A book intelligent recommendation method and system based on user portrait

Abstract

The present invention relates to the field of intelligent recommendation technology, and in particular to a method and system for intelligent book recommendation based on user portraits. A user portrait-based intelligent book recommendation system comprises: a life cycle classification module and a recommendation strategy setting module. The present invention constructs a multi-strategy recommendation architecture based on the division of book life cycle states, combines user portrait exploration willingness, group behavior similarity, social influence characteristics and external hot event perception, and sets matching recommendation strategies in the cold start period, growth period, outbreak period, decline period and sedimentation period respectively, to achieve dynamic switching and precise delivery of book recommendation strategies, so that the recommendation system can maintain high recommendation efficiency and content exposure value in multiple scenarios such as user interest transfer, book popularity evolution and platform content changes, thereby significantly improving the book life cycle management capabilities and the overall user satisfaction of the platform.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation technology, and in particular to a user portrait-based intelligent book recommendation method and system. Background Art

[0002] Among existing book recommendation technologies, mainstream methods mostly use collaborative filtering, content matching, deep learning, and other means. These methods primarily start from the user's perspective and make recommendations based on their historical behavior, interests, and similar user profiles. These methods have achieved certain success in alleviating cold start issues and capturing interest, but they generally lack awareness of the life cycle characteristics of the book itself. Current recommendation systems often treat all books as equivalent candidates, ignoring the natural evolution of books from the cold start period, growth period, explosive period, decline period, and sedimentation period. They fail to dynamically adjust recommendation strategies based on the stage of the book's life cycle, leading to insufficient exposure of cold books, insufficient diffusion of popular books, and the ineffectiveness of long-tail content. This, in turn, affects the balance of the content ecosystem and the continuous optimization of the user recommendation experience. Summary of the Invention

[0003] The present invention constructs a multi-strategy recommendation architecture based on the division of book life cycle states, combines user portrait exploration willingness, group behavior similarity, social influence characteristics and external hot event perception, and sets matching recommendation strategies in the cold start period, growth period, outbreak period, decline period and sedimentation period respectively, to achieve dynamic switching and precise delivery of book recommendation strategies, so that the recommendation system can maintain high recommendation efficiency and content exposure value in multiple scenarios such as user interest transfer, book popularity evolution and platform content changes, thereby significantly improving the book life cycle management capabilities and the overall user satisfaction of the platform.

[0004] The present invention provides a user portrait-based intelligent book recommendation method, comprising:

[0005] At the current monitoring time point, the book behavior feature data corresponding to all books in the time window are collected, and the book behavior feature data corresponding to the books are sent to the book life cycle classification model for processing, and the life cycle classification labels corresponding to the books are output. The life cycle classification labels include cold start period, growth period, outbreak period, decline period and sedimentation period. Based on the life cycle classification labels corresponding to the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library;

[0006] For all books in the cold start recommendation library, set a book recommendation strategy based on the user's exploration intention;

[0007] For all books in the growth recommendation library, set book recommendation strategies based on similar users;

[0008] For all books in the explosive recommendation library, set book recommendation strategies based on user influence;

[0009] For all books in the Decline Recommendation Library, set the same recommendation strategy as that for the Growth Recommendation Library.

[0010] For all books in the sedimentation recommendation library, a recommendation strategy is built based on hot information.

[0011] As a preferred aspect, for all books in the cold start recommendation library, a book recommendation strategy is set based on the user's exploration intention, specifically:

[0012] The books to be recommended are referred to as cold start recommended books, and the book feature vectors corresponding to the cold start recommended books are extracted;

[0013] Traverse all users and obtain the selected user's exploration score for the selected user. The exploration score represents the user's exploration willingness. Obtain the book feature vector corresponding to the book that the selected user last purchased, recorded as the interest book feature vector. Calculate the similarity between the interest book feature vector and the book feature vector corresponding to the cold start recommended book, recorded as the first user compatibility score. Then add the selected user's exploration score and the first user compatibility score to obtain the cold start recommendation score. Arrange all users from largest to smallest according to the cold start recommendation score to form a cold start recommendation set, and recommend the cold start recommended book to the top N users in the cold start recommendation set.

