, information label) should gradually increase with the increase of user historical behavior data, and the value of w should always be in the range of [0,1]. Based on the above requirements, the calculation formula of w is as follows:

w=tanh(a*number of information browsed by the current user within the preset time period T)

Among them, tanh is the hyperbolic tangent function, a is a constant greater than 0, when the amount of information browsed by the current user within the preset time period T remains unchanged, the larger a is, the larger w is, and the value of a can be determined according to the actual situation. It is set according to the application situation. For reference, a can be set to 0.05.

The calculated value of the jth dimension in the current user's interest degree vector V (user, information label) of the information tag represents the final user's degree of interest in the tag T _j calculated by this method, and the value of the jth dimension The higher the value, the greater the possibility that the user is interested in the tag T _j . The calculated interest degree vector V(user, information label) of the current user on the information label completely reflects the current user's interest degree on all the labels in the information label library. Select the tags with the highest values in V (user, information tag) to push information.

For example, when the information tag library is [financial, sports, military, entertainment, life, education, health, technology, culture, travel, others], if the current user’s interest degree vector V(user, information tag ) is [0.42, 0.08, 0.02, 0.20, 0.01, 0.06, 0.33, 0.41, 0.05, 0.19, 0.14], from the user-information label interest degree vector, we can know that the current user’s interest in each information label estimated by this method is given by High to low are finance, technology, health, entertainment, travel, others, sports, education, culture, military, and life. If the information of the three tags that the current user is most interested in is preferentially selected for personalized push, the corresponding information tags of the pushed information are finance, technology, and health.

Further, the steps of creating and maintaining a tagged candidate information base include:

Select one or more tags that best match the content of the information from the preset information tag library as the tag of the information, and add the tagged information to the candidate information library containing tags; for each information in the candidate information library , according to the information tag corresponding to the information, each information is represented by an m-dimensional information vector, m is the total number of tags in the default information tag library; when the information contains a tag T _j , the j-th dimension of the m-dimensional information vector The value is 1, otherwise the value of the jth dimension is 0;

Regularly maintain the tagged candidate information base, add new information, and remove out-of-date information.

Further, the user demographic information includes, but is not limited to: one or more of the obtainable gender, age and/or geographical information to classify users into several groups.

Further, said obtaining the user attribute-information label interest degree vector according to the user demographic information includes:

The user attribute-information tag interest degree vector H _ij of the i-th group G _i to the j-th information tag T _j is:

The value of H _ij is between [0, 1].

Further, the acquisition of historical behavior data of multiple users within a preset time period and preprocessing, inputting the deep learning model for training to obtain a trained deep learning model, including:

Obtain the information vector corresponding to each information browsed in the historical behavior data of multiple users within the preset time period, and input the information vector into the recurrent neural network model in the order of browsing information time sequence to train the recurrent neural network model, and obtain A trained deep learning model.

Further, the acquisition of the historical behavior data of the current user and preprocessing, using the trained deep learning model to calculate the current user's user behavior-information label interest degree vector, including:

Obtain the information vector corresponding to each information browsed in the current user's historical behavior data, and arrange them in chronological order;

Input the information vectors corresponding to each information in the current user's historical behavior data into the trained deep learning model in chronological order. After all the information vectors corresponding to each information in the historical behavior data are input, the The m-dimensional prediction vector obtained by the trained deep learning model is the current user's user behavior-information tag interest degree vector.

Further, the user-information tag interest degree vector is calculated according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector, and finally determines several information tags that the user is most interested in, including:

According to the calculated user attribute-information tag interest degree vector and user behavior-information tag interest degree vector of the current user, the m-dimensional interest degree vector of the current user to the information tag can be calculated, and the calculation formula is as follows:

V (user, information label) = (1-w) * V (user attribute, information label) + w*V (user behavior, information label)

Among them, V (user, information label) is the m-dimensional interest degree vector of the current user to the information label; V (user behavior, information label) is the user behavior-information label interest degree vector; V (user attribute, information label) is the current The user's user attribute - information label interest degree vector; w represents the weight of V (user behavior, information label) in calculating the current user's interest degree vector for information label, and the value of w should always be in [0,1] within range.

Further, the calculation formula of w is as follows:

w=tanh(a*number of information browsed by the current user within the preset time period T)

Among them, tanh is the hyperbolic tangent function, and a is a constant greater than 0.

A method for estimating the interest degree of information tags provided by the present invention innovatively proposes a method of combining user attributes-information tag interest degree vectors and user behavior-information tag interest degree vectors to obtain user-information tag interest degree vectors, so that When the user has not browsed the information, use the demographic information of the user to find the interest of the information label in the group to avoid the cold start problem, and when the information that the user browses gradually increases, the weight of the user behavior-information label interest degree vector Gradually increasing, because the user behavior-information tag interest degree vector is calculated by the deep learning model based on the recurrent neural network, using the time series features in the user's historical behavior data, so when the training samples are sufficient and the recurrent neural network parameters are adjusted appropriately In some cases, the effect is better than that of commonly used non-deep learning models.

User attribute-information tag interest degree vector can make up for the cold start problem when the user has not browsed the information. Simply using the deep learning model based on the recurrent neural network to estimate the interest degree of the information tag cannot be estimated when the user has not browsed the information. interest.

The present invention establishes and maintains a tagged candidate information base, avoids the problem of low information quality that often occurs when information is directly selected from the Internet, and reduces the amount of calculation for the user interest degree estimation problem due to the screening of information . Adding tags to the filtered information can be used for subsequent operations of estimating the user's interest in the information tags.

The present invention makes full use of the demographic information of the user, the historical behavior data of the user and the time sequence information of the information in the historical behavior data. By establishing a user attribute-information tag interest degree vector, the cold start problem of user interest degree estimation can be solved. By establishing a user behavior-information tag interest degree vector, the advantages of the deep learning model are used.

