CN111400305A - Traceability and visualization method based on feature engineering blood relationship - Google Patents

Abstract

The invention discloses a traceable and visualized method based on a characteristic engineering blood relationship, which comprises the following steps: step one, establishing a blood relationship of database wide table integration, step two, establishing a data preprocessing blood relationship, step three, integrating a blood relationship, step four, establishing visual interaction, and establishing a traceable visual interaction interface in a multi-level manner based on the whole blood relationship. The invention solves the problems that the variable of the characteristic engineering result of the current industry is difficult to backtrack and the source of error data is difficult to search, and the blood-related relation solves the problems that the characteristic engineering experiment and production are disjointed and the variable processing is difficult to reproduce in the production environment, so that the whole process can be quickly reproduced based on the blood-related relation structure and can be quickly applied to the production environment.

Description

Traceability and visualization method based on feature engineering blood relationship

technical field

The present invention is in the technical field of data processing through artificial intelligence, and in particular, relates to a method of constructing a blood relationship of feature engineering by using a computer, and performing retrospective and visual interactive construction of features.

Background technique

In the development of big data and artificial intelligence, the requirements for model development speed and model effect are getting higher and higher. Correspondingly, the efficiency requirements for data collection, aggregation and processing are also higher and higher.

Data integration and processing are the most important factors that limit the efficiency of model development. Machine learning modeling is a process of continuously adjusting the parameters of the model and inputting data to adjust. Therefore, it is necessary to construct the blood relationship for the processing flow of data features, and to create a traceable and visualized blood relationship structure.

Constructing the blood relationship of features enables users to quickly adjust the input features of the model during the modeling process, reproduce the processing flow from source features to final features, provide users with a clear feature source path, and ultimately facilitate data processing. Cross-platform processing.

During the model launch process, the traceable blood relationship structure is helpful for users to trace back the erroneous data, locate the problem in time, and quickly correct it.

At present, there are few solutions on the market, and the existing solutions have the following problems:

1) The record of blood relationship is only for part of the process, and there is no record of the whole process, and it is impossible to reproduce the source to the result.

2) The construction of blood relationship takes the table as the dimension, and can only provide a simple display. Users cannot learn about a feature in detail according to the display content. .

3) The underlying parameters of feature processing methods such as normalization and outlier correction are not fully incorporated into the blood relationship system. Therefore, when performing feature engineering reproduction, the generated data set can only be used as a training set, not a test set.

Patent application 201610127589.0 discloses a method and device for determining a feature engineering strategy. The method obtains multiple eigenvalues of a preset dimension feature used for training a preset model; and determines multiple quantiles according to the ordering of the multiple eigenvalues interval; obtain the positive sample ratio of the number of eigenvalues as positive samples in each quantile interval and the number of all eigenvalues in the interval; calculate the sum of the positive sample proportions in any two adjacent quantile intervals According to the positive sample change rate between the positive sample ratios corresponding to all quantile intervals, the target feature engineering strategy for processing the preset dimension feature can be determined.

Another example is the patent application 201810669281.8 which discloses a method and system for constructing a machine learning modeling process. The method includes: displaying the constructed machine learning modeling process in a graphical interface for constructing the machine learning modeling process; and, in response to a user operation for running at least one step in the machine learning modeling process, running all the machine learning modeling procedures. at least one step; while running the at least one step, receiving a modification operation of the user for modifying the machine learning modeling process; in response to the modification operation, modifying the machine learning modeling process, Wherein, when running to the modified part of the machine learning modeling process, the running is based on the modified machine learning modeling process.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention provides a traceability and visualization method based on the blood relationship of feature engineering. The method can help users to trace the source of the result variables after the feature engineering is performed, and the process of the feature engineering can be completely visualized and interacted. show.

Another object of the present invention is to provide a traceable and visualized method based on the blood relationship of feature engineering, which can completely record the data processing flow based on the whole process, solve the problem that users are confused about the source of the final generated feature, and at the same time, The clear blood relationship structure with features as the dimension enables users to reproduce the logic through the blood relationship structure when selecting the result variables that are useful for modeling, and can develop models clearly and quickly, and launch them across platforms.

