Artificial Intelligence-Based Information Recommendation Method and System

Abstract

The present application provides an information recommendation method and system based on artificial intelligence. The method includes: obtaining user preference data and user browsing data, where the user preference data reveals the data types of user preferences, and the user browsing data includes long-term browsing data and real-time browsing data. The long-term browsing data reveals the long-term browsing behavior of the user, and the real-time browsing data reveals the real-time browsing behavior of the user; determining the user characteristics of the present user according to the user preference data and the user browsing data; matching the user characteristics of the present user with the user characteristics of other users to determine the target user; determining the target information to be recommended based on the target user, and recommending the target information to the present user. By performing feature mining on the long-term and real-time browsing data of users through clustering analysis and combining the user interest preferences to construct feature vectors, it is not only possible to take into account the user interests, but also to overcome the cold start problem as much as possible, effectively improving the information recommendation effect.

Description

Technical Field

[0001] This application relates to the technical field of information recommendation, and more specifically, to an information recommendation method and system based on artificial intelligence. Background Art

[0002] With the development of the information age, traditional information recommendation solutions have gradually become difficult to adapt to the load brought about by the rapid increase in data volume, resulting in the existing information recommendation solutions becoming increasingly difficult to meet the requirements.

[0003] Traditional recommendation solutions include popularity-based recommendation solutions, similarity-based recommendation solutions, user collaborative filtering-based recommendation solutions, etc. Traditional recommendation solutions have indeed promoted the development of information recommendation services to a certain extent, but they each have limitations. For example, popularity-based recommendation solutions often only focus on the most popular content, but ignore the personalized needs of users; similarity-based recommendation solutions, although considering users' historical preferences, may fall into an "information cocoon" due to over-reliance on historical data and are difficult to explore users' potential interests. User-based collaborative filtering usually has a cold start problem. When encountering new users, due to the lack of sufficient historical data, the recommendation effect is not good.

[0004] As a service platform, pushing more information that matches the reading interests of platform subscribers (such as various types and fields of information like novel documents, factual news, entertainment news, scientific exploration information, etc.) is an effective means to improve the recommendation effect and enhance user stickiness. Therefore, how to provide a more effective information recommendation solution is a problem that the service platform needs to solve. Summary of the Invention

[0005] The purpose of the embodiments of this application is to provide an information recommendation method and system based on artificial intelligence, which can perform feature mining on users' long-term and real-time browsing data through clustering analysis, and at the same time, combined with users' interest preferences, jointly construct feature vectors as the basis for collaborative filtering analysis, which can not only take into account users' interests, but also overcome the cold start problem as much as possible and effectively improve the information recommendation effect.

[0006] To achieve the above purpose, the embodiments of this application are implemented as follows:

[0007] In a first aspect, an embodiment of the present application provides an information recommendation method based on artificial intelligence, including: obtaining user preference data and user browsing data, where the user preference data reveals the data types preferred by the user, and the user browsing data includes long-term browsing data and real-time browsing data. The long-term browsing data is used to reveal the user's long-term browsing behavior, and the real-time browsing data is used to reveal the user's real-time browsing behavior; determining the user characteristics of the present user according to the user preference data and the user browsing data; matching the user characteristics of the present user with the user characteristics of other users to determine the target user; determining the target information to be recommended based on the target user, and recommending the target information to the present user.

[0008] In combination with the first aspect, in the first possible implementation manner of the first aspect, determining the user characteristics of the present user according to the user preference data and the user browsing data includes: performing one-hot encoding on the user preference data to obtain user preference features; preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine long-term browsing features; preprocessing and feature extraction on the real-time browsing data in the user browsing data to determine real-time browsing features; combining the user preference features, long-term browsing features, and real-time browsing features into user characteristics.

[0009] In combination with the first possible implementation manner of the first aspect, in the second possible implementation manner of the first aspect, the long-term browsing data contains several pieces of long-term browsing information. Preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine long-term browsing features includes: preprocessing each piece of long-term browsing information in the long-term browsing data, where each piece of long-term browsing information corresponds to a piece of news information, and the long-term browsing information includes the time node, browsing duration, browsing entry, and operation frequency of the user browsing the corresponding news information; for each piece of long-term browsing information after preprocessing: performing feature extraction on the time node, browsing duration, browsing entry, and operation frequency in this piece of long-term browsing information to determine the browsing feature components, and fusing the browsing feature components with the label feature components of the corresponding news information to obtain the long-term browsing vector corresponding to this piece of long-term browsing information; inputting the long-term browsing vectors corresponding to all long-term browsing information into the first clustering analysis model, and outputting the distances between each long-term browsing vector and each clustering center as the optimized long-term vectors corresponding to each long-term browsing vector; based on each optimized long-term vector, extracting the overall distribution characteristics of all optimized long-term vectors as long-term browsing features.

[0010] Combined with the second possible implementation manner of the first aspect, in the third possible implementation manner of the first aspect, based on each optimized long-term vector, the overall distribution characteristics of all optimized long-term vectors are extracted as the long-term browsing characteristics, including: for each dimension in the optimized long-term vector: calculate the mean and standard deviation of all optimized long-term vectors in this dimension respectively, and correspondingly obtain the first eigenvalue and the second eigenvalue of this dimension; determine the number of optimized long-term vectors with the smallest absolute value and the largest absolute value in this dimension respectively, and correspondingly obtain the third eigenvalue and the fourth eigenvalue of this dimension; determine the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension respectively, and correspondingly obtain the fifth eigenvalue and the sixth eigenvalue of this dimension; based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue and the sixth eigenvalue of this dimension, determine the characteristics of this dimension, and further integrate the characteristics of all dimensions to obtain the long-term browsing characteristics.

[0011] Combined with the third possible implementation manner of the first aspect, in the fourth possible implementation manner of the first aspect, based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue and the sixth eigenvalue of this dimension, determine the characteristics of this dimension, including: based on the sixth eigenvalue of each dimension, determine several target dimensions with the largest sixth eigenvalue value; for each target dimension: determine the target optimized long-term vectors with values lower than the set value in the current target dimension, and count the mean and standard deviation of all target optimized long-term vectors in each dimension in the current target dimension as the supplementary eigenvalue of the corresponding dimension; for each dimension: based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue, the sixth eigenvalue and all supplementary eigenvalues of this dimension, determine the characteristics of this dimension.

[0012] Combined with the first possible implementation manner of the first aspect, in the fifth possible implementation manner of the first aspect, the real-time browsing data includes several pieces of real-time browsing information within a recent period. The real-time browsing data in the user browsing data is preprocessed and feature-extracted to determine the real-time browsing features, including: preprocessing each piece of real-time browsing information in the real-time browsing data, where each piece of real-time browsing information corresponds to a piece of information. The real-time browsing information includes the time node when the user browses the corresponding information, the browsing duration, the browsing entry, the operation frequency, the bounce rate, the interaction behavior, the subscription status, the reading completion rate, the repeated access rate, the page scrolling depth, and the click conversion rate; for each piece of preprocessed real-time browsing information: feature-extracting the time node, the browsing duration, the browsing entry, the operation frequency, the bounce rate, the interaction behavior, the subscription status, the reading completion rate, the repeated access rate, the page scrolling depth, and the click conversion rate in this piece of real-time browsing information to determine the browsing feature components, and fusing the browsing feature components with the title feature components, content feature components, label feature components, and popularity feature components of the corresponding information to obtain the real-time browsing vector corresponding to this piece of real-time browsing information; inputting the real-time browsing vectors corresponding to all real-time browsing information into the second clustering analysis model, and outputting the distances between each real-time browsing vector and each clustering center as the optimized real-time vectors corresponding to each real-time browsing vector; based on each optimized real-time vector, extracting the overall distribution features of all optimized real-time vectors as the real-time browsing features.

