Personalized news information recommendation method and system based on multi-path recall

Abstract

The invention discloses a multi-recall-based personalized news information recommendation method, which comprises the following steps of: when a user is a new user, providing a user interest tag selection function, and if the user selects an interest tag, recommending tag column information according to the user interest tag; if the user refuses the interest label selection, cold start hotspot information recommendation is carried out; when the user is not a new user, judging whether historical behavior data information of the user exists or not, and if the historical behavior data information does not exist, performing interest label judgment according to the new user and performing cold start hot spot information recommendation or column recommendation; and if the historical behavior data exist, carrying out comprehensive information recommendation by integrating similar user preference information recommendation and time sequence content-based information recommendation. According to the method, the defects that an existing personalized information recommendation algorithm is high in dependence on historical data of users in the field of news recommendation, and recommendation is homogenized are overcome.

Description

Technical Field

[0001] The present invention relates to the field of information processing, and in particular to a personalized news information recommendation method and system based on multi-channel recall. Background Art

[0002] Personalized news information recommendation refers to the process of analyzing and mining user behaviors, discovering users' personalized news information needs and interest characteristics, and recommending news information that meets users' intentions or may be of interest to users. The task of news information recommendation is to display and push news information to users through recommendation algorithms under the premise of regular news data collection. In essence, news recommendation is an information filtering system, which is an interactive process of user historical data, personal information, recommendation algorithms and user behavior feedback. Its working principle can be summarized into three steps: (1) Obtain a large amount of user news behavior history data from various channels such as mobile terminals, web pages, and social networks, and generate effective user log data through cleaning and screening; (2) The recommendation algorithm mines user characteristics from user log data, models user interest preferences, and predicts the user's preference for news content; (3) Sort the news content according to the predicted user's preference for news content, and recommend the higher-ranked information content to users. Users browse, click or search for recommended news information, forming new user interaction records, which are used to update user interest preferences. Information recommendation mainly includes two process modules: recall and sorting. Recall is to obtain a set of candidate data that users may be interested in from massive information through different recommendation strategies, and then sort the data through the specified sorting strategy or user click prediction probability, and recommend the top-ranked data to the user.

[0003] Generally, information recommendation methods based solely on a single recall have the following defects: (1) They require a high amount of user historical behavior data. The more historical behavior data a user has, the better the recommendation performance. (2) They are unable to model the interests of cold-start users. (3) Information cocoons are prone to occur, and the recommended information becomes increasingly homogeneous, making it impossible to explore other potential interests of users.

[0004] In view of the defects of the above single recall information recommendation method and the difficulties faced by current information recommendation in specific fields, if a fast news information recommendation method can be provided to avoid information cocoons and tap into users' "new" interests and preferences, it can provide faster and more effective services for user profiling, personalized recommendations, and information retrieval. Summary of the invention

[0005] The purpose of the present invention is to propose a personalized news information recommendation method and system that integrates multiple recall methods such as hot spot recommendation, column recommendation, user similarity recommendation, and time-based content recommendation, so as to overcome the defects of existing personalized information recommendation algorithms in the field of news recommendation, such as strong dependence on user historical data and homogeneous recommendations.

[0006] The technical solution to achieve the purpose of the present invention is: a personalized news information recommendation method based on multi-channel recall, comprising the following steps:

[0007] Step 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with an interest tag selection function. If the user selects an interest tag, recommend the tag column information according to the user's interest tag; if the user refuses to select an interest tag, recommend cold start hot information; if the user is not a new user, determine whether there is user historical behavior data information. If there is no historical data, determine the interest tag according to the new user and recommend cold start hot information or columns; if there is historical behavior data, recommend comprehensive information by combining similar user preference information and information recommendation based on time series content;

[0008] Step 2: Comprehensively collect the cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation;

[0009] Step 3: Record the user's behavior log for recommended news, and update the recommendation data based on the user's latest behavior record when the user refreshes the recommendation.

[0010] Furthermore, the cold start hot information recommendation calculates the news popularity, sorts the news according to the news popularity, and preferentially recommends the top N pieces of data with the highest popularity, wherein the news popularity is jointly determined by the initial popularity and the interactive popularity, and the news popularity decays exponentially over time, as follows:

[0011] Initial heat h ini Determined by the data source and content attributes, the calculation formula is as follows:

[0012] h ini =w c *x c +w s *x s +w l *x l +x r × r +b ini

[0013] Among them, x c represents the basic score of news category, wc is the initial heat weight value of the news category, x s represents the news source base score, w s is the initial heat weight value of the news source, x l represents the news text length fraction, w l is the initial heat weight value of the news length, x r represents the news content richness score, w r is the initial heat weight value of news content richness, b ini is the initial heat deviation;