For achieving the above object, technical scheme of the present invention is as follows:

A traceability and visualization method based on feature engineering blood relationship, the method includes:

Step 1. Consanguinity construction of database wide table integration: The database wide table integration is based on mechanisms such as aggregation, association, extraction, and derivation, and according to predefined relationships between tables, multiple tables are integrated into wide tables for modeling. process. The aggregation is to use aggregation calculation for other feature columns according to the unique key of the data table; the association is to horizontally connect the table and the table with a common column; the extraction is the seat table and the right table, when one-to-many association is used. , extract a record in the right table and associate it with the left table; the derivation is to perform aggregation of characteristic columns or combination calculation between columns and columns according to business rules; the fields and tables involved in the above processing are based on the table to which the field belongs and the content of the operation. bloodline construction;

Further, in the process of wide table integration, according to the different ways of aggregation and derivation of business rules for each feature, each single feature is used as a dimension to record the aggregation and derivation rules of each variable and the upper-level features, and the output standard can be Backtracking data structure.

Step 2, data preprocessing bloodline construction: the data is the wide table data formed by the database output in step 1, and the data preprocessing is the processing of common feature engineering processing methods for the features of the wide table, including but not Limited to variable deletion, normalization, filling of missing values, outlier correction, one-hot encoding, standardization, multi-class binning and custom derivation, limited by different operation contents, the blood relationship in this step mainly involves the operation content and the underlying parameters of the operation ;

Further, the processing method of the feature engineering of this system is visual processing, that is, the processing process records the underlying parameters of each operation involved in turn according to the different processing methods and principles, according to a single feature dimension, and outputs a standard traceable data structure.

Step 3. Consanguinity integration: Since data processing is generally carried out in blocks, taking the above as an example, the two parts of the data structure are finally matched and summarized. The source variables of data preprocessing are the result variables of database aggregation and derivation. The result variable of data preprocessing is dimensional blood relationship structure data.

Step 4. Visual interaction construction: Based on the complete bloodline output in step 3, visual interaction construction is carried out at multiple levels and levels, including table level (what table the target field comes from, reflecting the relationship between the field and table), field level (target field From what field, reflecting the source relationship between fields), record level (what record the target record comes from, reflecting the record and the record), from beginning to end (from the source field to find the target field), from the end to the beginning (by the result field find the source field), the level selection is selected interactively by the user.

The invention solves the problems that the current industry feature engineering result variables are difficult to trace back and the source of the wrong data is difficult to find, and at the same time solves the problem that the feature engineering experiment and production are disconnected, and the variable processing is difficult to reproduce in the production environment and display problems, and finally the overall process can be based on The blood relationship structure can be quickly displayed and reproduced, and it can be quickly applied to the production environment.

The scheme steps realized by the present invention are as follows:

Step 1. The bloodline construction of database wide table integration; specifically, the source of data lies in each table of the database, and the construction of the table is atomic, so to form a wide table, it needs to be aggregated and associated with multiple tables, including:

101. When using aggregation and association, it is necessary to pre-define the association relationship between multiple tables, and define the business relationship between features; association relationships include but are not limited to internal associations, left associations, right associations and other methods. Business relationship types include but are not limited to transaction flow types, call record types, SMS record types, etc.;

102. Use database aggregation (the processing methods of database aggregation features mainly include: maximum value aggregation, minimum value aggregation, summation aggregation, count aggregation, standard deviation aggregation, mean aggregation, etc., taking the customer as the dimension, respectively representing a certain Summarize the data by means of the sum, non-null quantity, maximum value, minimum value, standard deviation, mean) of a certain characteristic of the customer;

103. The method of database extraction is adopted. Only when the left table is associated with the right table, the right table record is not unique compared to the associated field, and a certain right table record is extracted to be associated with the left table;

104. When adopting the method of database derivation, according to the interactive calculation rules between the preset features, the features are derived to form new features, and the new features have relevant business meanings and have practical industry significance;

When performing database aggregation and association, each feature is processed in different ways. According to the difference, the blood relationship of each feature is constructed separately. Each current feature is a feature of the upper-level table, which is obtained through aggregation operations. , so which is the upper-level table and which features are recorded, this is the blood relationship, and can be reproduced in other ways according to the recorded information (that is, the blood relationship). The blood relationship includes the upper-level table of all features, the upper-level feature, the aggregation type, the associated fields of the upper-level table, and the upper-level association table. The blood relationship includes but is not limited to the above contents. Environment visualization as the basis.