[0013] Combined with the fifth possible implementation manner of the first aspect, in the sixth possible implementation manner of the first aspect, based on each optimized real-time vector, extracting the overall distribution features of all optimized real-time vectors as the real-time browsing features, including: for each dimension in the optimized real-time vectors: respectively calculating the mean and standard deviation of all optimized real-time vectors in this dimension, and correspondingly obtaining the first feature quantity and the second feature quantity of this dimension; respectively determining the number of optimized real-time vectors with the minimum absolute value and the maximum absolute value in this dimension, and correspondingly obtaining the third feature quantity and the fourth feature quantity of this dimension; respectively determining the number of optimized real-time vectors with values lower than the mean and lower than a set value in this dimension, and correspondingly obtaining the fifth feature quantity and the sixth feature quantity of this dimension; based on the first feature quantity, the second feature quantity, the third feature quantity, the fourth feature quantity, the fifth feature quantity, and the sixth feature quantity of this dimension, determining the features of this dimension, and further integrating the features of all dimensions to obtain the real-time browsing features.

[0014] Combined with the first possible implementation manner of the first aspect, in the seventh possible implementation manner of the first aspect, matching the user characteristics of this user with the user characteristics of other users to determine the target user, including: for each other user: calculating the similarity between the user preference characteristics of this user and the user preference characteristics of this other user to obtain a preference similarity; calculating the similarity between the long-term browsing characteristics of this user and the long-term browsing characteristics of this other user to obtain a long-term similarity; calculating the similarity between the real-time browsing characteristics of this user and the real-time browsing characteristics of this other user to obtain a real-time similarity; calculating the user similarity between this user and this other user based on the preference similarity, long-term similarity, and real-time similarity; determining the target user based on the user similarity corresponding to each other user.

[0015] Combined with the first aspect, in the eighth possible implementation manner of the first aspect, determining the target information to be recommended based on the target user, including: obtaining the information recommendation list of each target user in the recent period; counting the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information; calculating the recommendation index of each piece of information based on the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information; determining the target information to be recommended according to the recommendation index of each piece of information.

[0016] In a second aspect, an information recommendation system based on artificial intelligence provided by an embodiment of the present application includes: a data acquisition module, configured to acquire user preference data and user browsing data, where the user preference data reveals the data types of user preferences, and the user browsing data includes long-term browsing data and real-time browsing data, the long-term browsing data is used to reveal the long-term browsing behavior of the user, and the real-time browsing data is used to reveal the real-time browsing behavior of the user; a feature extraction module, configured to determine the user characteristics of this user according to the user preference data and the user browsing data; a user matching module, configured to match the user characteristics of this user with the user characteristics of other users to determine the target user; an information recommendation module, configured to determine the target information to be recommended based on the target user and recommend the target information to this user.

[0017] Beneficial effects:

[0018] (1) This solution determines user characteristics of the current user by obtaining user preference data (determined based on the user's settings, access, sharing, and favorite records, generally generated by the platform based on the user's information, and the user preference data in the platform can be directly obtained in this solution) and user browsing data (including long-term browsing data and real-time browsing data, where the long-term browsing data is used to reveal the user's long-term browsing behavior and the real-time browsing data is used to reveal the user's real-time browsing behavior). By one-hot encoding the user preference data, user preference features are obtained; the long-term browsing data in the user browsing data is preprocessed and feature extracted (including clustering analysis) to determine long-term browsing features; the real-time browsing data in the user browsing data is preprocessed and feature extracted (including clustering analysis) to determine real-time browsing features, which are finally combined into user characteristics. The user characteristics of the current user are matched with those of other users (by performing corresponding similarity calculations) to determine target users, and then, based on the information recommendation list of the target users (analyzing several indicators such as the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information), the target information to be recommended is determined and recommended to the current user. Through this solution, user preferences, long-term browsing data, and short-term browsing data (i.e., real-time browsing data) can be comprehensively analyzed, feature mining can be performed on the user's long-term and real-time browsing data through clustering analysis, and at the same time, combined with the user's interest preferences, a feature vector is jointly constructed as the basis for collaborative filtering analysis. This method can not only take into account the user's interests but also overcome the cold start problem as much as possible, effectively improving the information recommendation effect.

[0019] (2) In the feature extraction of long-term browsing data, each long-term browsing information (including the time node when the user browsed the corresponding information information, browsing time, browsing entry and operation frequency) is pre-processed, and then preliminary feature extraction is performed (for example, unified format and standardization processing are performed to obtain the corresponding browsing feature component, and then it is merged with the label feature component of the corresponding information information) to obtain the long-term browsing vector corresponding to this long-term browsing information, and the long-term browsing vectors corresponding to all long-term browsing information are input into the first clustering analysis model (the clustering analysis model of this embodiment is the K-Means clustering algorithm, which is obtained through training and regularly optimized through the long-term browsing data of all users in the platform), and cluster analysis is performed to output the distance between each long-term browsing vector and each cluster center (a multi-dimensional vector will be formed, and the dimension corresponds to the number of cluster centers in the first clustering analysis model) as the optimized long-term vector corresponding to each long-term browsing vector, and further extract the overall distribution characteristics of all optimized long-term vectors as the long-term browsing feature. Specifically, the mean and standard deviation of all optimized long-term vectors in this dimension are calculated, and the first and second eigenvalues of this dimension are obtained; the number of optimized long-term vectors with the minimum absolute value and the maximum absolute value in this dimension are determined, and the third and fourth eigenvalues of this dimension are obtained; the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension are determined, and the fifth and sixth eigenvalues of this dimension are obtained; based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue and the sixth eigenvalue in this dimension, the features in this dimension are determined, and the features in all dimensions are further integrated to obtain the long-term browsing features. In this way, not only can the first clustering analysis model be used to perform clustering analysis on the large number of long-term browsing information of this user with different quantities, and further mine potential features, but also the distribution statistics of the optimized long-term vectors corresponding to the output of each long-term browsing information can be performed (reflected in the form of quantity or percentage, and normalization can be performed when the quantity is used for statistics), so as to overcome the cold start problem caused by the small amount of data as much as possible.When constructing long-term browsing features, data on some key dimensions (key clustering centers, which can reflect user personalization to a certain extent) are also analyzed. For example, based on the sixth eigenvalue on each dimension (i.e., the number of optimized long-term vectors that are lower than the set value on a certain dimension, which can reflect the concentration degree, that is, a large number of long-term browsing information aggregates at this clustering center), several target dimensions with the largest sixth eigenvalue values are determined (for example, the top three are selected); for each target dimension: target optimized long-term vectors with values lower than the set value on the current target dimension are determined, and the mean and standard deviation of all target optimized long-term vectors on the current target dimension on each dimension are calculated as the supplementary eigenvalue on the corresponding dimension; for each dimension: based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, sixth eigenvalue, and all supplementary eigenvalues on this dimension, the feature on this dimension is determined. Thus, not only the overall distribution of the user's long-term browsing data is considered, but also the potentially personalized features of the user are considered, which can further improve the reflection ability of the long-term browsing features, so as to more accurately match similar users for the user (perform user-based collaborative filtering) and improve the information recommendation effect.