[0014] Interaction heat act It is determined by the user's behavior data and the calculation formula is as follows:

[0015] h act =w i *x i +w k *x k +w m *x m +w h *x h +w e × e +b act

[0016] Among them, x i represents the number of news clicks, w i is the weight of news clicks, x k Indicates the number of news likes, w k is the weight of the news likes, x m represents the amount of news comments, w m is the weight of news comments, x h represents the news sharing amount, w h is the weight of news sharing, x e represents the amount of news collection, w e is the weight of news collection, b act is the interaction heat deviation;

[0017] Heat time decay factor h ～ This means that the news popularity is exponentially decreasing over time, and the calculation formula is as follows:

[0018]

[0019] Among them, κ is the adjustment parameter, T p is the current time, T0 is the initial time of news release;

[0020] The news heat is determined by the above three factors, and the calculation formula is as follows:

[0021]

[0022] Sort by popularity and recommend the top N most popular news in news set S. heat_rec For the current user, the formula is represented as follows:

[0023] News heat_rec =Rank(Set(S:Heat)) N

[0024] S is the news set, Heat is the news heat, Rank(Set(S:Heat)) N Indicates sorting by news popularity and taking the top N news as News heat_rec A collection of highly recommended news.

[0025] Furthermore, the tag column recommendation, in the news collection and storage stage, classifies news categories and labels them with keywords according to news titles, keywords and content to form a news category and tag set, matches user interest tags with news categories and tags with keywords, and recommends news information with consistent categories and tags, and the recommended news is recorded as News. label_rec .

[0026] Furthermore, the similar user preference information recommendation is based on user interest tags, personal information, and historical behavior data, and integrates user-based collaborative filtering and content-based collaborative filtering algorithms to model the user's interests, calculate user similarities, and recommend news information preferred by other users with similar interests to the user, as follows:

[0027] The user to be recommended information is u. User-based collaborative filtering constructs the user interest vector based on the onehot interest vector construction method according to the news categories, data sources and user interest tags that the user is concerned about.

[0028]

[0029] Content-based collaborative filtering uses the text vectors mapped to all news titles browsed by the user as the user interest vector. Suppose all news titles browsed by the user are ST = {ST1, ST2, …, ST n}, for News headlines Perform word segmentation, stop word filtering, and word form restoration operations to obtain The word vector for words appearing in the word2vec model is recorded as For unregistered words that do not appear in word2vec, zero vectors are used for filling. The mapped text vector representation is as follows:

[0030]

[0031] User interest vector Characterization is as follows:

[0032]

[0033] User similarity calculation: Calculate two interest vectors of user u Two interest vectors with other users u' Similarity;

[0034]

[0035] Take the top ζ user sets U1 and U2 in similarity ranking respectively, merge the set U = U1 ∪ U2, get the top k similar users, and extract all the news S browsed by the top k similar users k , filter the news S that the current user u has browsed u , the remaining news is sorted by news popularity, and the top N news are recommended. sim_rec For the current user, the formula is represented as follows:

[0036] News si m _rec =Rank(Set(S k -S u :Heat)) N .

[0037] Furthermore, the information recommendation based on time series content combines the user's historical behavior data and the historical behavior time series information, models the user's interests based on the normalized word frequency model and the word vector model, and performs vector representation learning on the news data, calculates the similarity between the user's interests and the news, and recommends the news that the user is interested in, as follows:

[0038] (41) User interest modeling is performed based on the normalized word frequency model, and vector representation learning is performed on news data to calculate the similarity between user interest and news, and recommend news that users are interested in. The details are as follows:

[0039] According to the recorded user behavior logs, the historical behavior data of a single user is constructed. Since different user behaviors have different importance, the behavior weight is set according to the importance of user behavior. Secondly, user interests change dynamically, and the temporal changes of user behaviors represent the evolution of user interests. The closer the behavior time is, the more it can reflect the user's recent interests. Therefore, the temporal weight is set according to the user behavior time.

[0040] User behaviors and their weight settings are shown in Table 1;

[0041] Table 1 User historical behavior and weight

[0042] User behavior identification User Conduct Behavior Weight Timing Weight read Read and browse <![CDATA[W read ]]> <![CDATA[Q read ]]> favorite collect <![CDATA[W favorite ]]> <![CDATA[Q favorite ]]> like like <![CDATA[W like ]]> <![CDATA[Q like ]]> repost share <![CDATA[W repost ]]> <![CDATA[Q repost ]]> click_source View source link <![CDATA[W click_source ]]> <![CDATA[Q click_source ]]> search search <![CDATA[W search ]]> <![CDATA[Q search ]]> unlike dislike <![CDATA[W unlike ]]> <![CDATA[Q unlike ]]> remove delete <![CDATA[W remove ]]> <![CDATA[Q remove ]]>

[0043] Based on user behavior data, behavior time series information and user behavior weight, user interest is modeled. Let user i-th behavior H i , the behavior weight is W i , the time when the behavior occurs T i , the temporal weight Q of the behavior i , T p is the current time, where:

[0044] Positive behaviors of users are positive, including reading, browsing, collecting, liking, sharing, viewing source links, and searching. The larger the weight, the higher the value of the news to the user.