[0014] The user's exploration score is calculated as follows: obtain all books purchased by the user in the past, record them as historically purchased books, and obtain the book feature vectors corresponding to all historically purchased books. Send the book feature vectors corresponding to the historically purchased books into the style label analysis network for processing to obtain the style labels corresponding to the historically purchased books. The reciprocal of the total number of evaluations corresponding to all historically purchased books, the variance corresponding to the historically purchased book feature vectors corresponding to all historically purchased books, and the distribution entropy of all style labels are added together to obtain the user's exploration score.

[0015] As a preferred aspect, for all books in the growth recommendation library, a book recommendation strategy is set based on similar users, specifically:

[0016] The books to be recommended are designated as growth-recommended books. The corresponding book feature vectors are obtained. Users who purchased growth-recommended books within the time window are grouped into a to-be-recommended set. For each user in the to-be-recommended set, a user profile feature vector is constructed. Cluster analysis is performed on all user profile feature vectors to obtain several clusters. The cluster center is obtained for each cluster.

[0017] Traverse all users except those in the set to be recommended, and for the selected user, obtain the user portrait feature vector corresponding to the selected user. Calculate the similarity between the user portrait feature vector corresponding to the selected user and each cluster center, and obtain the simulated recommendation score corresponding to the selected user. Then use the largest simulated recommendation score as the growth recommendation score corresponding to the selected user. Arrange all users in the set to be recommended from largest to smallest according to their growth recommendation scores to form a growth recommendation set, and recommend the growth recommendation books to the top N users in the growth recommendation set.

[0018] The specific method of constructing the user portrait feature vector is: forming the user portrait feature vector based on the user's basic information and the book feature vector corresponding to the book that the user last purchased.

[0019] As an optimal aspect, for all books in the explosive recommendation library, a book recommendation strategy is set based on user influence, specifically:

[0020] The book to be recommended is called an explosive recommended book. All users are traversed to obtain the interest book feature vector corresponding to the selected user. The similarity between the interest book feature vector and the book feature vector corresponding to the explosive recommended book is calculated and recorded as the second user compatibility score. The second user compatibility score is then added to the user's influence value to obtain the explosive recommendation score. All users are arranged from large to small according to the explosive recommendation score to form a explosive recommendation set, and the explosive recommended books are recommended to the top N users in the explosive recommendation set.

[0021] As an optimal aspect, for all books in the sedimentation recommendation library, a recommendation strategy is constructed based on hot information, specifically:

[0022] Continuously monitor hotspot information and obtain the hotspot information feature vector corresponding to the hotspot information;

[0023] The books in the sedimentation recommendation library are referred to as sedimentation recommended books. All sedimentation recommended books in the sedimentation recommendation library are traversed, and the similarity between the hot information feature vector and each sedimentation recommended book in the sedimentation recommendation library is calculated respectively, and recorded as the sedimentation recommendation score corresponding to the sedimentation recommended book. The sedimentation recommended book corresponding to the largest sedimentation recommendation score is recommended to all users.

[0024] As a preferred aspect, the book life cycle classification model is established based on the random forest model, and the training of the book life cycle classification model specifically includes the following steps:

[0025] Obtain several book life cycle classification training samples, which include book behavior feature data corresponding to any book. Label the book life cycle classification training samples with life cycle classification labels, and form a book life cycle classification training set with all labeled book life cycle classification training samples. Train the book life cycle classification model with the book life cycle classification training set.

[0026] As a preferred aspect, the style label analysis network is established based on the BP neural network, and training the style label analysis network specifically includes the following steps:

[0027] Obtain several style label analysis training samples, each of which includes a book feature vector corresponding to the book. Label the style label analysis training samples using style labels, form a style label analysis training set with all labeled style label analysis training samples, and train the style label analysis network using the style label analysis training set.

[0028] The present invention also provides a book intelligent recommendation system based on user portraits, comprising:

[0029] The life cycle classification module is used to collect the book behavior feature data corresponding to all books within the time window at the current monitoring time point, send the book behavior feature data corresponding to the books into the book life cycle classification model for processing, and output the life cycle classification label corresponding to the books. The life cycle classification label includes cold start period, growth period, outbreak period, decline period and sedimentation period. Based on the life cycle classification label corresponding to the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library;

[0030] The recommendation strategy setting module is used to set book recommendation strategies based on the user's exploration intention for all books in the cold start recommendation library, set book recommendation strategies based on similar users for all books in the growth recommendation library, set book recommendation strategies based on user influence for all books in the outbreak recommendation library, set recommendation strategies consistent with the growth recommendation library for all books in the decline recommendation library, and build recommendation strategies based on hot information for all books in the sedimentation recommendation library.