[0020] (3) In the feature extraction of real-time browsing data, each real-time browsing information in the real-time browsing data is preprocessed. Each real-time browsing information corresponds to a piece of news information. The real-time browsing information includes the time node, browsing duration, browsing entry, operation frequency, bounce rate, interactive behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click-through conversion rate of the user browsing the corresponding news information. Feature extraction is performed to determine the browsing feature component, and the browsing feature component is fused with the title feature component, content feature component, label feature component, and heat feature component of the corresponding news information to obtain the real-time browsing vector corresponding to the real-time browsing information. The real-time browsing information contains more detailed indicators (compared to the long-term browsing information), which can well reflect the user's real-time reading and operation status, and the corresponding news information contains richer feature information (including title feature component, content feature component, label feature component, heat feature component, etc., which can be obtained by obtaining the relevant data of the corresponding news information through the platform), so as to better reflect the user's current status and interests. Afterwards, the real-time browsing vectors corresponding to all the real-time browsing information are input into the second cluster analysis model (obtained by training using the K-Means clustering algorithm), and the distance between each real-time browsing vector and each cluster center is output as the optimized real-time vector corresponding to each real-time browsing vector; based on each optimized real-time vector, the overall distribution characteristics of all optimized real-time vectors are extracted as real-time browsing features. For each dimension in the optimized real-time vector: the mean and standard deviation of all optimized real-time vectors in this dimension are calculated respectively, and the first feature quantity and the second feature quantity of this dimension are obtained; the number of optimized real-time vectors with the minimum absolute value and the maximum absolute value in this dimension are determined respectively, and the third feature quantity and the fourth feature quantity of this dimension are obtained respectively; the number of optimized real-time vectors with values lower than the mean and lower than the set value in this dimension are determined respectively, and the fifth feature quantity and the sixth feature quantity of this dimension are obtained respectively; based on the first feature quantity, the second feature quantity, the third feature quantity, the fourth feature quantity, the fifth feature quantity and the sixth feature quantity in this dimension, the features in this dimension are determined, and the features in all dimensions are further integrated to obtain the real-time browsing features. In this way, not only can the second clustering analysis model be used to perform cluster analysis on different quantities of real-time browsing information of the user to further explore potential features, but also distribution statistics can be performed based on the optimized real-time vectors output corresponding to each real-time browsing information (which can also be reflected in the form of quantity or percentage, and normalization can be performed when using quantity for statistics), so as to overcome the cold start problem caused by the small amount of data as much as possible.

[0021] (4) When performing user matching and information recommendation, calculate the preference similarity between the user preference features of the current user and those of other users; calculate the long-term similarity between the long-term browsing features of the current user and those of other users; calculate the real-time similarity between the real-time browsing features of the current user and those of other users; based on the preference similarity, long-term similarity, and real-time similarity, calculate the user similarity between the current user and other users; based on the user similarity corresponding to each other user, determine the target users. In this way, the target users that are more matched in all aspects can be determined by integrating features from multiple aspects, which is very suitable for platforms with an increasingly large number of users. When performing information recommendation, obtain the list of recommended information for each target user within the recent period; count the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information, calculate the recommendation index of each piece of information, and determine the target information to be recommended. In this way, information can be better recommended to users, effectively improving the recommendation effect.

[0022] To make the above objects, features, and advantages of the present application more obvious and understandable, the following specifically presents preferred embodiments and, in conjunction with the accompanying drawings, provides a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] To more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of an information recommendation method based on artificial intelligence provided by an embodiment of the present application.

[0025] Figure 2 It is a schematic diagram for obtaining access-related data of some users.

[0026] Figure 3 It is a schematic diagram for subscription-related data of some users.

[0027] Figure 4 It is a structural block diagram of an information recommendation system based on artificial intelligence.

[0028] Reference numerals: 10 - Information recommendation system based on artificial intelligence; 11 - Data acquisition module; 12 - Feature extraction module; 13 - User matching module; 14 - Information recommendation module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] Next, the technical solutions in the embodiments of the present application will be described in conjunction with the accompanying drawings in the embodiments of the present application.

[0030] To implement information recommendation for platform users, this embodiment provides an information recommendation method based on artificial intelligence. Please refer to Figure 1 , Figure 1 which is a flowchart of the information recommendation method based on artificial intelligence. In this embodiment, the information recommendation method based on artificial intelligence may include step S10, step S20, step S30, and step S40.

[0031] First, when a user browses information, relevant information can be recommended to the user. At this time, step S10 can be executed.

[0032] Step S10: Obtain user preference data and user browsing data. Among them, user preference data reveals the data types preferred by the user, and user browsing data includes long-term browsing data and real-time browsing data. Long-term browsing data is used to reveal the user's long-term browsing behavior, and real-time browsing data is used to reveal the user's real-time browsing behavior.

[0033] In this embodiment, user preference data and user browsing data can be obtained.

[0034] User preference data reveals the data types preferred by the user, which can be determined through the user's settings, access, sharing, subscriptions, favorite records, etc. Generally, the platform generates a user portrait based on the user's information, and determines the corresponding user preference data based on the user portrait. In this embodiment, the platform can directly obtain the user's user preference data, which will not be elaborated.

[0035] User browsing data includes long-term browsing data and real-time browsing data. Long-term browsing data is used to reveal the user's long-term browsing behavior, and real-time browsing data is used to reveal the user's real-time browsing behavior.

[0036] The long-term browsing data may include several pieces of long-term browsing information. Each piece of long-term browsing information corresponds to a piece of information. The long-term browsing information includes the time node, browsing duration, browsing entry, and operation frequency of the user browsing the corresponding information. For the information corresponding to the long-term browsing information, each piece of information may include the corresponding label feature component (each piece of information in the platform actually has many features, such as title features, several content features, label features, popularity features, etc. However, when analyzing long-term browsing information, so many features are not required, and only the label feature component needs to be obtained).

[0037] The real-time browsing data may include several pieces of real-time browsing information within a recent period (such as the most recent ten minutes, the most recent half hour, etc. Of course, in other solutions, the number of pieces of information to be browsed can also be fixed, such as the most recent 10 pieces. However, this solution is prone to ignoring the time coherence and has certain analysis defects. To overcome this, the continuity of these browsing information in the time dimension still needs to be considered. Therefore, this solution is illustrated by taking a period as an example). Each piece of real-time browsing information corresponds to a piece of information. The real-time browsing information includes the time node when the user browses the corresponding information, the browsing duration, the browsing entry (i.e., how to enter the browsing, such as entering through search, historical record, recommendation, etc.), the operation frequency (the number of operations such as clicks, swipes, scrolls, etc. of the user on the page, reflecting the user's activity and participation), the bounce rate (the proportion of users who exit after only browsing one page, reflecting the attractiveness of the page content to users), interactive behaviors (such as liking, commenting, sharing, reporting, etc.), subscription status (reflecting the user's subscription information. This platform launches various subscription packages, which can largely reflect the user's preference for information), the reading completion rate (the proportion of users who read the entire article or watch the entire video or listen to the entire music, measuring the attractiveness of the information content and the user's reading depth and immersion depth), the repeat visit rate (the proportion of users who visit the same piece of information multiple times, reflecting the user's interest or confusion about the content), the page scroll depth (the proportion or distance of the user's page scrolling, or the duration progress of the user watching the video or listening to the music, reflecting the user's attention degree and browsing depth to the content), the click-through conversion rate (the proportion of users who click on a specific button, such as reading the full text, sharing, commenting, etc., measuring the attractiveness of the information and the user's participation), etc. When analyzing the real-time browsing information, the characteristics of the information need to be considered. Therefore, as much information as possible that can be obtained, such as the title feature component, the content feature component, the tag feature component, the popularity feature component, etc., needs to be obtained.

[0038] As Figure 2 and Figure 3 shown, Figure 2 and Figure 3 show a schematic diagram of some user access and subscription related data (the interface is from the subscription data management subsystem independently developed by the applicant, which is used to provide some data support for the platform). Many data are difficult to be directly displayed through the display interface of the platform system and are not suitable for display. Therefore, only a small part of examples are given, and this part of data has been protected for user privacy and the user privacy information has not been exposed.

[0039] After obtaining the user preference data and the user browsing data, step S20 can be executed.

[0040] Step S20: Determine the user characteristics of this user according to the user preference data and the user browsing data.

[0041] In this embodiment, according to the user preference data and the user browsing data, the user characteristics of this user are determined.