Step 2: Preprocessing of wide table feature data and construction of blood relationship, including:

105. The wide table generated in step 1 is the input of step 2, obtain the wide table processed by the database, and perform data splitting in the form of feature columns;

Splitting by feature column is mainly for parallel processing. The splitting rule is to ensure that the running time of each parallel task is consistent as much as possible, and at the same time to ensure that the performance of the server runs parallel tasks at the same time without downtime or memory overflow. In order to parallelize the system and speed up the processing process, the splitting rules are based on the principles of accelerating the processing speed and the average processing time.

Further, for each different feature column, the data preprocessing operations that need to be performed for the feature are respectively predefined. Specifically, the content of each operation is recorded by feature, and the preprocessing operation flow of each feature is constructed. Since the current operation flow is pre-defined, the data preprocessing operation is not started. Here, all the features are the original features of their respective features, so that the original features of each feature after preprocessing can be known.

106. Start feature preprocessing. During feature preprocessing, record the underlying parameters involved in data preprocessing operations in real time. When data preprocessing is completed, summarize feature processing operations and parameter streams to form the current feature flow for each feature. blood relationship structure.

In particular, in the process of data preprocessing, a single source feature may generate one or more intermediate features after a certain preprocessing. When recording the operations and parameter flow of these intermediate features, the operations and parameters of the single source feature should be inherited. Stream, record and build on this basis. For the result feature, its bloodline record should be complete, from source to result.

The third step is the integration of blood relationship. After each feature is processed in steps 1 and 2, two types of blood relationship will be generated. The difference is that due to the two different processing methods, the recorded blood relationship information and structure will be different. However, the wide table output from step 1 is the input of step 2, so there is a relationship between the original feature of step 2 and the result feature of step 1. According to this relationship, from the result feature of step 2, the original feature of step 2 can be found. Since the original feature belongs to the result feature of step 1, the original feature of step 1 can be found again. Specifically include:

107. Through the matching and splicing of the result features in the two steps, the blood relationship of each feature is finally successfully constructed, and the corresponding source features can be found for the result features of each step.

The blood relationship reflects the whole process of a result feature from a source feature of a certain table in the database to the final result feature, and it has considerable completeness. This structure can be used for the development of visual display and the reproduction of production feature processing.

Step 4. Feature blood relationship backtracking and visual interaction. Due to the completeness of the characteristic blood relationship and the rationality of the blood relationship structure, blood relationship can be quickly applied to development and visual interaction. The completeness and rationality of the blood relationship structure is reflected in a certain original feature of a certain table. According to the blood relationship structure, it can be separated from the database, and the result feature can be obtained from the original feature using the processing defined by the blood relationship structure.

108. By quickly reproducing the processing logic in any computer programming language, from the original features, the resulting features can be quickly generated, and backtracking can be performed, and then a multi-level architecture provides a visual interactive display.

The visual interactive display is mainly based on the complete blood relationship in step 3. The interactive display covers multiple levels and multiple orders, and the switching between levels and orders is determined by the interaction; the level includes: table level, field level, and record level; the visualization order includes Trace back to the source field from the result field, locate the result field from the source field.

Since blood relationship is dimensioned by each feature, if a user needs a certain feature, the blood relationship of the feature can be extracted separately to reproduce the feature. The design and storage of blood relationship follows the structure that can be parsed by most computer programming languages, which is convenient for parsing and production reproduction in various languages.

Compared with the prior art, the present invention adopts the blood relationship constructed by this scheme, which can completely record the data processing flow based on the whole process, solves the problem of user confusion about the source of the final generated feature, and facilitates the user to trace the source. At the same time, the visual interactive display enables users to quickly locate data problems, and the clear blood relationship structure with features as the dimension enables users to reproduce the feature processing logic through the visual interactive interface when selecting the result variables that are useful for modeling. The specific result variables required for processing and generation, and finally achieve a clear and rapid model development, and cross-platform launch.

Description of drawings

FIG. 1 is a schematic diagram of field blood relationship construction implemented by the present invention.

FIG. 2 is a basic architecture diagram of a bloodline visualization interactive interface implemented by the present invention.