[0031] The present invention has the following advantages:

[0032] The present invention constructs a multi-strategy recommendation architecture based on the division of book life cycle states, combines user portrait exploration willingness, group behavior similarity, social influence characteristics and external hot event perception, and sets matching recommendation strategies in the cold start period, growth period, outbreak period, decline period and sedimentation period respectively, to achieve dynamic switching and precise delivery of book recommendation strategies, so that the recommendation system can maintain high recommendation efficiency and content exposure value in multiple scenarios such as user interest transfer, book popularity evolution and platform content changes, thereby significantly improving the book life cycle management capabilities and the overall user satisfaction of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a structural diagram of the user portrait-based intelligent book recommendation system adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0035] Example 1, a user portrait-based intelligent book recommendation method, comprising:

[0036] At the current monitoring time point, the book behavior feature data corresponding to all books in the time window are collected. The time window here generally refers to the past 7 days. The book behavior feature data includes the total sales, views, evaluations, sales growth rate, view growth rate and evaluation growth rate, etc., among which the sales growth rate is the growth rate of the total sales in the current time window compared with the total sales in the past 7 days, the view growth rate is the growth rate of the views in the current time window compared with the views in the past 7 days, the views are the total number of times users click on the book details page on the e-commerce platform, the evaluation growth rate is the growth rate of the evaluations in the current time window compared with the evaluations in the past 7 days, and the evaluations are the number of comments made by users on the book on the e-commerce platform. The book behavior feature data corresponding to the book is sent to the book life cycle classification model for processing, and the output is the book corresponding Life cycle classification labels, which include cold start period, growth period, outbreak period, decline period and sedimentation period. The cold start period is when the book is just put on the shelf or there are few long-term reviews, the book behavior data is relatively sparse and there is less feedback. The growth period is the period when the book behavior feature data of the book is rising but not higher than the threshold set in advance. The outbreak period is the period when the book behavior feature data of the book is higher than the threshold set in advance. The decline period is the period when the book behavior feature data of the book is declining. The sedimentation period is the period when the book behavior feature data of the book is stable. Based on the corresponding life cycle classification labels of the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library. It should be noted that the time interval between adjacent monitoring time points here is generally set to 1 day.

[0037] For all books in the cold start recommendation library, set a book recommendation strategy based on the user's exploration intention, specifically:

[0038] The book to be recommended is called a cold-start recommended book, and the book feature vector corresponding to the cold-start recommended book is extracted. The specific method of extracting the book feature vector here is: using the TF-IDF method to extract keywords from the book title, book subtitle, book introduction, chapter preview, author introduction and comments, and then performing a word embedding operation on the keywords to construct the first keyword vector. Then, a weighted sum and average operation is performed on all the first keyword vectors and the corresponding TF-IDF values ​​to obtain the book feature vector;

[0039] Traverse all users, here targeting users registered on the e-commerce platform, and obtain the selected user's exploration score for the selected user. The exploration score represents the user's exploration willingness. Obtain the book feature vector corresponding to the book that the selected user last purchased, recorded as the interest book feature vector, calculate the similarity between the interest book feature vector and the book feature vector corresponding to the cold start recommended book, recorded as the first user compatibility score, and then add the selected user's exploration score and the first user compatibility score to obtain the cold start recommendation score. Arrange all users from large to small according to the cold start recommendation score to form a cold start recommendation set, and recommend the cold start recommended books to the top N users in the cold start recommendation set; whenever a user makes a purchase, the interest book feature vector will be reconstructed;

[0040] It should be noted that the cold start recommended books are recommended to the first N users in the cold start recommendation set. The specific implementation method is to set a recommendation area on the user's home page interface and display the recommended books in the recommendation area.