[0042] First, the user preference data can be one-hot encoded to obtain user preference features. This part of the technology is very simple, which is equivalent to providing a full-label data. If the user has a certain label, the value is determined to be 1 on the corresponding dimension, otherwise it is 0. The obtained feature vector is this user preference feature.

[0043] Secondly, the long-term browsing data in the user browsing data can be preprocessed and feature extracted to determine the long-term browsing features.

[0044] Exemplarily, each piece of long-term browsing information in the long-term browsing data can be preprocessed. Here, each piece of long-term browsing information corresponds to a piece of news information, and the long-term browsing information includes the time node, browsing duration, browsing entry, and operation frequency of the user browsing the corresponding news information.

[0045] Then, for each piece of long-term browsing information after preprocessing: the time node, browsing duration, browsing entry, and operation frequency in this piece of long-term browsing information can be feature extracted to determine the browsing feature components, and the browsing feature components are fused with the label feature components of the corresponding news information to obtain the long-term browsing vector corresponding to this piece of long-term browsing information.

[0046] Specifically, the time nodes in each piece of long-term browsing information can be unified, and the browsing time nodes are converted into a unified format (such as UNIX timestamp), and time-related features such as the time period of the day (morning, afternoon, evening), weekdays and weekends, months, seasons, etc. can be extracted as needed to capture the user's browsing preferences at different time periods; and, the browsing duration is statistically calculated (accurate to seconds), the browsing entry is assigned a value (different values are assigned according to different browsing entries, for example, the search entry is assigned 1, the history record entry is assigned 2, the interface recommendation entry is assigned 3, the associated recommendation entry is assigned 4, etc.) and the operation frequency (the number of operations per second or per minute is statistically calculated). The initially extracted browsing feature components are fused with the label feature components of the corresponding news information (such as splicing) to obtain the long-term browsing vector corresponding to each piece of long-term browsing information. Each long-term browsing vector obtained after such processing is in a unified format for subsequent clustering analysis.

[0047] After obtaining the long-term browsing vectors corresponding to each piece of long-term browsing information, the long-term browsing vectors corresponding to all the long-term browsing information of this user can be input into the first clustering analysis model (using the K-Means clustering algorithm, obtained through training, and regularly optimized with the long-term browsing data of all users within the platform, for example, updated once a month). Then, through clustering analysis of each long-term browsing vector, the distances between the long-term browsing vector and each cluster center in the first clustering analysis model can be calculated, thereby obtaining a vector as the optimized long-term vector corresponding to this long-term browsing vector. Assuming that the number of cluster centers of the first clustering analysis model is k and the number of long-term browsing vectors is n, n optimized long-term vectors can be obtained accordingly.

[0048] After that, based on each optimized long-term vector, the overall distribution characteristics of all optimized long-term vectors can be extracted as the long-term browsing characteristics.

[0049] Exemplarily, for each dimension in the optimized long-term vector:

[0050] The mean and standard deviation of all optimized long-term vectors in this dimension can be calculated respectively to obtain the first eigenvalue and the second eigenvalue of this dimension. For example, the i-th optimized long-term vector is denoted as r i =(x i1 ,x i2 ,…,x ij ,…,x ik ), i ∈ [1, n]. Then, for the j-th dimension, j ∈ [1, k], the following formulas are used to calculate the mean and standard deviation of all the optimized long-term vectors of this user in this dimension j:

[0051]

[0052] where μ j is the mean of all the optimized long-term vectors of this user in dimension j, n is the number of the long-term browsing vectors of this user, x ij is the value of the i-th optimized long-term vector in dimension j, and σ j is the standard deviation of all the optimized long-term vectors of this user in dimension j.

[0053] In addition, the number of optimized long-term vectors with the smallest absolute value and the largest absolute value in this dimension can be determined respectively to obtain the third eigenvalue and the fourth eigenvalue of this dimension. For example, determine the dimension where the smallest absolute value and the largest absolute value of each optimized long-term vector (in this embodiment, the elements of the optimized long-term vector are all non-negative, so the absolute value is itself) are located. Then, for each dimension (taking the j-th dimension as an example), count the number a j of the optimized long-term vectors with the smallest absolute value in this dimension (the proportion can also be calculated to weaken the difference in quantity, for example ) and the number b of optimized long-term vectors that are the maximum absolute values in this dimension j (The proportion can also be calculated to weaken the difference in quantity. For example ) as the third eigenvalue (proportion α j as the seventh eigenvalue) and the fourth eigenvalue (proportion β j as the eighth eigenvalue).

[0054] The number of optimized long-term vectors that are lower than the mean value and lower than the set value (for example, the mean value calculated from the minimum absolute value elements of each optimized long-term vector) in this dimension can also be determined respectively, and the fifth eigenvalue and the sixth eigenvalue of this dimension are correspondingly obtained. For example, for each dimension (taking the jth dimension as an example), count the number c of optimized long-term vectors that are lower than the mean value μ j in this dimension j (The proportion can also be calculated to weaken the difference in quantity. For example ) as the fifth eigenvalue (proportion γ j as the ninth eigenvalue), count the number d of optimized long-term vectors that are lower than the set value in this dimension j (The proportion can also be calculated to weaken the difference in quantity. For example ) as the sixth eigenvalue (proportion δ j as the tenth eigenvalue).

[0055] After that, based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue and the sixth eigenvalue in this dimension, the feature in this dimension can be determined, and the long-term browsing features are further integrated by integrating the features of all dimensions.

[0056] Exemplarily, based on the sixth eigenvalue of each dimension, several target dimensions with the largest sixth eigenvalue value can be determined (for example, 1, 3, 5, etc., taking 3 as an example in this embodiment).

[0057] Then, for each target dimension: the target optimized long-term vectors that are lower than the set value (the set value of the current target dimension) in the current target dimension can be determined (for example, there are y target optimized long-term vectors), and the mean value and the standard deviation of all target optimized long-term vectors (that is, these y target optimized long-term vectors) in each dimension in the current target dimension are counted as the supplementary eigenvalue of the corresponding dimension.

[0058] Finally, for each dimension: based on the first eigenvalue, the second eigenvalue, the third eigenvalue (and the seventh eigenvalue), the fourth eigenvalue (and the eighth eigenvalue), the fifth eigenvalue (and the ninth eigenvalue), the sixth eigenvalue (and the tenth eigenvalue), and all supplementary eigenvalues on this dimension, the feature on this dimension can be determined. Finally, the long-term browsing feature containing a total of k dimensions with z elements in each dimension can be obtained.

[0059] After determining the long-term browsing feature of the user, the real-time browsing data in the user browsing data can be preprocessed and feature extracted to determine the real-time browsing feature.

[0060] Exemplarily, each piece of real-time browsing information in the real-time browsing data can be preprocessed. Here, each piece of real-time browsing information corresponds to a piece of news information, and the real-time browsing information includes the time node when the user browses the corresponding news information, the browsing duration, the browsing entrance, the operation frequency, the bounce rate, the interaction behavior, the subscription status, the reading completion rate, the repeat visit rate, the page scroll depth, and the click conversion rate.

[0061] Then, for each piece of preprocessed real-time browsing information: feature extraction is performed on the time node, browsing duration, browsing entrance, operation frequency, bounce rate, interaction behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click conversion rate in this piece of real-time browsing information to determine the browsing feature component, and the browsing feature component is fused with the title feature component, content feature component, tag feature component, and popularity feature component of the corresponding news information to obtain the real-time browsing vector corresponding to this piece of real-time browsing information.