FIG. 3 is a schematic diagram of feature engineering reproduction from original features to result features implemented by the invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

As shown in Figures 1-3, the solution of the present invention mainly includes four parts, the first part is that the data is from the original database, through the method of aggregation, the wide table is generated, and the blood relationship related to database aggregation is constructed; the second part is the first part. Part of the follow-up is to preprocess the feature data of the database output wide table, and construct the blood relationship related to the preprocessing; the third part is the integration of blood relationship. On the basis of completing the first part and the second part, integrate the blood relationship between the two. , forming a blood relationship structure with each feature as the dimension, which is reasonably visualized and traceable. The fourth part is the visual interactive interface display of the blood relationship and the rapid recurrence and retrospect of the characteristics based on the blood relationship. The processing flow is quickly replicated in the production environment.

Figure 1 shows the schematic diagram of field blood relationship construction. Figure 1 shows the three steps of field blood relationship construction. Figure 1 briefly shows various information including tables, table features, and feature processing types.

Since blood relationship is dimensioned by each feature, if a user needs a certain feature, the blood relationship of the feature can be extracted separately to reproduce the feature. The design and storage of blood relationship follows the structure that can be parsed by most computer programming languages, which is convenient for parsing and production reproduction in various languages.

Specifically, in conjunction with Fig. 2 and Fig. 3, the scheme steps realized by the present invention are as follows:

Step 1: Database aggregation and bloodline construction.

First of all, the source of data lies in each table of the database. The construction of the table is atomic. Therefore, to form a wide table, it needs to be aggregated and associated with multiple tables. When using aggregation and association, it is necessary to pre-define the association relationship between multiple tables, including but not limited to inner association, left association, right association, etc., as well as defining the business relationship between features (between features For example, a certain column represents the unit price of an item, and a certain column represents the customer's purchase quantity. Multiplying the customer can get the total price spent by the customer. And so on, summarize some similar business relationships, including but not Limited to transaction flow type, call record type, SMS record type, etc., business relationship is the general term for the combination of such features and features with business meaning), characteristic business relationship types include but are not limited to transaction flow type, call record type, SMS record types and other types, using database aggregation (the processing methods of database aggregation features mainly include: summation aggregation, count aggregation, maximum value aggregation, minimum value aggregation, standard deviation aggregation, mean aggregation, etc. Dimension, which represents the sum, non-empty quantity, maximum value, minimum value, standard deviation, mean value of a certain characteristic of a customer (can be total price or quantity or other), and summarizes the data. The summary calculation includes But not limited to sum, count, maximum value, minimum value, standard deviation, mean and other methods.

When performing database aggregation and association, each feature is processed in different ways. According to the difference, the blood relationship of each feature is constructed separately. Each current feature is a feature of the upper-level table, which is obtained through aggregation operations. , so which is the upper-level table and which is the characteristic must be recorded, and then can be reproduced in other ways according to the recorded information. For example: a customer's transaction records are multiple records (here can be understood as a table), after summation and aggregation, we know how many records the customer has in total, then the total number of records is a new feature, It is a new feature formed by summation aggregation over multiple records. The blood relationship includes all features (including intermediate processing features and final features: intermediate features may be processed from original features or upper-level intermediate features, and final features may be processed from intermediate features or original features). First-level features, aggregation type (aggregation type is actually the above-mentioned aggregation feature processing method), upper-level table associated fields and upper-level associated tables, blood relationship includes but not limited to the above content, all traceable by blood relationship (clearly understand based on how a certain final feature is obtained) and visualization in the production environment (according to the blood relationship, the process can be reproduced without the current development environment).

Step 2: Preprocessing of wide table feature data and construction of blood relationship.

From the perspective of the whole scheme, the wide table generated in step 1 is the pre-step of step 2. Step 2 cannot be performed without step 1. For example, step 1 produces a table about customers, each row represents a customer, and each row represents a customer. A column (feature) represents the basic information of the customer (age, gender, etc.), if the age of a customer in this table is missing (not in the database), then it needs to be filled (a kind of data preprocessing), will The age of all customers is averaged and filled with this value.

The wide table generated in step 1 is the input of step 2. Obtain the wide table processed by the database, and split the data in the form of feature columns. Splitting by feature column is mainly for parallel processing. The splitting rule is to ensure that the running time of each parallel task is consistent as much as possible, and at the same time to ensure that the performance of the server can run parallel tasks at the same time without downtime or memory overflow , in order to parallelize the system and speed up the processing process. The splitting rules are based on the principle of accelerating the processing speed and the average processing time.

