Information pushing method and device, computer device and computer readable storage medium

Abstract

This invention relates to the field of artificial intelligence technology and discloses an information push method. The method includes: classifying all viewed content by all users to obtain multiple content categories; calculating a first probability that a target user will view each of the content to be recommended based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the target user's viewed content; calculating a second probability that each content to be recommended will be recommended to the target user based on the comment information of the user category to which the target user belongs; obtaining a popularity rating for each content to be recommended based on the first and second probabilities; and recommending the content to the target user based on the popularity rating. Through this method, this invention improves the accuracy of content recommendation.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, specifically to an information push method, apparatus, computer device, and computer-readable storage medium. Background Technology

[0002] Currently, content is typically pushed through collaborative filtering by analyzing the similarity of content based on user relationships or content viewed by a user's friends, or simply based on trending topics.

[0003] In the process of implementing the embodiments of the present invention, the inventors discovered that existing methods for analyzing the probability of a user viewing content do not comprehensively consider the relationship between the user and the content, the relationship between users, and the user's evaluation of the effectiveness of the viewed content, resulting in low accuracy of information recommendation. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide an information push method, apparatus, computer device, and computer-readable storage medium to solve the technical problem of low accuracy of information recommendation in the prior art.

[0005] According to one aspect of the present invention, an information push method is provided, the method comprising:

[0006] The viewed content of all users is categorized into multiple content categories;

[0007] Based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user, a first probability is calculated for the target user to view each of the content to be recommended; the content to be recommended is any one of the multiple content categories excluding the content already viewed by the target user.

[0008] Based on the comment information of the content to be recommended to the target user according to the user category to which the target user belongs, calculate the second probability that each piece of content to be recommended is recommended to the target user;

[0009] Based on the first probability and the second probability, the popularity score of each piece of content to be recommended is obtained;

[0010] The content to be recommended to the target user is based on the popularity rating.

[0011] In one optional approach, classifying all viewed content by all users to obtain multiple content categories includes: selecting K content as category centers from all viewed content; calculating the correlation between all viewed content and each of the category centers; obtaining a correlation matrix between the viewed content based on the correlation; iteratively updating the category centers and the correlation matrix until an iteration threshold is met to obtain multiple content categories.

[0012] In one alternative approach, calculating the first probability that the target user will watch each of the content to be recommended based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user includes: determining each of the content already viewed by the target user and the category center of each of the content categories; and calculating the target user's preference for each of the content categories based on the target user's viewed content and the category center of each of the content categories.

[0013] In one alternative approach, a first probability that the target user will view each of the content to be recommended is calculated based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user. This includes: determining a first feature vector of the content already viewed by the target user; determining a second feature vector of the content to be recommended; and determining the similarity between the content to be recommended and the content already viewed by the target user based on the first feature vector and the second feature vector.

[0014] In one optional approach, the step of calculating a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended according to the user category to which the target user belongs includes: obtaining viewing information of all users; obtaining feature vectors of each user based on the viewing information and a word vector model; and performing clustering based on the feature vectors to obtain the user category to which each user belongs.

[0015] In one optional approach, calculating the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the user category to which the target user belongs includes: determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs based on the comment information of the content to be recommended; calculating the Euclidean distance between the target user and other users in the user category; and calculating the second probability that each piece of content to be recommended is recommended to the target user based on the Euclidean distance and the sentiment classification.

[0016] In one optional approach, determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs, based on the comment information of the content to be recommended by the user category to which the target user belongs, includes: obtaining the comment information of the content to be recommended by the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs; wherein the sentiment classification model is pre-trained based on sentiment classification samples.

[0017] According to another aspect of the present invention, an information push device is provided, comprising:

[0018] The content categorization module is used to classify all the content viewed by all users, resulting in multiple content categories;

[0019] The first probability calculation module is used to calculate the first probability that the target user will watch each of the recommended contents based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user; the content to be recommended is any one of the multiple content categories except the content already viewed by the target user;

[0020] The second probability calculation module is used to calculate the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended by the user category to which the target user belongs;

[0021] The popularity rating module is used to obtain a popularity rating for each piece of content to be recommended based on the first probability and the second probability.