[0062] Similarly, the time nodes in each real-time browsing information can be unified, the time nodes can be converted into a unified format (such as UNIX timestamp), and time-related features can be extracted as needed, such as the time period of the day (morning, afternoon, evening), weekdays and weekends, months, seasons, etc.; and, the browsing duration can be counted (accurate to seconds), the browsing entrance can be assigned a value (different values are assigned according to different browsing entrances. For example, for the search entrance, the value 1 is assigned, for the history record entrance, the value 2 is assigned, for the interface recommendation entrance, the value 3 is assigned, for the associated recommendation entrance, the value 4 is assigned, etc.) and the operation frequency (count the number of operations per second or per minute). Also, the bounce rate can be assigned a value (here, since it is counted for each real-time browsing information, for those who bounce after browsing one page, the value 1 is assigned, and for other cases, the value 0 is assigned), interactive behavior (if there is no interactive behavior, the value 0 is assigned, if there is interactive behavior, corresponding values are assigned according to different interactive behaviors. If there are multiple interactive behaviors, they are combined and assigned according to the existing behaviors. For example, if there are cases where the values 1 and 2 exist simultaneously, it is recorded as 12), subscription status (if there is a subscription, the value 1 is assigned, if there is no subscription, the value 0 is assigned), reading completion rate (if the reading is completed, the value 1 is assigned, if the reading is not completed, the value 0 is assigned), repeat visit rate (whether it is a repeated visit to the information, if so, it is recorded as 1, otherwise it is recorded as 0), page scroll depth (assigned according to the proportion of the scroll depth to the total depth), click-through conversion rate (if the relevant button is not clicked, the value 0 is recorded, if the relevant button is clicked, different values are assigned according to different button types. For those with multiple click behaviors, they are combined and assigned according to the existing click behaviors. For example, if there are cases where the values 1 and 2 exist simultaneously, it is recorded as 12) for feature extraction, and then fused (such as concatenated) with the title feature component, content feature component, label feature component, and popularity feature component of the corresponding information to obtain the real-time browsing vector corresponding to this real-time browsing information. Each real-time browsing vector obtained after such processing is in a unified format for subsequent clustering analysis.

[0063] Accordingly, the real-time browsing vectors corresponding to all real-time browsing information can be input into the second clustering analysis model (also using the K-Means clustering algorithm, obtained through training, but the training data is different from the first clustering analysis model), and the distances between each real-time browsing vector and each cluster center are output as the optimized real-time vectors corresponding to each real-time browsing vector.

[0064] Exemplarily, for optimizing each dimension in the real-time vector: The mean and standard deviation of all optimized real-time vectors in this dimension can be calculated respectively, and the first feature quantity and the second feature quantity of this dimension are correspondingly obtained. The number of optimized real-time vectors with the minimum absolute value and the maximum absolute value in this dimension is determined respectively, and the third feature quantity and the fourth feature quantity of this dimension are correspondingly obtained. The number of optimized real-time vectors with values lower than the mean and lower than a set value in this dimension is determined respectively, and the fifth feature quantity and the sixth feature quantity of this dimension are correspondingly obtained. For the specific calculation and determination methods, refer to the processing process of the optimized long-term vector in the previous text, which will not be elaborated here.

[0065] After that, based on the first feature quantity, the second feature quantity, the third feature quantity, the fourth feature quantity, the fifth feature quantity, and the sixth feature quantity in this dimension, the feature in this dimension can be determined, and further integrating the features in all dimensions to obtain the real-time browsing feature. The difference from the long-term browsing feature in the previous text here is that since the optimized real-time vector itself considers many real-time detailed features (such as user operations, relevant features of information, etc.), there is no need to calculate and integrate supplementary feature values.

[0066] After obtaining the real-time browsing feature of this user, the user preference feature, the long-term browsing feature, and the real-time browsing feature can be combined into a user feature (no splicing is required because there are differences in the number of dimensions or attributes among the user preference feature, the long-term browsing feature, and the real-time browsing feature).

[0067] After obtaining the user feature, step S30 can be executed.

[0068] Step S30: Match the user feature of this user with the user features of other users to determine the target user.

[0069] In this embodiment, the user feature of this user can be matched with the user features of other users. Then, for each other user:

[0070] First, the similarity between the user preference feature of this user and the user preference feature of this other user can be calculated (for example, by calculating the cosine similarity) to obtain the preference similarity.

[0071] Second, the similarity between the long-term browsing feature of this user and the long-term browsing feature of this other user can be calculated: Specifically, the similarity score of each dimension can be measured by the cosine similarity, and then the similarity scores of all dimensions are comprehensively weighted and summed (the weights can be measured according to the set value of each dimension, and the weights are allocated according to the ratio of the set values of each dimension, which can further strengthen the personalized features of the user to match more similar users) to obtain the comprehensive similarity score as the long-term similarity. For example,

[0072] Furthermore, the similarity between the real-time browsing features of this user and those of this other user can be calculated: Similarly, specifically, the similarity scores in each dimension can be measured by cosine similarity, and then the similarity scores in each dimension are comprehensively considered and weighted and summed (the weights can be measured according to the set values of each dimension, and the weights are allocated according to the ratio of the set values of each dimension, which can further strengthen the personalized features of the user to match more similar users), and the comprehensive similarity score is obtained as the real-time similarity.

[0073] Finally, based on the preference similarity, long-term similarity, and real-time similarity, the user similarity between this user and this other user is calculated. Here, the overall similarity is obtained by weighted summation as the user similarity. The weight design can be preset. For example, the preference similarity accounts for 0.3, the long-term similarity accounts for 0.3, and the real-time similarity accounts for 0.4. It can also be a dynamically adjusted weight design, where the weights are dynamically allocated according to the importance of each similarity or a certain indicator. For example, the preference similarity accounts for 0.3, and the remaining 0.7 is dynamically allocated according to the size of the preference similarity to allocate corresponding weights to the long-term similarity and the real-time similarity. This is not limited here.

[0074] After that, based on the user similarity corresponding to each other user, the target users can be determined (for example, 100 other users with the highest user similarity are determined as the target users).

[0075] After the target users are determined, step S40 can be executed.

[0076] Step S40: Based on the target users, determine the target information to be recommended and recommend the target information to this user.

[0077] In this embodiment, the information recommendation list of each target user in the recent period can be obtained (for example, in the recent 1 hour, or the 100 pieces of information browsed recently). Then, the number of recommendation times of each piece of information (that is, the number of times each piece of information is recommended to these target users), click-through rate (the ratio of the number of times the user clicks on this piece of information to the number of recommendation times of this piece of information), bounce rate (the probability that after the information is recommended to the target user each time and is clicked, the user exits after browsing no more than one page or less than a certain time), reading completion rate (the probability that after the information is recommended to the target user each time and is clicked, the browsing is completed), repeat visit rate (the probability that the information is recommended to the user non-first time and is clicked by the user), and click conversion rate (the probability that after the information is recommended to the target user each time and is clicked, there are further clicks such as viewing the full text, continuing to watch, etc.) can be counted.

[0078] The number of recommended times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information can be determined, and the recommendation index of each piece of information can be comprehensively calculated. The calculation method of this recommendation index is related to the specific types of information (such as text information, videos, music, real-time news, popular science documents, etc.). Different types of information have different calculation methods. The measurement methods of these recommendation indexes are not the focus of the present invention and can also be flexibly changed under the influence of various factors such as different requirements and application scenarios, which will not be elaborated here one by one.

[0079] After determining the recommendation index of each piece of information, several pieces with the highest recommendation index (such as 5 pieces, 10 pieces, 30 pieces) can be determined as the target information to be recommended and recommended to the user.

[0080] Please refer to Figure 4 , based on the same inventive concept, this embodiment also provides an information recommendation system 10 based on artificial intelligence, including:

[0081] A data acquisition module 11, configured to acquire user preference data and user browsing data. Among them, the user preference data reveals the data types preferred by the user, and the user browsing data includes long-term browsing data and real-time browsing data. The long-term browsing data is used to reveal the long-term browsing behavior of the user, and the real-time browsing data is used to reveal the real-time browsing behavior of the user.