[0041] The user's exploration score is calculated as follows: obtain all books purchased by the user in the past, record them as historically purchased books, and obtain the book feature vectors corresponding to all historically purchased books. Send the book feature vectors corresponding to the historically purchased books into the style label analysis network for processing to obtain the style labels corresponding to the historically purchased books. Style labels generally include science fiction, suspense, and history. The reciprocal of the total number of evaluations corresponding to all historically purchased books, the variance corresponding to the historically purchased book feature vectors corresponding to all historically purchased books, and the distribution entropy of all style labels are added together to obtain the user's exploration score. It should be noted that when calculating the variance corresponding to the historically purchased book feature vectors corresponding to all historically purchased books, first sum and average the historically purchased book feature vectors corresponding to all historically purchased books to obtain the average historically purchased book feature vector F. The variance D corresponding to the historically purchased book feature vectors corresponding to all historically purchased books is calculated by the following formula: -1 ∑ i [sim(F,H i )] 2 , where sim() is similarity calculation, H i is the historical book purchase feature vector corresponding to the i-th historical book purchase, i=1, 2, 3, ..., M, M is the total number of all historical books purchased by the user; when calculating the distribution entropy of all style labels, it is achieved through the following formula: , where P is the distribution entropy of all style labels, p jis the frequency of occurrence of the jth style tag, j=1, 2, 3, …, J, J is the total number of style tags; since historical purchase data and reviews of books in the cold start period are small, users will ignore the recommendation information due to lack of trust if these historical purchase books are routinely recommended, resulting in poor recommendation effect. Therefore, we need to find users who dare to try new content, that is, users with high exploration willingness. These users are also more likely to buy books with little historical purchase data and few reviews. Correspondingly, recommending books in the cold start period to these users with high exploration willingness can achieve better recommendation effects. Once a purchase occurs, the historical purchase data and the number of reviews can also be increased.

[0042] For all books in the growth recommendation library, set a book recommendation strategy based on similar users, specifically:

[0043] Since the books in the growth recommendation library already have some purchase records and evaluation numbers, it is possible to find users similar to the users recorded in the purchase records and make recommendations to improve the conversion rate and diffusion effect. Therefore, the books to be recommended are called growth recommendation books, and the book feature vectors corresponding to the growth recommendation books are obtained. The users who have purchased the growth recommendation books within the time window are formed into a to-be-recommended set. It should be noted that in the specific implementation process, the to-be-recommended set stores the serial numbers corresponding to the users. For each user in the to-be-recommended set, a user portrait feature vector is constructed, and cluster analysis is performed on all user portrait feature vectors. Specifically, the k-means algorithm can be used to obtain several clusters, and the cluster center is obtained for each cluster. It should be noted that in the process of executing the k-means algorithm, the similarity between two user portrait feature vectors is used as the distance between the two user portrait feature vectors, and the cluster center is the average value of the entire cluster.

[0044] Traverse all users except those in the set to be recommended, and for the selected user, obtain the user portrait feature vector corresponding to the selected user, calculate the similarity between the user portrait feature vector corresponding to the selected user and the center of each cluster, and obtain the simulated recommendation score corresponding to the selected user. The similarity calculation generally adopts the cosine similarity calculation method, and then use the largest simulated recommendation score as the growth recommendation score corresponding to the selected user. Arrange all users in the set to be recommended from large to small according to the growth recommendation score to form a growth recommendation set, and recommend growth recommendation books to the top N users in the growth recommendation set.

[0045] The specific method of constructing the user portrait feature vector is: based on the user's basic information and the book feature vector corresponding to the book that the user last purchased, the user portrait feature vector is composed. The basic information here includes age, gender, and region; every time the user makes a purchase, the user portrait feature vector will be reconstructed.

[0046] For all books in the explosive recommendation library, set a book recommendation strategy based on user influence, specifically:

[0047] The books in the burst recommendation library already have enough purchase records and evaluation numbers, and have sufficient influence. The recommendation strategy at this time should be to continue to expand the recommendation effect, and users with sufficient influence are good recommendation targets, which can further improve the recommendation effect with the help of the user's influence. Therefore, the book to be recommended is called a burst recommendation book, and all users are traversed to obtain the interest book feature vector corresponding to the selected user. The similarity between the interest book feature vector and the book feature vector corresponding to the burst recommendation book is calculated, and recorded as the second user compatibility score. The second user compatibility score is then added to the user's influence value to obtain the burst recommendation score. The user's influence value here is generally determined by the total number of all accounts associated with the user. All users are arranged from large to small according to the burst recommendation score to form a burst recommendation set, and the burst recommended books are recommended to the top N users in the burst recommendation set. It should be noted that when the recommendation is actually performed for all books in the burst recommendation library, reward measures can be set for the users who perform the recommendation.