[0022] The recommendation module is used to recommend the content to target users based on the popularity score.

[0023] According to another aspect of the present invention, a computer device is provided, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0024] The memory is used to store at least one executable instruction, which causes the processor to perform the above-described information push method.

[0025] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the operation of the above-described information push method.

[0026] This invention, through its embodiments, categorizes all viewed content by users into multiple content categories. Then, based on the user's preference for each content category and the similarity between the content to be recommended and the target user's viewed content, a first probability is calculated for the target user to view each of the content to be recommended. Next, based on the comment information of the target user's user category regarding the content to be recommended, a second probability is calculated for each content to be recommended to the target user. Based on the first and second probabilities, a popularity score is obtained for each content to be recommended. Finally, the content to be recommended is recommended to the target user based on the popularity score, effectively improving the accuracy of information recommendation.

[0027] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0028] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0029] Figure 1 A flowchart illustrating the information push method provided in an embodiment of the present invention is shown;

[0030] Figure 2 A schematic diagram of the information push device provided in an embodiment of the present invention is shown;

[0031] Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0032] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0033] Currently, content recommendation relies on collaborative filtering based on user relationship content similarity or content viewed by a user's friends, or simply on trending topics. Existing solutions primarily generate content feature vectors based on user ratings, neglecting factors like playback time and training effects. When identifying trending content, they simply count the number of trending keywords, failing to account for content without such keywords. Alternatively, they calculate trending content based on user search behavior, again relying on quantitative statistics. Generating trending content based on quantity doesn't comprehensively reflect content popularity; user viewing behavior provides a more accurate picture. Furthermore, analyzing the probability of a user viewing content fails to comprehensively consider relationships between users and content, relationships between users, and the effectiveness of viewed content.

[0034] Figure 1 A flowchart of an information push method provided in an embodiment of the present invention is shown. This method is executed by a computer device. The computer device may be a computer, tablet computer, mobile phone, watch, audio / video playback device, wearable device, etc., and the embodiments of the present invention do not impose specific limitations. Figure 1 As shown, the method includes the following steps:

[0035] Step 110: Categorize all the content viewed by all users to obtain multiple content categories.

[0036] In this embodiment of the invention, before classifying the viewed content of all users, user behavior data is pre-acquired and processed to obtain the viewed content for each user. Specifically, user behavior logs are sent to Kafka through the business backend, and user behavior data from the Kafka (distributed streaming media platform) user behavior logs is collected through Flume (a log collection system), and then filtered and extracted. Interceptors can be set in Flume to filter abnormal data in the user behavior data, such as empty usernames or abnormal field values. After filtering, data information is extracted using Spark Streaming (a real-time computing framework built on Spark). First, user viewing behavior data is obtained, and then viewing information such as user ID, viewed content ID, content playback duration, and content playback time are extracted from the viewing behavior data. Finally, the obtained viewing information is stored in a Hive (data warehouse tool) database.

[0037] After obtaining all the viewed content from all users, the viewed content is converted into feature vectors V using a word vector model. After obtaining the feature vectors V for all viewed content, these feature vectors V are used to classify all viewed content, resulting in multiple content categories. This includes the following steps:

[0038] a. Select K content items from all viewed content as category centers.

[0039] b. Calculate the correlation between all viewed content and each of the aforementioned category centers. Specifically, calculate the correlation between the i-th viewed content and each of the selected category centers. i With each classification center v j Correlation:

[0040]

[0041] Where, p ij c represents the viewed content of all users for the i-th element. i With the j-th classification center v j The correlation; d(c i ,v j ) represents the i-th viewed content among all users' viewed content. i With the j-th classification center v j The Euclidean distance; t is a variable, taking values ​​from 0 to k.

[0042] c. Based on the aforementioned correlations, obtain the correlation matrix among the viewed content. This correlation matrix is ​​a k*n matrix with viewed content as columns and category centers as rows, where n is the number of viewed content items for all users.

[0043] d. Iteratively update the classification center and the correlation matrix until the iteration threshold is met to obtain multiple content classifications.