[0082] A feature extraction module 12, configured to determine the user features of this user according to the user preference data and the user browsing data.

[0083] A user matching module 13, configured to match the user features of this user with the user features of other users to determine the target user.

[0084] An information recommendation module 14, configured to determine the target information to be recommended based on the target user and recommend the target information to this user.

[0085] In this embodiment, the feature extraction module 12 is specifically configured to: perform one-hot encoding on the user preference data to obtain user preference features; perform preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine long-term browsing features; perform preprocessing and feature extraction on the real-time browsing data in the user browsing data to determine real-time browsing features; and combine the user preference features, long-term browsing features, and real-time browsing features into user features.

[0086] In this embodiment, the feature extraction module 12, where the long-term browsing data includes several pieces of long-term browsing information, is specifically configured to: preprocess each piece of long-term browsing information in the long-term browsing data, where each piece of long-term browsing information corresponds to a piece of news information, and the long-term browsing information includes the time node, browsing duration, browsing entry, and operation frequency of the user browsing the corresponding news information; for each piece of long-term browsing information after preprocessing: extract features from the time node, browsing duration, browsing entry, and operation frequency in this piece of long-term browsing information to determine browsing feature components, and fuse the browsing feature components with the label feature components of the corresponding news information to obtain a long-term browsing vector corresponding to this piece of long-term browsing information; input the long-term browsing vectors corresponding to all long-term browsing information into the first clustering analysis model, and output the distances between each long-term browsing vector and each cluster center as the optimized long-term vectors corresponding to each long-term browsing vector; based on each optimized long-term vector, extract the overall distribution characteristics of all optimized long-term vectors as the long-term browsing features.

[0087] In this embodiment, the feature extraction module 12 is specifically configured to: for each dimension in the optimized long-term vector: calculate the mean and standard deviation of all optimized long-term vectors in this dimension respectively to obtain the first eigenvalue and the second eigenvalue of this dimension; determine the number of optimized long-term vectors with the minimum absolute value and the maximum absolute value in this dimension respectively to obtain the third eigenvalue and the fourth eigenvalue of this dimension; determine the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension respectively to obtain the fifth eigenvalue and the sixth eigenvalue of this dimension; based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, and sixth eigenvalue of this dimension, determine the feature of this dimension, and further integrate the features of all dimensions to obtain the long-term browsing features.

[0088] In this embodiment, the feature extraction module 12 is specifically configured to: based on the sixth eigenvalue of each dimension, determine several target dimensions with the largest sixth eigenvalue value; for each target dimension: determine the target optimized long-term vectors with values lower than the set value in the current target dimension, and statistically calculate the mean and standard deviation of all target optimized long-term vectors in each dimension in the current target dimension as the supplementary eigenvalue of the corresponding dimension; for each dimension: based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, sixth eigenvalue, and all supplementary eigenvalues of this dimension, determine the feature of this dimension.

[0089] In this embodiment, the real-time browsing data includes several pieces of real-time browsing information within a recent period. The feature extraction module 12 is specifically configured to: preprocess each piece of real-time browsing information in the real-time browsing data, where each piece of real-time browsing information corresponds to a piece of news information, and the real-time browsing information includes the time node when the user browses the corresponding news information, browsing duration, browsing entry, operation frequency, bounce rate, interaction behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click conversion rate; for each piece of preprocessed real-time browsing information: extract features from the time node, browsing duration, browsing entry, operation frequency, bounce rate, interaction behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click conversion rate in this piece of real-time browsing information to determine browsing feature components, and fuse the browsing feature components with the title feature components, content feature components, tag feature components, and popularity feature components of the corresponding news information to obtain a real-time browsing vector corresponding to this piece of real-time browsing information; input the real-time browsing vectors corresponding to all real-time browsing information into the second clustering analysis model, and output the distances between each real-time browsing vector and each cluster center as the optimized real-time vectors corresponding to each real-time browsing vector; based on each optimized real-time vector, extract the overall distribution features of all optimized real-time vectors as the real-time browsing features.

[0090] In this embodiment, the feature extraction module 12 is specifically configured to: for each dimension in the optimized real-time vector: calculate the mean and standard deviation of all optimized real-time vectors in this dimension respectively to obtain the first feature quantity and the second feature quantity of this dimension; determine the number of optimized real-time vectors with the minimum absolute value and the maximum absolute value in this dimension respectively to obtain the third feature quantity and the fourth feature quantity of this dimension; determine the number of optimized real-time vectors with values lower than the mean and lower than a set value in this dimension respectively to obtain the fifth feature quantity and the sixth feature quantity of this dimension; based on the first feature quantity, second feature quantity, third feature quantity, fourth feature quantity, fifth feature quantity, and sixth feature quantity of this dimension, determine the features of this dimension, and further integrate the features of all dimensions to obtain the real-time browsing features.

[0091] In this embodiment, the user matching module 13 is specifically configured to: for each other user: calculate the similarity between the user preference features of this user and the user preference features of this other user to obtain the preference similarity; calculate the similarity between the long-term browsing features of this user and the long-term browsing features of this other user to obtain the long-term similarity; calculate the similarity between the real-time browsing features of this user and the real-time browsing features of this other user to obtain the real-time similarity; based on the preference similarity, long-term similarity, and real-time similarity, calculate the user similarity between this user and this other user; based on the user similarity corresponding to each other user, determine the target user.

[0092] In this embodiment, the information recommendation module 14 is specifically configured to: obtain the information recommendation list of each target user in the recent period; count the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information; calculate the recommendation index of each piece of information based on the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information; and determine the target information to be recommended according to the recommendation index of each piece of information.

[0093] In summary, the embodiment of the present application provides an information recommendation method and system based on artificial intelligence. By obtaining user preference data (determined according to the user's settings, access, sharing, and collection records, generally generated by the platform based on the user's information, and the user preference data in this solution can be directly obtained from the platform) and user browsing data (including long-term browsing data and real-time browsing data, where the long-term browsing data is used to reveal the user's long-term browsing behavior and the real-time browsing data is used to reveal the user's real-time browsing behavior), the user characteristics of the current user are determined according to the user preference data and the user browsing data (by performing one-hot encoding on the user preference data to obtain user preference features; preprocessing and feature extraction (including clustering analysis) on the long-term browsing data in the user browsing data to determine long-term browsing features; preprocessing and feature extraction (including clustering analysis) on the real-time browsing data in the user browsing data to determine real-time browsing features, and finally combining them into user characteristics); matching the user characteristics of the current user with the user characteristics of other users (performing corresponding similarity calculations) to determine the target user, and then determining the target information to be recommended according to the information recommendation list of the target user (analyzing several indicators, such as the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information), and recommending the target information to the current user. Through this solution, the user preferences, long-term browsing data, and short-term browsing data (i.e., real-time browsing data) can be comprehensively analyzed, the long-term and real-time browsing data of the user can be mined for features through clustering analysis, and at the same time, combined with the user's interest preferences, a feature vector is jointly constructed as the basis for collaborative filtering analysis. This method can not only take into account the user's interests but also overcome the cold start problem as much as possible, effectively improving the information recommendation effect.