[0048] For all books in the declining recommendation library, set a recommendation strategy consistent with that of the growing recommendation library. Since the purchase data of books in the declining recommendation library shows a downward trend, it means that most users who have the desire to buy have already executed the purchase behavior. At this time, users similar to those recorded in the purchase records can be returned for recommendation.

[0049] For all books in the sedimentation recommendation library, a recommendation strategy is constructed based on hot information, specifically:

[0050] Continuously monitor hot information, which can be hot search content, highly praised articles, or descriptions of popular TV series, and obtain the corresponding hot information feature vector. Specifically, the hot information feature vector is obtained by extracting keywords from the hot information using the TF-IDF method, then embedding the keywords to construct a second keyword vector, and then performing a weighted sum and average operation on all the second keyword vectors and the corresponding TF-IDF values ​​to obtain the hot information feature vector.

[0051] The books in the sedimentation recommendation library are referred to as sedimentation recommended books. All sedimentation recommended books in the sedimentation recommendation library are traversed, and the similarity between the hot information feature vector and each sedimentation recommended book in the sedimentation recommendation library is calculated respectively, and recorded as the sedimentation recommendation score corresponding to the sedimentation recommended book. The sedimentation recommended book corresponding to the largest sedimentation recommendation score is recommended to all users. It should be noted that the recommendation conversion rate of books in the sedimentation recommendation library is low. Therefore, recommending similar books through the latest hot information can significantly improve the recommendation conversion rate of books in the sedimentation recommendation library.

[0052] The book life cycle classification model is built based on the random forest model. Training the book life cycle classification model includes the following steps:

[0053] Obtain several book life cycle classification training samples. The book life cycle classification training samples include the book behavior feature data corresponding to any book. The book behavior feature data here is obtained based on actual monitoring. The book life cycle classification training samples are labeled with life cycle classification labels. The labeled life cycle classification labels here are manually implemented by the operator. All labeled book life cycle classification training samples are combined into a book life cycle classification training set. The book life cycle classification model is trained with the book life cycle classification training set. The training target is the labeled life cycle classification labels. It is judged whether the accuracy of the book life cycle classification model meets expectations. If the accuracy of the book life cycle classification model meets expectations, the trained book life cycle classification model is output. Otherwise, the book life cycle classification model is continued to be trained with the book life cycle classification training set.

[0054] The style label analysis network is built based on the BP neural network. The training of the style label analysis network includes the following steps:

[0055] Obtain several style label analysis training samples, which include book feature vectors corresponding to books. The book feature vectors here are obtained by the operator based on actual books. The style label analysis training samples are labeled with style labels. The labeled style labels here are manually divided by the operator based on actual reading. All labeled style label analysis training samples are combined into a style label analysis training set. The style label analysis network is trained with the style label analysis training set. The training target is the labeled style labels. It is judged whether the accuracy of the style label analysis network meets expectations. If the accuracy of the style label analysis network meets expectations, the trained style label analysis network is output. Otherwise, the style label analysis network continues to be trained with the style label analysis training set.

[0056] This application constructs a multi-strategy recommendation architecture based on the division of book life cycle status, combines user portrait exploration willingness, group behavior similarity, social influence characteristics and external hot event perception, and sets matching recommendation strategies in the cold start period, growth period, outbreak period, decline period and sedimentation period respectively, to achieve dynamic switching and precise delivery of book recommendation strategies, so that the recommendation system can maintain high recommendation efficiency and content exposure value in multiple scenarios such as user interest transfer, book popularity evolution and platform content changes, thereby significantly improving the book life cycle management capabilities and the overall user satisfaction of the platform.