Secondly, for each different feature column, pre-define the data preprocessing operations that need to be performed for the feature. Record the content of each operation by feature, and build a stream of preprocessing operations for each feature. As mentioned above, a feature can be filled, and other operations can be performed after filling, such as outlier removal, etc. A feature has its own operation flow. Since the current operation flow is pre-defined, the data preprocessing operation is not started. Here, all the features are the original features of their respective features, so that the original features of each feature after preprocessing can be known.

Begin feature preprocessing. During feature preprocessing, the underlying parameters involved in the data preprocessing operation are recorded in real time according to the data preprocessing operations performed by the feature (the underlying parameters are related to the actual data, so they cannot be defined in advance. record), when the data preprocessing is completed, the processing operations and parameter flow of the features are summarized to form the current blood relationship structure of each feature.

In particular, in the process of data preprocessing, a single source feature may generate one or more intermediate features after a certain preprocessing. When recording the operations and parameter flow of these intermediate features, the operations and parameters of the single source feature should be inherited. Stream, record and build on this basis. For the result feature, its bloodline record should be complete, from source to result.

The third step is the integration of blood relationship.

After each feature is processed in steps 1 and 2, two types of blood relationship will be generated. The difference is that due to the two different processing methods, the recorded blood relationship information and structure will be different.

However, the wide table output from step 1 is the input of step 2, so there is a relationship between the original feature of step 2 and the result feature of step 1. According to this relationship, from the result feature of step 2, the original feature of step 2 can be found. Since the original feature belongs to the result feature of step one, the original feature given to step one can also be found.

According to the above relationship, through the matching and splicing of the result features in the two steps, the blood relationship of each feature is finally successfully constructed, and the corresponding source feature can be found for the result feature of each step. The blood relationship reflects the whole process of a result feature from a source feature of a certain table in the database to the final result feature, and it has considerable completeness. This structure can be used for the development of visual display and the reproduction of production feature processing.

Step 4. Feature blood relationship backtracking and visual interaction.

Due to the completeness of the characteristic blood relationship and the rationality of the blood relationship structure, blood relationship can be quickly applied to development and visual interaction. The completeness and rationality of the blood relationship structure is reflected in a certain original feature of a certain table. According to the blood relationship structure, it can be separated from the database, and the result feature can be obtained from the original feature using the processing defined by the blood relationship structure. Therefore, by quickly reproducing the processing logic in any computer programming language, the resulting features can be quickly generated from the original features, and backtracking can be performed, and then a visual interactive display can be provided in a multi-level architecture.

In a word, the present invention adopts the blood relationship constructed by this scheme, which can completely record the data processing flow based on the whole process, solve the problem that users are confused about the source of the final generated feature, and facilitate the user to trace the source. At the same time, the clear blood relationship structure with features as the dimension enables users to reproduce the logic through the blood relationship structure when selecting the result variables that are useful for modeling, and only process and generate the specific result variables required, and finally achieve a clear and fast result. Model development and cross-platform launch.

The preferred embodiments of the present invention are listed above, but are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention. .

Claims (9)