[0044] After obtaining the correlation matrix, the classification centers are updated using the following method:

[0045]

[0046] Where, p ij This represents the relevance between the i-th viewed content and the j-th viewed content as category centers; d j Let be the feature vector of the viewed content j, n represent the total number of viewed content items for all users, and i represent the i-th viewed content item.

[0047] By iteratively updating the classification centers, the iteration stops when the Euclidean distance between the current classification center and the previous classification center is less than the threshold e, resulting in k content classifications.

[0048] Step 120: Calculate the first probability that the target user will watch each of the recommended contents based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user.

[0049] The content to be recommended is any one of the multiple content categories, excluding the content already viewed by the target user.

[0050] Specifically, after obtaining k content categories, the target user's viewed content and the category center of each content category are determined; based on the target user's viewed content and the category centers of each content category, the target user's preference for each content category is calculated. Specifically, the calculation method can be:

[0051]

[0052] Among them, I(u,c i ) represents the target user u's choice of the i-th content category c out of k content categories. i The degree of preference; s represents the total number of views of content corresponding to the target user u; X j This represents the j-th content that the target user u has viewed.

[0053] In this embodiment of the invention, the similarity between the content to be recommended and each viewed content of the target user u is also calculated, including: determining a first feature vector of the viewed content of the target user; determining a second feature vector of the content to be recommended; and determining the similarity between the content to be recommended and the viewed content of the target user based on the first feature vector and the second feature vector. Specifically, the calculation formula can be:

[0054]

[0055] Among them, W ij V represents the similarity between the content category of the i-th viewed content and the j-th content to be recommended among the s viewed content of the target user u; im V represents the feature vector of the i-th content viewed by the target user u; im Let represent the feature vector of the j-th content to be recommended. After obtaining the user's preference level and the similarity between the content to be recommended and the content already viewed by the target user, the first probability of the target user viewing each of the content to be recommended is calculated based on the target user's preference level for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user. Specifically, this includes the following methods:

[0056]

[0057] Where, p uj1 The first probability that target user u will watch the j-th recommended content; s is the number of content items that target user u has already watched; W ij Let P be the similarity between the i-th viewed content and the j-th content to be recommended by target user u. When the content to be recommended is content already viewed by target user u, P... uj1 =1, meaning the probability of watching is 1. j The category center for the content category to which the j-th content to be recommended belongs.

[0058] Step 130: Calculate the second probability that each piece of content to be recommended will be recommended to the target user based on the comment information of the target user's user category.

[0059] In this embodiment of the invention, user pre-classification includes: acquiring viewing information of all users; obtaining feature vectors for each user based on the viewing information and a word vector model; and clustering the feature vectors to obtain the user category to which each user belongs. Specifically, the feature vectors W of all users can be obtained by training word2Vec. After obtaining the user feature vectors W, all feature vectors are clustered. Specifically, steps a to d, similar to those in the content classification described above, can be used to calculate the group classification of all users, resulting in c user categories.

[0060] After obtaining the user category to which the target user belongs, the sentiment classification of the content to be recommended by other users in the target user's user category is determined based on the comment information of the target user's user category regarding the content to be recommended. The Euclidean distance between the target user and other users in the target user's user category is calculated. Based on the Euclidean distance and the sentiment classification, a second probability is calculated for each piece of content to be recommended to the target user. Specifically, determining the sentiment classification of the content to be recommended by other users in the target user's user category based on the comment information of the target user's user category is achieved as follows: obtaining the comment information of the user's user category regarding the content to be recommended; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the target user's user category; the sentiment classification model is pre-trained based on sentiment classification samples.