[0057] Example 2, a book intelligent recommendation system based on user portrait, see Figure 1 ,include:

[0058] The life cycle classification module is used to collect the book behavior feature data corresponding to all books in the time window at the current monitoring time point. The time window here generally refers to the past 7 days. The book behavior feature data includes the total sales, views, evaluations, sales growth rate, view growth rate and evaluation growth rate, etc., among which the sales growth rate is the growth rate of the total sales in the current time window compared with the total sales in the past 7 days, the view growth rate is the growth rate of the views in the current time window compared with the views in the past 7 days, the views are the total number of times users click on the book details page on the e-commerce platform, the evaluation growth rate is the growth rate of the evaluations in the current time window compared with the evaluations in the past 7 days, and the evaluations are the number of comments made by users on the book on the e-commerce platform. The book behavior feature data corresponding to the book are sent to the book life cycle classification model for processing and input The life cycle classification labels corresponding to the books are released. The life cycle classification labels include cold start period, growth period, outbreak period, decline period and sedimentation period. The cold start period is when the book is just put on the shelf or there are few long-term reviews. The book behavior data is relatively sparse and there is less feedback. The growth period is the period when the book behavior feature data of the book is rising but not higher than the threshold set in advance. The outbreak period is the period when the book behavior feature data of the book is higher than the threshold set in advance. The decline period is the period when the book behavior feature data of the book is declining. The sedimentation period is the period when the book behavior feature data of the book is stable. Based on the life cycle classification labels corresponding to the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library. It should be noted that the time interval between adjacent monitoring time points here is generally set to 1 day.

[0059] The recommendation strategy setting module is used to set book recommendation strategies based on the user's exploration intention for all books in the cold start recommendation library, set book recommendation strategies based on similar users for all books in the growth recommendation library, set book recommendation strategies based on user influence for all books in the outbreak recommendation library, set recommendation strategies consistent with the growth recommendation library for all books in the decline recommendation library, and build recommendation strategies based on hot information for all books in the sedimentation recommendation library.

[0060] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A book intelligent recommendation method based on user portrait, characterized in that: include: At the current monitoring time point, the book behavior feature data corresponding to all books in the time window are collected, and the book behavior feature data corresponding to the books are sent to the book life cycle classification model for processing, and the life cycle classification labels corresponding to the books are output. The life cycle classification labels include cold start period, growth period, outbreak period, decline period and sedimentation period. Based on the life cycle classification labels corresponding to the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library; For all books in the cold start recommendation library, set a book recommendation strategy based on the user's exploration intention; For all books in the growth recommendation library, set book recommendation strategies based on similar users; For all books in the explosive recommendation library, set book recommendation strategies based on user influence; For all books in the Decline Recommendation Library, set the same recommendation strategy as that for the Growth Recommendation Library. For all books in the sedimentation recommendation library, build a recommendation strategy based on hot information; For all books in the cold start recommendation library, set a book recommendation strategy based on the user's exploration intention, specifically: The books to be recommended are referred to as cold start recommended books, and the book feature vectors corresponding to the cold start recommended books are extracted; Traverse all users and obtain the selected user's exploration score for the selected user. The exploration score represents the user's exploration willingness. Obtain the book feature vector corresponding to the book that the selected user last purchased, recorded as the interest book feature vector. Calculate the similarity between the interest book feature vector and the book feature vector corresponding to the cold start recommended book, recorded as the first user compatibility score. Then add the selected user's exploration score and the first user compatibility score to obtain the cold start recommendation score. Arrange all users from largest to smallest according to the cold start recommendation score to form a cold start recommendation set, and recommend the cold start recommended book to the top N users in the cold start recommendation set. The user's exploration score is calculated as follows: obtain all books purchased by the user in the past, record them as historically purchased books, and obtain the book feature vectors corresponding to all historically purchased books. Send the book feature vectors corresponding to the historically purchased books into the style label analysis network for processing to obtain the style labels corresponding to the historically purchased books. The reciprocal of the total number of evaluations corresponding to all historically purchased books, the variance corresponding to the historically purchased book feature vectors corresponding to all historically purchased books, and the distribution entropy of all style labels are added together to obtain the user's exploration score.