[0094] In the feature extraction of long-term browsing data, each long-term browsing information (including the time node when the user browsed the corresponding information information, browsing time, browsing entry and operation frequency) is preprocessed, and then preliminary feature extraction is performed (for example, unified format and standardization processing are performed to obtain the corresponding browsing feature components, and then merged with the label feature components of the corresponding information information) to obtain the long-term browsing vector corresponding to this long-term browsing information, and the long-term browsing vectors corresponding to all long-term browsing information are input into the first clustering analysis model (the clustering analysis model of this embodiment is the K-Means clustering algorithm, which is obtained through training and regularly optimized through the long-term browsing data of all users in the platform), and clustering analysis is performed to output the distance between each long-term browsing vector and each cluster center (a multidimensional vector will be formed, and the dimension corresponds to the number of cluster centers in the first clustering analysis model) as the optimized long-term vector corresponding to each long-term browsing vector, and further extract the overall distribution characteristics of all optimized long-term vectors as long-term browsing features. Specifically, the mean and standard deviation of all optimized long-term vectors in this dimension are calculated, and the first and second eigenvalues of this dimension are obtained; the number of optimized long-term vectors with the minimum absolute value and the maximum absolute value in this dimension are determined, and the third and fourth eigenvalues of this dimension are obtained; the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension are determined, and the fifth and sixth eigenvalues of this dimension are obtained; based on the first eigenvalue, the second eigenvalue, the third eigenvalue, the fourth eigenvalue, the fifth eigenvalue and the sixth eigenvalue in this dimension, the features in this dimension are determined, and the features in all dimensions are further integrated to obtain the long-term browsing features. In this way, not only can the first clustering analysis model be used to perform clustering analysis on the large number of long-term browsing information of this user with different quantities, and further mine potential features, but also the distribution statistics of the optimized long-term vectors corresponding to the output of each long-term browsing information can be performed (reflected in the form of quantity or percentage, and normalization can be performed when the quantity is used for statistics), so as to overcome the cold start problem caused by the small amount of data as much as possible.When constructing long-term browsing features, data on some key dimensions (key clustering centers, which can reflect user personalization to a certain extent) are also analyzed. For example, based on the sixth eigenvalue on each dimension (i.e., the number of optimized long-term vectors that are lower than the set value on a certain dimension, which can reflect the concentration degree, that is, a large number of long-term browsing information is concentrated at this clustering center), several target dimensions with the largest sixth eigenvalue values are determined (for example, the top three are selected); for each target dimension: target optimized long-term vectors with values lower than the set value on the current target dimension are determined, and the mean and standard deviation of all target optimized long-term vectors on the current target dimension on each dimension are calculated as the supplementary eigenvalue on the corresponding dimension; for each dimension: based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, sixth eigenvalue, and all supplementary eigenvalues on this dimension, the feature on this dimension is determined. Thus, not only the overall distribution of the user's long-term browsing data is considered, but also the potentially personalized features of the user are considered, which can further improve the reflection ability of the long-term browsing features, so as to more accurately match similar users for the user (perform user-based collaborative filtering) and improve the information recommendation effect.

[0095] In the feature extraction of real-time browsing data, each real-time browsing information in the real-time browsing data is preprocessed. Each real-time browsing information corresponds to an information information. The real-time browsing information includes the time node, browsing duration, browsing entrance, operation frequency, bounce rate, interactive behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click-through conversion rate of the user browsing the corresponding information information. Feature extraction is performed to determine the browsing feature component, and the browsing feature component is fused with the title feature component, content feature component, label feature component, and heat feature component of the corresponding information information to obtain the real-time browsing vector corresponding to this real-time browsing information. The real-time browsing information contains more detailed indicators (compared with long-term browsing information), which can well reflect the user's real-time reading and operation situation, and the corresponding information information contains richer feature information (including title feature component, content feature component, label feature component, heat feature component, etc., which can be obtained by obtaining relevant data of the corresponding information information through the platform), so as to better reflect the user's current status and interests. Afterwards, the real-time browsing vectors corresponding to all the real-time browsing information are input into the second cluster analysis model (obtained by training using the K-Means clustering algorithm), and the distance between each real-time browsing vector and each cluster center is output as the optimized real-time vector corresponding to each real-time browsing vector; based on each optimized real-time vector, the overall distribution characteristics of all optimized real-time vectors are extracted as real-time browsing features. For each dimension in the optimized real-time vector: the mean and standard deviation of all optimized real-time vectors in this dimension are calculated respectively, and the first feature quantity and the second feature quantity of this dimension are obtained; the number of optimized real-time vectors with the minimum absolute value and the maximum absolute value in this dimension are determined respectively, and the third feature quantity and the fourth feature quantity of this dimension are obtained respectively; the number of optimized real-time vectors with values lower than the mean and lower than the set value in this dimension are determined respectively, and the fifth feature quantity and the sixth feature quantity of this dimension are obtained respectively; based on the first feature quantity, the second feature quantity, the third feature quantity, the fourth feature quantity, the fifth feature quantity and the sixth feature quantity in this dimension, the features in this dimension are determined, and the features in all dimensions are further integrated to obtain the real-time browsing features. In this way, not only can the second clustering analysis model be used to perform cluster analysis on different quantities of real-time browsing information of the user to further explore potential features, but also distribution statistics can be performed based on the optimized real-time vectors output corresponding to each real-time browsing information (which can also be reflected in the form of quantity or percentage, and normalization can be performed when using quantity for statistics), so as to overcome the cold start problem caused by the small amount of data as much as possible.

[0096] When performing user matching and information recommendation, calculate the preference similarity between the user preference features of this user and those of other users; calculate the long-term similarity between the long-term browsing features of this user and those of other users; calculate the real-time similarity between the real-time browsing features of this user and those of other users; based on the preference similarity, long-term similarity, and real-time similarity, calculate the user similarity between this user and other users; based on the user similarity corresponding to each other user, determine the target users. In this way, the characteristics in multiple aspects can be integrated to determine the target users who are more matched in all aspects, which is very suitable for platforms with an increasingly large number of users. When performing information recommendation, obtain the information recommendation list of each target user in the recent period; count the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each piece of information, calculate the recommendation index of each piece of information, and determine the target information to be recommended. In this way, information can be better recommended to users, effectively improving the recommendation effect.

[0097] In this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0098] The above are only examples of the present application and are not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that, Including: Obtain user preference data and user browsing data. Among them, the user preference data reveals the data types of user preferences, and the user browsing data includes long-term browsing data and real-time browsing data. The long-term browsing data is used to reveal the long-term browsing behavior of the user, and the real-time browsing data is used to reveal the real-time browsing behavior of the user; Determine the user characteristics of this user according to the user preference data and the user browsing data; Match the user characteristics of this user with the user characteristics of other users to determine the target user; Based on the target user, determine the target information to be recommended and recommend the target information to this user; Determine the user characteristics of this user according to the user preference data and the user browsing data, including: Perform one-hot encoding on the user preference data to obtain user preference features; Perform preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine the long-term browsing features; Perform preprocessing and feature extraction on the real-time browsing data in the user browsing data to determine the real-time browsing features; Based on the user preference features, long-term browsing features and real-time browsing features, combine them into user characteristics; The long-term browsing data contains several pieces of long-term browsing information. Perform preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine the long-term browsing features, including: Perform preprocessing on each piece of long-term browsing information in the long-term browsing data. Each piece of long-term browsing information corresponds to a piece of news information, and the long-term browsing information includes the time node, browsing duration, browsing entry and operation frequency of the user browsing the corresponding news information; For each piece of long-term browsing information after preprocessing: perform feature extraction on the time node, browsing duration, browsing entry and operation frequency in this piece of long-term browsing information to determine the browsing feature components, and fuse the browsing feature components with the label feature components of the corresponding news information to obtain the long-term browsing vector corresponding to this piece of long-term browsing information; Input the long-term browsing vectors corresponding to all long-term browsing information into the first clustering analysis model, and output the distance between each long-term browsing vector and each clustering center as the optimized long-term vector corresponding to each long-term browsing vector; Based on each optimized long-term vector, extract the overall distribution characteristics of all optimized long-term vectors as the long-term browsing features; Based on each optimized long-term vector, extract the overall distribution characteristics of all optimized long-term vectors as the long-term browsing features, including: For each dimension in the optimized long-term vector: Calculate the mean and standard deviation of all optimized long-term vectors in this dimension respectively, and correspondingly obtain the first eigenvalue and the second eigenvalue of this dimension; Determine the number of optimized long-term vectors with the minimum absolute value and the maximum absolute value in this dimension respectively, and correspondingly obtain the third eigenvalue and the fourth eigenvalue of this dimension; Determine the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension respectively, and correspondingly obtain the fifth eigenvalue and the sixth eigenvalue of this dimension, where the set value represents the mean calculated from the minimum absolute value elements of each optimized long-term vector; Based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, and sixth eigenvalue in this dimension, determine the feature in this dimension, and further integrate the features in all dimensions to obtain the long-term browsing feature; Based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, and sixth eigenvalue in this dimension, determine the feature in this dimension, including: Based on the sixth eigenvalue in each dimension, determine several target dimensions with the largest sixth eigenvalue value; For each target dimension: determine the target optimized long-term vectors with values lower than the set value in the current target dimension, and calculate the mean and standard deviation of all target optimized long-term vectors in each dimension in the current target dimension as the supplementary eigenvalue in the corresponding dimension; For each dimension: based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, sixth eigenvalue, and all supplementary eigenvalues in this dimension, determine the feature in this dimension.