[0061] In this embodiment of the invention, the process of training the sentiment classification model is as follows: The sentiment classification model is trained using the BERT model (an open-source pre-trained model for natural language understanding). During the pre-training process of BERT, the method of randomly removing one or more words in the Masked LM task is modified to Mask specific words, allowing Masked LM to learn the representation of words in context. The calculation method is as follows: Sub-step a: First, select a subset of seed words and label their parts of speech (positive or negative). For example, positive words include "like," "good," etc., while negative words include "dislike," "bad," "awkward," etc. Sub-step b: Mine more sentiment attribute words from the selected seed words. The mining method is as follows: First, segment all comment content using an open-source word segmentation tool (Stanford CoreNLP) and obtain the part of speech for each word. After obtaining the set of all words and the part of speech for each word, obtain all adjectives. Then, calculate the correlation between all adjectives and the selected seed words. The calculation formula is as follows: Where P(w1,w2) represents the probability of words w1 and w2 appearing simultaneously; P(w1) represents the probability of word w1 appearing; and P(w2) represents the probability of word w2 appearing. The relevance of each word to the selected seed word is calculated using this formula, yielding the relevance PP to positive words and the relevance PN to negative words. Then, the difference between PP and PN (PP-PN) is calculated. If the difference is positive, the adjective is positive; if negative, it is negative. This process extracts all positive and negative adjectives from all words. Sub-step c: Using the part-of-speech tags of all words obtained in sub-step b, the nouns and adjectives in each comment are selected and paired. Sub-step d: The sample comment information (with special symbols removed) from the training data is pre-trained using the BERT pre-trained model. BERT is a multi-task model whose tasks consist of two self-supervised tasks: MLM and NSP. MLM refers to randomly masking some words from the input corpus during training and then predicting the word based on the context. In the MLM task, the adjectives and noun-adjective word pairs with parts of speech obtained in sub-steps b and c are masked. The remaining training steps are the same as the training process of the BERT model, thus obtaining a trained sentiment classification model.

[0062] After obtaining the sentiment classification model, the sentiment classification of each user's comments on each viewed content (i.e., content to be recommended) is calculated based on the obtained sentiment classification model. A "positive" sentiment classification indicates a recommendation, while a "negative" sentiment classification indicates a non-recommendation.

[0063] After determining the sentiment category, calculate the Euclidean distance u between the target user u and other users b in the same user category. ubBased on the Euclidean distance and the sentiment classification, a second probability is calculated for each piece of content to be recommended to the target user. The specific calculation formula can be expressed as:

[0064]

[0065] Where, p uj2 This represents the second probability of recommending the j-th content to target user u; U ub Let be the Euclidean distance between target user u and user b; r be user b's sentiment classification towards recommended content j, where 1 represents a recommendation, -1 a dislike, and 0 a no-comment. q be the number of users in the user category to which target user u belongs.

[0066] Step 140: Based on the first probability and the second probability, obtain the popularity score for each piece of content to be recommended.

[0067] After obtaining the first and second probabilities, the target probability for user u to view the j-th recommended content is determined according to the following formula:

[0068]

[0069] After obtaining the target probability, the popularity score of each piece of content to be recommended is determined based on the target probability. The specific popularity score can be calculated using the following formula:

[0070]

[0071] Among them, Score j Let C be the popularity score of the j-th content to be recommended, and C be the total number of all users.

[0072] Step 150: Recommend the content to be submitted to the target user based on the popularity score.

[0073] After obtaining the popularity scores of each piece of content to be recommended, the most popular content can be identified based on these scores and recommended to the corresponding target users. Specifically, the content to be recommended can be sorted according to its popularity score, and the top N pieces can be recommended to the relevant target users.

[0074] This invention categorizes all viewed content by users into multiple content categories. Then, based on the target user's preference for each content category and the similarity between the content to be recommended and the target user's viewed content, a first probability is calculated for the target user to view each of the content to be recommended. Next, based on the comment information of the target user's user category regarding the content to be recommended, a second probability is calculated for each content to be recommended to the target user. Based on the first and second probabilities, a popularity score is obtained for each content to be recommended. Finally, the content to be recommended is recommended to the target user based on the popularity score, effectively improving the accuracy of information recommendation.

[0075] Figure 2 A schematic diagram of the structure of the information push device provided in an embodiment of the present invention is shown. Figure 2 As shown, the device 200 includes:

[0076] The content classification module 210 is used to classify the content viewed by all users, resulting in multiple content categories;

[0077] The first probability calculation module 220 is used to calculate the first probability that the target user will watch each of the recommended contents based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user; the content to be recommended is any one of the multiple content categories except the content already viewed by the target user.