1. a retrospective, visualization method based on feature engineering blood relationship, it is characterized in that described method comprises: Step 1. Consanguinity construction of database wide table integration: The database wide table integration is based on mechanisms such as aggregation, association, extraction, and derivation, and according to predefined relationships between tables, multiple tables are integrated into wide tables for modeling. process. The aggregation is to use aggregation calculation for other feature columns according to the unique key of the data table; the association is to horizontally connect the table and the table with a common column; the extraction is the left table and the right table, when a one-to-many association is used. , extract a record in the right table and associate it with the left table; the derivation is to perform aggregation of characteristic columns or combination calculation between columns and columns according to business rules; the fields and tables involved in the above processing are based on the table to which the field belongs and the content of the operation. Build bloodlines. Step 2, data preprocessing bloodline construction: the data is the wide table data formed by the database output in step 1, and the data preprocessing is the processing of common feature engineering processing methods for the features of the wide table, including but not Limited to variable deletion, normalization, filling of missing values, outlier correction, one-hot encoding, standardization, multi-class binning and custom derivation, limited by different operation contents, the blood relationship in this step mainly involves the operation content and the underlying parameters of the operation . Step 3. Consanguinity integration: Since data processing is generally carried out in blocks, taking the above as an example, the two parts of the data structure are finally matched and summarized. The source variables of data preprocessing are the result variables of database aggregation and derivation. The result variable of data preprocessing is dimensional blood relationship structure data. Step 4. Visual interaction construction: Based on the complete bloodline output in step 3, the visual interaction construction is carried out in multiple levels and in multiple orders, including table level, field level, record level, from beginning to end, from end to beginning, and the level selection is by user interaction choose. 2. The traceability and visualization method based on feature engineering blood relationship as claimed in claim 1, wherein in step 1, in the process of aggregation, association, extraction and derivation, according to the aggregation, association and derivative business of each feature Depending on the way of rules, each single feature is used as the dimension to record the aggregation, association and derivation rules of each variable, as well as the upper-level features, and output a standard traceable data structure. 3. the retrospective, visual method based on feature engineering blood relationship as claimed in claim 1, it is characterized in that in step 2, the processing mode of feature engineering is the processing of visualization, and the processing procedure is by single feature dimension, successively to the involved. Each operation and underlying parameters are recorded, and a standard traceable data structure is output. 4. the traceability, visualization method based on feature engineering blood relationship as claimed in claim 3, it is characterized in that the concrete steps of described method are as follows: Step 1. The bloodline construction of database wide table integration, including: 101. When using aggregation and association, it is necessary to pre-define the association relationship between multiple tables, and define the business relationship between features; the association relationship includes but is not limited to inner association, left association, right association, and characteristic business relationship type Including but not limited to transaction flow type, call record type, SMS record type; 102. Aggregate data by means of database aggregation. Aggregation calculations include but are not limited to maximum value aggregation, minimum value aggregation, summation aggregation, count aggregation, standard deviation aggregation, mean aggregation and other methods; 103. The method of database extraction is adopted, only when the left table is associated with the right table, and the records of the right table are not unique compared to the associated fields; 104. When the method of database derivation is adopted, according to the interactive calculation rules between the preset features, the features are derived to form new features; Step 2: Preprocessing of wide table feature data and construction of blood relationship, including: 105. Obtain the wide table processed by the database, and split the data in the form of characteristic columns; 106. Start feature preprocessing. During feature preprocessing, record the underlying parameters involved in data preprocessing operations in real time. When data preprocessing is completed, summarize feature processing operations and parameter streams to form the current feature flow for each feature. blood relationship structure. Step 3. Consanguinity integration, including: 107. Through the matching and splicing of the result features in the two steps, the blood relationship of each feature is finally successfully constructed, and the corresponding source features can be found for the result features of each step; Step 4. Visual interactive display of characteristic blood relationship, including: 108. Quickly reproduce processing logic through computer programming language, quickly generate resulting features from original features, and backtrack, and then provide visual interactive display in a multi-level architecture. 5. The traceability and visualization method based on feature engineering blood relationship as claimed in claim 4, characterized in that when performing database aggregation and association, each current feature is a certain feature of an upper-level table, and after aggregation This is the blood relationship. The blood relationship includes but is not limited to the previous-level table, previous-level feature, aggregation type, and previous-level table of all features. Associated fields and upper-level associated tables. 6. The retrospective and visualized method based on feature engineering blood relationship as claimed in claim 4, wherein in the step 105, for each different feature column, predefine the data predefinition that belongs to the feature. Processing operation; is to record the content of each operation by feature, and construct the preprocessing operation flow of each feature. 7. The traceability and visualization method based on feature engineering blood relationship as claimed in claim 4, characterized in that in the step 106, in the data preprocessing process, a single source feature may generate one or more after a certain preprocessing. For multiple intermediate features, when recording the operations and parameter flows of these intermediate features, they should inherit the operations and parameter flows of the single source feature, and record and construct on this basis. 8. The traceability and visualization method based on feature engineering blood relationship as claimed in claim 4, characterized in that in step 107, the blood relationship reflects a certain result feature from a certain source feature of a certain table in the database to the final The whole process of the result feature is quite complete, and the structure can be used for the development of visual display and the reproduction of production feature processing. 9. The traceability, visualization method based on feature engineering blood relationship as claimed in claim 4, it is characterized in that in described 108 steps, carry out visual interactive display according to the complete blood relationship in step 3 as the main, interactive display covers many. Hierarchy and multi-sequence, the switching of hierarchy and order is determined by interaction; the hierarchy includes: table level, field level, record level; visualization sequence includes tracing back to the source field from the result field, and locating the result field from the source field.

Images