2. The intelligent book recommendation method based on user portrait according to claim 1, characterized in that: For all books in the growth recommendation library, set a book recommendation strategy based on similar users, specifically: The books to be recommended are designated as growth-recommended books. The corresponding book feature vectors are obtained. Users who purchased growth-recommended books within the time window are grouped into a to-be-recommended set. For each user in the to-be-recommended set, a user profile feature vector is constructed. Cluster analysis is performed on all user profile feature vectors to obtain several clusters. The cluster center is obtained for each cluster. Traverse all users except those in the set to be recommended, and for the selected user, obtain the user portrait feature vector corresponding to the selected user. Calculate the similarity between the user portrait feature vector corresponding to the selected user and each cluster center, and obtain the simulated recommendation score corresponding to the selected user. Then use the largest simulated recommendation score as the growth recommendation score corresponding to the selected user. Arrange all users in the set to be recommended from largest to smallest according to their growth recommendation scores to form a growth recommendation set, and recommend the growth recommendation books to the top N users in the growth recommendation set. The specific method of constructing the user portrait feature vector is: forming the user portrait feature vector based on the user's basic information and the book feature vector corresponding to the book that the user last purchased.

3. The intelligent book recommendation method based on user portrait according to claim 2 is characterized in that: For all books in the explosive recommendation library, set a book recommendation strategy based on user influence, specifically: The book to be recommended is called an explosive recommended book. All users are traversed to obtain the interest book feature vector corresponding to the selected user. The similarity between the interest book feature vector and the book feature vector corresponding to the explosive recommended book is calculated and recorded as the second user compatibility score. The second user compatibility score is then added to the user's influence value to obtain the explosive recommendation score. All users are arranged from large to small according to the explosive recommendation score to form a explosive recommendation set, and the explosive recommended books are recommended to the top N users in the explosive recommendation set.

4. The intelligent book recommendation method based on user portrait according to claim 3 is characterized in that: For all books in the sedimentation recommendation library, a recommendation strategy is constructed based on hot information, specifically: Continuously monitor hotspot information and obtain the hotspot information feature vector corresponding to the hotspot information; The books in the sedimentation recommendation library are referred to as sedimentation recommended books. All sedimentation recommended books in the sedimentation recommendation library are traversed, and the similarity between the hot information feature vector and each sedimentation recommended book in the sedimentation recommendation library is calculated respectively, and recorded as the sedimentation recommendation score corresponding to the sedimentation recommended book. The sedimentation recommended book corresponding to the largest sedimentation recommendation score is recommended to all users.

5. The intelligent book recommendation method based on user portrait according to claim 4 is characterized in that: The book life cycle classification model is built based on the random forest model. Training the book life cycle classification model includes the following steps: Obtain several book life cycle classification training samples, which include book behavior feature data corresponding to any book. Label the book life cycle classification training samples with life cycle classification labels, and form a book life cycle classification training set with all labeled book life cycle classification training samples. Train the book life cycle classification model with the book life cycle classification training set.

6. The intelligent book recommendation method based on user portrait according to claim 5, characterized in that: The style label analysis network is built based on the BP neural network. The training of the style label analysis network includes the following steps: Obtain several style label analysis training samples, each of which includes a book feature vector corresponding to the book. Label the style label analysis training samples using style labels, form a style label analysis training set with all labeled style label analysis training samples, and train the style label analysis network using the style label analysis training set.

7. A book intelligent recommendation system based on user portrait, characterized in that: The system applies the user portrait-based intelligent book recommendation method according to any one of claims 1 to 6, including: The life cycle classification module is used to collect the book behavior feature data corresponding to all books within the time window at the current monitoring time point, send the book behavior feature data corresponding to the books into the book life cycle classification model for processing, and output the life cycle classification label corresponding to the books. The life cycle classification label includes cold start period, growth period, outbreak period, decline period and sedimentation period. Based on the life cycle classification label corresponding to the books, all books are divided into cold start recommendation library, growth recommendation library, outbreak recommendation library, decline recommendation library and sedimentation recommendation library; The recommendation strategy setting module is used to set book recommendation strategies based on the user's exploration intention for all books in the cold start recommendation library, set book recommendation strategies based on similar users for all books in the growth recommendation library, set book recommendation strategies based on user influence for all books in the outbreak recommendation library, set recommendation strategies consistent with the growth recommendation library for all books in the decline recommendation library, and build recommendation strategies based on hot information for all books in the sedimentation recommendation library.

Images