[0078] The second probability calculation module 230 is used to calculate the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended to the user belonging to the user category of the target user.

[0079] The popularity rating module 240 is used to obtain a popularity rating for each piece of content to be recommended based on the first probability and the second probability.

[0080] The recommendation module 250 is used to recommend the content to the target user based on the popularity score.

[0081] In one optional approach, classifying all viewed content by all users to obtain multiple content categories includes: selecting K content as category centers from all viewed content; calculating the correlation between all viewed content and each of the category centers; obtaining a correlation matrix between the viewed content based on the correlation; iteratively updating the category centers and the correlation matrix until an iteration threshold is met to obtain multiple content categories.

[0082] In one alternative approach, calculating the first probability that the target user will watch each of the content to be recommended based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user includes: determining each of the content already viewed by the target user and the category center of each of the content categories; and calculating the target user's preference for each of the content categories based on the target user's viewed content and the category center of each of the content categories.

[0083] In one alternative approach, a first probability that the target user will view each of the content to be recommended is calculated based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user. This includes: determining a first feature vector of the content already viewed by the target user; determining a second feature vector of the content to be recommended; and determining the similarity between the content to be recommended and the content already viewed by the target user based on the first feature vector and the second feature vector.

[0084] In one optional approach, the step of calculating a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended according to the user category to which the target user belongs includes: obtaining viewing information of all users; obtaining feature vectors of each user based on the viewing information and a word vector model; and performing clustering based on the feature vectors to obtain the user category to which each user belongs.

[0085] In one optional approach, calculating the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the user category to which the target user belongs includes: determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs based on the comment information of the content to be recommended; calculating the Euclidean distance between the target user and other users in the user category; and calculating the second probability that each piece of content to be recommended is recommended to the target user based on the Euclidean distance and the sentiment classification.

[0086] In one optional approach, determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs, based on the comment information of the content to be recommended by the user category to which the target user belongs, includes: obtaining the comment information of the content to be recommended by the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs; wherein the sentiment classification model is pre-trained based on sentiment classification samples.

[0087] The specific working process of the information push device in this embodiment of the invention is largely the same as the specific process steps of the above method embodiment, and will not be repeated here.

[0088] This invention categorizes all viewed content by users into multiple content categories. Then, based on the target user's preference for each content category and the similarity between the content to be recommended and the target user's viewed content, a first probability is calculated for the target user to view each of the content to be recommended. Next, based on the comment information of the target user's user category regarding the content to be recommended, a second probability is calculated for each content to be recommended to the target user. Based on the first and second probabilities, a popularity score is obtained for each content to be recommended. Finally, the content to be recommended is recommended to the target user based on the popularity score, effectively improving the accuracy of information recommendation.

[0089] Figure 3 The diagram shows a structural schematic of a computer device provided in an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0090] like Figure 3 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0091] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other network elements such as clients or other servers. The processor 402 executes program 410, specifically performing the relevant steps described above in the information push method embodiment.

[0092] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0093] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0094] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0095] Specifically, program 410 can be called by processor 402 to cause the computer device to perform the following operations:

[0096] The viewed content of all users is categorized into multiple content categories;

[0097] Based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the user, a first probability is calculated for the target user to view each of the content to be recommended; the content to be recommended is any one of the multiple content categories excluding the content already viewed by the target user.

[0098] Based on the comment information of the content to be recommended to the target user's user category, calculate the second probability that each piece of content to be recommended will be recommended to the user.

[0099] Based on the first probability and the second probability, the popularity score of each piece of content to be recommended is obtained;

[0100] The content to be recommended to the target user is based on the popularity rating.

[0101] In one optional approach, classifying all viewed content by all users to obtain multiple content categories includes: selecting K content as category centers from all viewed content; calculating the correlation between all viewed content and each of the category centers; obtaining a correlation matrix between the viewed content based on the correlation; iteratively updating the category centers and the correlation matrix until an iteration threshold is met to obtain multiple content categories.