2. The information recommendation method based on artificial intelligence according to claim 1, wherein The real-time browsing data contains several real-time browsing information in the recent period. Preprocess and extract features from the real-time browsing data in the user browsing data to determine the real-time browsing feature, including: Preprocess each real-time browsing information in the real-time browsing data. Each real-time browsing information corresponds to a piece of news information, and the real-time browsing information includes the time node, browsing duration, browsing entrance, operation frequency, bounce rate, interaction behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click conversion rate of the user browsing the corresponding news information; For each preprocessed real-time browsing information: extract features from the time node, browsing duration, browsing entrance, operation frequency, bounce rate, interaction behavior, subscription status, reading completion rate, repeat visit rate, page scroll depth, and click conversion rate in this real-time browsing information to determine the browsing feature component, and fuse the browsing feature component with the title feature component, content feature component, tag feature component, and popularity feature component of the corresponding news information to obtain the real-time browsing vector corresponding to this real-time browsing information; Input the real-time browsing vectors corresponding to all real-time browsing information into the second clustering analysis model, and output the distance between each real-time browsing vector and each clustering center as the optimized real-time vector corresponding to each real-time browsing vector; Based on each optimized real-time vector, extract the overall distribution feature of all optimized real-time vectors as the real-time browsing feature.

3. The information recommendation method based on artificial intelligence according to claim 2, wherein Based on each optimized real-time vector, extract the overall distribution feature of all optimized real-time vectors as the real-time browsing feature, including: For each dimension in the optimized real-time vector: Calculate the mean and standard deviation of all optimized real-time vectors in this dimension respectively, and correspondingly obtain the first feature quantity and the second feature quantity in this dimension; Determine the number of optimized real-time vectors with the smallest absolute value and the largest absolute value in this dimension respectively, and correspondingly obtain the third feature quantity and the fourth feature quantity in this dimension; Determine the number of optimized real-time vectors with values lower than the mean and lower than the set value in this dimension respectively, and correspondingly obtain the fifth feature quantity and the sixth feature quantity in this dimension; Based on the first feature quantity, second feature quantity, third feature quantity, fourth feature quantity, fifth feature quantity, and sixth feature quantity in this dimension, the features in this dimension are determined, and further, the features in all dimensions are integrated to obtain the real-time browsing features.

4. The information recommendation method based on artificial intelligence according to claim 1, wherein Match the user features of this user with the user features of other users to determine the target users, including: For each other user: Calculate the similarity between the user preference features of this user and the user preference features of this other user to obtain the preference similarity; Calculate the similarity between the long-term browsing features of this user and the long-term browsing features of this other user to obtain the long-term similarity; Calculate the similarity between the real-time browsing features of this user and the real-time browsing features of this other user to obtain the real-time similarity; Based on the preference similarity, long-term similarity, and real-time similarity, calculate the user similarity between this user and this other user; Based on the user similarity corresponding to each other user, determine the target users.

5. The information recommendation method based on artificial intelligence according to claim 1, wherein Based on the target users, determine the target information to be recommended, including: Obtain the information recommendation list of each target user in the recent period; Count the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each information; Based on the number of recommendation times, click-through rate, bounce rate, reading completion rate, repeat visit rate, and click conversion rate of each information, calculate the recommendation index of each information; Based on the recommendation index of each information, determine the target information to be recommended.

6. An information recommendation system based on artificial intelligence, characterized in that, Including: A data acquisition module for acquiring user preference data and user browsing data. Among them, the user preference data reveals the data types of user preferences, and the user browsing data includes long-term browsing data and real-time browsing data. The long-term browsing data is used to reveal the long-term browsing behavior of the user, and the real-time browsing data is used to reveal the real-time browsing behavior of the user; A feature extraction module for determining the user features of this user according to the user preference data and the user browsing data; A user matching module for matching the user features of this user with the user features of other users to determine the target users; An information recommendation module for determining the target information to be recommended based on the target users and recommending the target information to this user; The feature extraction module is specifically used for: performing one-hot encoding on the user preference data to obtain user preference features; preprocessing and feature extraction on the long-term browsing data in the user browsing data to determine the long-term browsing features; preprocessing and feature extraction on the real-time browsing data in the user browsing data to determine the real-time browsing features; combining the user preference features, long-term browsing features, and real-time browsing features into user features; The long-term browsing data contains several pieces of long-term browsing information. The feature extraction module is specifically used for: preprocessing each piece of long-term browsing information in the long-term browsing data, where each piece of long-term browsing information corresponds to a piece of news information, and the long-term browsing information includes the time node, browsing duration, browsing entry, and operation frequency of the user browsing the corresponding news information; for each piece of long-term browsing information after preprocessing: extracting features from the time node, browsing duration, browsing entry, and operation frequency in this piece of long-term browsing information to determine the browsing feature components, and fusing the browsing feature components with the label feature components of the corresponding news information to obtain the long-term browsing vector corresponding to this piece of long-term browsing information; inputting the long-term browsing vectors corresponding to all long-term browsing information into the first clustering analysis model, and outputting the distances between each long-term browsing vector and each cluster center as the optimized long-term vectors corresponding to each long-term browsing vector; based on each optimized long-term vector, extracting the overall distribution features of all optimized long-term vectors as the long-term browsing features; The feature extraction module is specifically used for: for each dimension in the optimized long-term vector: calculating the mean and standard deviation of all optimized long-term vectors in this dimension respectively to obtain the first eigenvalue and the second eigenvalue of this dimension; determining the number of optimized long-term vectors with the minimum absolute value and the maximum absolute value in this dimension respectively to obtain the third eigenvalue and the fourth eigenvalue of this dimension; determining the number of optimized long-term vectors with values lower than the mean and lower than the set value in this dimension respectively to obtain the fifth eigenvalue and the sixth eigenvalue of this dimension, where the set value represents the mean calculated from the minimum absolute value elements of each optimized long-term vector; based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, and sixth eigenvalue of this dimension, determining the feature of this dimension, and further integrating the features of all dimensions to obtain the long-term browsing features; The feature extraction module is specifically used for: based on the sixth eigenvalue of each dimension, determining several target dimensions with the largest sixth eigenvalue value; for each target dimension: determining the target optimized long-term vectors with values lower than the set value in the current target dimension, and statistically calculating the mean and standard deviation of all target optimized long-term vectors in each dimension in the current target dimension as the supplementary eigenvalue of the corresponding dimension; for each dimension: based on the first eigenvalue, second eigenvalue, third eigenvalue, fourth eigenvalue, fifth eigenvalue, sixth eigenvalue, and all supplementary eigenvalues of this dimension, determining the feature of this dimension.

Images