Establishing a candidate information base with tags avoids the problem of low information quality that often occurs when directly selecting information from the Internet, and reduces the amount of calculation for user interest estimation due to the screening of information.

The information in the candidate information base contains one or more information tags. Many other interest estimation methods only support one information tag per sample. In this method, the user behavior-information tag interest degree vector calculation is based on recursion The deep learning model of the neural network is calculated, and the input and output of the deep learning model can directly use vectors containing multi-label information, which is suitable for scenarios where each sample contains multiple labels.

The tanh hyperbolic tangent function used when calculating the user-information tag interest degree vector can convert the number [0, +∞) into a number between [0, 1), which just meets the user attribute-information tag interest in this method. The degree vector is combined with the user behavior-information tag interest degree vector, and the estimated value of each dimension of the calculated user-information tag interest degree vector is between [0,1] to reflect the user interest degree need.

The embodiment of the present invention proposes a method for estimating the interest degree of information tags. In this method, information on the network is screened and tags are added to the screened information to obtain a tagged candidate information base. For the current users who need to estimate the interest degree of information tags, obtain the demographic characteristics of the current users and organize and analyze the data, and use the demographic characteristics of the current users to calculate the user attributes of the current users-information tag interest degree vector. This method trains a deep learning model based on recurrent neural network, obtains the historical behavior data of the current user, organizes and analyzes the data, and inputs the historical behavior data of the current user into the trained deep learning model. Obtain the current user's user behavior-information tag interest degree vector. According to the proposed method, the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector are used to calculate the user-information tag interest degree vector. The calculated user-information tag interest degree vector of the current user reflects the estimation result of the current user's interest degree to each information tag by this method.

Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be realized by hardware, or by software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present invention can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be CD-ROM, U disk, mobile hard disk, etc.), including several The instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute the methods described in various embodiments of the present invention.

The above disclosures are only a few specific embodiments of the present invention, but the present invention is not limited to the above embodiments, and any changes conceivable by those skilled in the art should fall within the protection scope of the present invention.

Claims (8)

1. A method for estimating interest in information tags, characterized in that the method comprises: Create and maintain a tagged candidate repository; According to the demographic information of the user, the user attribute-information tag interest degree vector is obtained; Obtain and preprocess the historical behavior data of multiple users within a preset time period, input the deep learning model for training to obtain the trained deep learning model; Obtain and preprocess the historical behavior data of the current user, and use the trained deep learning model to calculate the current user's user behavior-information tag interest degree vector; Calculate the user-information tag interest degree vector according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector, and finally determine several information tags that the user is most interested in. 2. The method according to claim 1, wherein the step of creating and maintaining a tagged candidate information base comprises: Select one or more tags that best match the content of the information from the preset information tag library as the tag of the information, and add the tagged information to the candidate information library containing tags; for each information in the candidate information library , according to the information tag corresponding to the information, each information is represented by an m-dimensional information vector, m is the total number of tags in the default information tag library; when the information contains a tag T _j , the j-th dimension of the m-dimensional information vector The value is 1, otherwise the value of the jth dimension is 0; Regularly maintain the tagged candidate information base, add new information, and remove out-of-date information. 3. The method according to claim 1, wherein the user demographic information includes but is not limited to: one or more of the available gender, age and/or geographical information to classify users into several groups Information. 4. The method according to claim 1 or 3, wherein said obtaining user attribute-information label interest degree vector according to user demographic information comprises: The user attribute-information tag interest degree vector H _ij of the i-th group G _i to the j-th information tag T _j is: The value of H _ij is between [0, 1]. 5. The method according to claim 1, wherein the acquisition of historical behavior data of multiple users within a preset period of time and preprocessing, inputting a deep learning model for training to obtain a trained deep learning model, includes : Obtain the information vector corresponding to each information browsed in the historical behavior data of multiple users within the preset time period, and input the information vector into the recurrent neural network model in the order of browsing information time sequence to train the recurrent neural network model, and obtain A trained deep learning model. 6. The method according to claim 1 or 5, wherein the historical behavior data of the current user is acquired and pre-processed, and the user behavior of the current user-information tag interest degree is calculated by using the trained deep learning model vector, including: Obtain the information vector corresponding to each information browsed in the current user's historical behavior data, and arrange them in chronological order; Input the information vectors corresponding to each information in the current user's historical behavior data into the trained deep learning model in chronological order. After all the information vectors corresponding to each information in the historical behavior data are input, the The m-dimensional prediction vector obtained by the trained deep learning model is the current user's user behavior-information tag interest degree vector. 7. The method according to any one of claims 1-6, wherein the user-information tag interest degree vector is calculated according to the current user's user attribute-information tag interest degree vector and user behavior-information tag interest degree vector Vector, and finally determine several information tags that users are most interested in, including: According to the calculated user attribute-information tag interest degree vector and user behavior-information tag interest degree vector of the current user, the m-dimensional interest degree vector of the current user to the information tag can be calculated, and the calculation formula is as follows: V (user, information label) = (1-w) * V (user attribute, information label) + w*V (user behavior, information label) Among them, V (user, information label) is the m-dimensional interest degree vector of the current user to the information label; V (user behavior, information label) is the user behavior-information label interest degree vector; V (user attribute, information label) is the current The user's user attribute-information tag interest degree vector; w represents the weight of V (user behavior, information tag) in calculating the current user's interest degree vector for information tags, and the value of w should always be in [0,1] within range. 8. the method for claim 7, is characterized in that, the computing formula of described w is as follows: w=tanh(a*number of information browsed by the current user within the preset time period T) Among them, tanh is the hyperbolic tangent function, and a is a constant greater than 0.