[0102] In one optional approach, calculating the first probability that the target user will watch each of the content to be recommended based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the user includes: determining each of the target user's viewed content and the category center of each of the content categories; and calculating the target user's preference for each of the content categories based on the target user's viewed content and the category center of each of the content categories.

[0103] In one alternative approach, a first probability that the target user will view each of the content to be recommended is calculated based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user. This includes: determining a first feature vector of the content already viewed by the target user; determining a second feature vector of the content to be recommended; and determining the similarity between the content to be recommended and the content already viewed by the target user based on the first feature vector and the second feature vector.

[0104] In one optional approach, the step of calculating a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended according to the user category to which the target user belongs includes: obtaining viewing information of all users; obtaining feature vectors of each user based on the viewing information and a word vector model; and performing clustering based on the feature vectors to obtain the user category to which each user belongs.

[0105] In one optional approach, calculating the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended by the user category to which the target user belongs includes: determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs based on the comment information of the content to be recommended by the user category to which the target user belongs; calculating the Euclidean distance between the user and other users in the user category; and calculating the second probability that each piece of content to be recommended is recommended to the target user based on the Euclidean distance and the sentiment classification.

[0106] In one optional approach, determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs, based on the comment information of the content to be recommended by the user category to which the target user belongs, includes: obtaining the comment information of the content to be recommended by the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs; wherein the sentiment classification model is pre-trained based on sentiment classification samples.

[0107] This invention categorizes all viewed content by users into multiple content categories. Then, based on the target user's preference for each content category and the similarity between the content to be recommended and the target user's viewed content, a first probability is calculated for the target user to view each of the content to be recommended. Next, based on the comment information of the target user's user category regarding the content to be recommended, a second probability is calculated for each content to be recommended to the target user. Based on the first and second probabilities, a popularity score is obtained for each content to be recommended. Finally, the content to be recommended is recommended to the target user based on the popularity score, effectively improving the accuracy of information recommendation.

[0108] This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a computer device, causes the computer device to perform the information push method described in any of the above method embodiments.

[0109] Executable instructions can be used to cause computer devices to perform the following operations:

[0110] The viewed content of all users is categorized into multiple content categories;

[0111] Based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user, a first probability is calculated for the target user to view each of the content to be recommended; the content to be recommended is any one of the multiple content categories excluding the content already viewed by the target user.

[0112] Based on the comment information of the content to be recommended to the target user according to the user category to which the target user belongs, calculate the second probability that each piece of content to be recommended is recommended to the target user;

[0113] Based on the first probability and the second probability, the popularity score of each piece of content to be recommended is obtained;

[0114] The content to be recommended to the target user is based on the popularity rating.

[0115] In one optional approach, classifying all viewed content by all users to obtain multiple content categories includes: selecting K content as category centers from all viewed content; calculating the correlation between all viewed content and each of the category centers; obtaining a correlation matrix between the viewed content based on the correlation; iteratively updating the category centers and the correlation matrix until an iteration threshold is met to obtain multiple content categories.

[0116] In one alternative approach, calculating the first probability that a user will watch each of the content to be recommended based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user includes: determining each of the target user's viewed content and the category center of each of the content categories; and calculating the target user's preference for each of the content categories based on the target user's viewed content and the category center of each of the content categories.

[0117] In one alternative approach, a first probability that the target user will view each of the content to be recommended is calculated based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the user. This includes: determining a first feature vector of the content already viewed by the target user; determining a second feature vector of the content to be recommended; and determining the similarity between the content to be recommended and the content already viewed by the target user based on the first feature vector and the second feature vector.

[0118] In one optional approach, the step of calculating a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended according to the user category to which the target user belongs includes: obtaining viewing information of all users; obtaining feature vectors of each user based on the viewing information and a word vector model; and performing clustering based on the feature vectors to obtain the user category to which each user belongs.

[0119] In one optional approach, calculating the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the user category to which the target user belongs includes: determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs based on the comment information of the content to be recommended; calculating the Euclidean distance between the target user and other users in the user category; and calculating the second probability that each piece of content to be recommended is recommended to the target user based on the Euclidean distance and the sentiment classification.

[0120] In one optional approach, determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs, based on the comment information of the content to be recommended by the user category to which the target user belongs, includes: obtaining the comment information of the content to be recommended by the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs; wherein the sentiment classification model is pre-trained based on sentiment classification samples.

[0121] This invention categorizes all viewed content by users into multiple content categories. Then, based on the target user's preference for each content category and the similarity between the content to be recommended and the target user's viewed content, a first probability is calculated for the target user to view each of the content to be recommended. Next, based on the comment information of the target user's user category regarding the content to be recommended, a second probability is calculated for each content to be recommended to the target user. Based on the first and second probabilities, a popularity score is obtained for each content to be recommended. Finally, the content to be recommended is recommended to the target user based on the popularity score, effectively improving the accuracy of information recommendation.

[0122] This invention provides an information push device for executing the above-described information push method.

[0123] This invention provides a computer program that can be called by a processor to cause a computer device to execute the information push method in any of the above method embodiments.

[0124] This invention provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed on a computer, cause the computer to perform the information push method described in any of the above method embodiments.

[0125] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0126] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0127] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0128] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0129] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. An information push method, characterized in that, The method includes: The viewed content of all users is categorized into multiple content categories; Based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user, a first probability is calculated for the target user to view each of the content to be recommended; the content to be recommended is any one of the multiple content categories except for the content already viewed by the target user. Based on the comment information of the content to be recommended to the target user according to the user category to which the target user belongs, calculate the second probability that each piece of content to be recommended is recommended to the target user; The step of calculating the second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the user category to which the target user belongs includes: determining the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs based on the comment information of the content to be recommended, including: obtaining the comment information of the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the user category to which the target user belongs; the sentiment classification model is pre-trained based on sentiment classification samples, wherein the sentiment classification model is trained using the BERT model, and the process of training the sentiment classification model is as follows: during the pre-training of BERT, the method of randomly removing one or more words in the MaskedLM task is modified to Mask specific words, and MaskedLM is used to perform the training. LM learns the representation of words in context; the calculation method is as follows: Sub-step a: First, select a portion of seed words and label their parts of speech; Sub-step b: Mine more sentiment attribute words from the selected seed words, the mining method is as follows: First, segment all comment content using an open-source word segmentation tool and obtain the part of speech of each word; after obtaining the set of all words and the part of speech of each word, obtain all adjectives; then, calculate the correlation between all adjectives and the selected seed words, the calculation formula is as follows: Where P(w1,w2) represents the probability that words w1 and w2 appear simultaneously; P(w1) represents the probability that word w1 appears; P(w2) represents the probability that word w2 appears. The correlation between all words and the selected seed words is calculated using this formula, yielding the correlation PP with positive words and the correlation PN with negative words. The difference between PP and PN is calculated; if the difference is positive, the adjective is positive; if it is negative, it is negative. This process extracts all positive and negative adjectives from all words. Sub-step c: Using the parts of speech of all words obtained in sub-step b, the nouns and adjectives for each comment are selected. The process involves: forming word pairs between nouns and adjectives; sub-step d: pre-training the sample comment information (with special symbols removed) from the training data using the BERT pre-training model. In the BERT pre-training model's MLM task, the adjectives with part-of-speech tags obtained in sub-steps b and c are masked with the noun-adjective word pairs. The remaining training steps are the same as the BERT model's training process, resulting in a trained sentiment classification model; calculating the Euclidean distance between the target user and other users in their respective user category; and calculating the second probability that each piece of content to be recommended will be recommended to the user based on the Euclidean distance and the sentiment classification. Based on the first probability and the second probability, the popularity score of each piece of content to be recommended is obtained; The content to be recommended is recommended to the target user based on the popularity score.

2. The method according to claim 1, characterized in that, The process of categorizing all viewed content by all users yields multiple content categories, including: Select K content items from all viewed content as category centers; Calculate the correlation between all viewed content and each of the aforementioned category centers; Based on the aforementioned correlations, a correlation matrix is ​​obtained among the various viewed content items; The classification center and the correlation matrix are iteratively updated until the iteration threshold is met, resulting in multiple content classifications.

3. The method according to claim 1, characterized in that, The step of calculating the first probability that the target user will view each of the recommended contents based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user includes: Determine the category centers for each piece of content viewed by the target user and for each of the content categories. Based on the target user's viewed content and the category center of each content category, calculate the target user's preference for each content category.

4. The method according to any one of claims 1-3, characterized in that, Based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user, a first probability of the target user viewing each of the content to be recommended is calculated, including: Determine the first feature vector of the content viewed by the target user; Determine the second feature vector of the content to be recommended; Based on the first feature vector and the second feature vector, the similarity between the content to be recommended and the content already viewed by the target user is determined.

5. The method according to any one of claims 1-3, characterized in that, The step of calculating a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the content to be recommended according to the user category to which the target user belongs includes: Get all users' viewing information; Based on the viewing information and word vector model, the feature vectors of each user are obtained; Clustering is performed based on the feature vectors to obtain the user category to which each user belongs.

6. An information push device, characterized in that, The device includes: The content categorization module is used to classify all the content viewed by all users, resulting in multiple content categories; The first probability calculation module is used to calculate the first probability that the target user will watch each of the recommended contents based on the target user's preference for each of the content categories and the similarity between the content to be recommended and the content already viewed by the target user; the content to be recommended is any one of the multiple content categories except the content already viewed by the target user; The second probability calculation module is used to calculate a second probability that each piece of content to be recommended is recommended to the target user based on the comment information of the user category to which the target user belongs. This calculation includes: determining the sentiment classification of the content to be recommended by other users in the target user's user category based on the comment information of the content to be recommended; obtaining the comment information of the user category to which the target user belongs; inputting the comment information into a sentiment classification model to obtain the sentiment classification of the content to be recommended by other users in the target user's user category; the sentiment classification model is pre-trained based on sentiment classification samples, wherein the BERT model is used to train the sentiment classification model. The training process of the sentiment classification model is as follows: during the pre-training of BERT, the method of randomly removing one or more words in the Masked LM task is modified to Mask specific words, and Masked LM is used to perform the training. LM learns the representation of words in context; the calculation method is as follows: Sub-step a: First, select a portion of seed words and label their parts of speech; Sub-step b: Mine more sentiment attribute words from the selected seed words, the mining method is as follows: First, segment all comment content using an open-source word segmentation tool and obtain the part of speech of each word; after obtaining the set of all words and the part of speech of each word, obtain all adjectives; then, calculate the correlation between all adjectives and the selected seed words, the calculation formula is as follows: Where P(w1,w2) represents the probability that words w1 and w2 appear simultaneously; P(w1) represents the probability that word w1 appears; P(w2) represents the probability that word w2 appears. The correlation between all words and the selected seed words is calculated using this formula, yielding the correlation PP with positive words and the correlation PN with negative words. The difference between PP and PN is calculated; if the difference is positive, the adjective is positive; if it is negative, it is negative. This process extracts all positive and negative adjectives from all words. Sub-step c: Using the parts of speech of all words obtained in sub-step b, the nouns and adjectives for each comment are selected. The process involves: forming word pairs between nouns and adjectives; sub-step d: pre-training the sample comment information (with special symbols removed) from the training data using the BERT pre-training model. In the BERT pre-training model's MLM task, the adjectives with part-of-speech tags obtained in sub-steps b and c are masked with the noun-adjective word pairs. The remaining training steps are the same as the BERT model's training process, resulting in a trained sentiment classification model; calculating the Euclidean distance between the target user and other users in their respective user category; and calculating the second probability that each piece of content to be recommended will be recommended to the user based on the Euclidean distance and the sentiment classification. The popularity rating module is used to obtain a popularity rating for each piece of content to be recommended based on the first probability and the second probability. The recommendation module is used to recommend the content to be recommended to the target user based on the popularity score.

7. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation of the information push method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the information push method as described in any one of claims 1-5.

Images