[0045] Negative behaviors of users are negative, including dislike and deletion. The larger the absolute value of the weight, the higher the user's aversion to the news;

[0046] The temporal weight of behavior is related to time. The weight of short-term behavior is larger, and the weight of long-term behavior is smaller. The temporal weight of behavior decays exponentially over time:

[0047]

[0048] ω is the behavior temporal weight adjustment parameter;

[0049] Assume that user i-th behavior H i The news set for the following operation is S i ={S i1 ,S i2 ,…,S in}, set S i The number of news is |S i |, S ik Represents the set S i The kth news in S i Each news in the D is segmented, stop word filtered, and word form restored to obtain D i ={d i1 ,di2 ,…,d im}, in S i In the statistics of words d ij The frequency of occurrence is recorded as:

[0050]

[0051] For S i The frequency of all words in the word multiplied by the weight W of the user's i-th behavior i and the behavior timing weight Q i , get word d ij The weight frequency of:

[0052]

[0053] There are two cases for fusing word weight frequencies of different behaviors:

[0054] (a) Words appear repeatedly in different behavior data. The weighted frequencies of the same words in different behavior data are added and combined with the repetition penalty factor Multiply, Related to the number of occurrences, we get the final weighted frequency of the word:

[0055]

[0056] (b) A word only appears in a certain type of behavior data. The fused word weight frequency is the unique weight frequency of the word δ'(d) = δ(d ij );

[0057] The fused word weight frequency set is used as the user's interest feature set UF:

[0058] UF=Set(d:δ′(d))

[0059] d represents a specific word, δ'(d) is the sum of the weighted frequencies of word d in different behavior data, which are aggregated to generate the user's characteristic behavior vocabulary and weights, and the characteristic keyword set constructed by the user's historical behavior data is saved. The characteristic keyword set constructed by all positive behavior data is recorded as D + , the set of characteristic keywords constructed by all negative behavior data is D - ;

[0060] In order to avoid the influence of excessive weight values ​​of some words on subsequent similarity calculations, the user weight frequency is normalized. The normalization formula is:

[0061]

[0062] The user's positive interest vector is:

[0063]

[0064] Among them, d i ∈D + ;

[0065] The user's negative interest vector is:

[0066]

[0067] Among them, d j ∈D - ;

[0068] Based on the positive behavior data feature keyword set D + And the negative behavior data feature keyword set D - , count the keywords and word frequencies in the news data to be recommended, and normalize the word frequencies to convert them into news positive word frequency vectors DV + and news negative word frequency vector DV - ;

[0069] Calculate the user's positive interest vector UFV respectively + and the news positive word frequency vector DV + Similarity, user negative interest vector UFV - and the news negative word frequency vector DV - Similarity;

[0070]

[0071] When sim(UFV + ,DV + )>sim(UFV - ,DV - ), it indicates that the news is close to the user's interests, and the user is more likely to be interested and click. + ,DV + )≤sim(UFV - ,DV - ), it indicates that the user is not interested in the content. Therefore, sim(UFV + ,DV + )>sim(UFV - ,DV - )'s news S is added to the recommendation candidate list;

[0072] Based on user interests and news popularity, candidate news are ranked and the top N news are recommended. com_rec For the current user, the formula is represented as follows:

[0073] News con_rec=Rank(Set(S:sim(UFV + ,DV + )+αHeat)) N

[0074] Where α is the heat adjustment parameter;

[0075] (42) Modeling user interests based on word vector models, learning vector representations of news data, calculating the similarity between user interests and news, and recommending news that users are interested in;

[0076] The user's interest vector UWV is expressed as:

[0077]

[0078] in Indicates news i Chinese word d ij The word vector, W i is the behavior weight, Q i is the behavior time series weight, user behavior H i The behavior weight is W i , the corresponding behavior operation data set is

[0079] S i ={S i1 ,S i2 ,…,S in}

[0080] The news text to be recommended is Its vector It is expressed as:

[0081]

[0082] in Indicates recommended news text Chinese words The word vector of The news text to be recommended;

[0083] Calculate user interest vector UWV and news vector Similarity;

[0084]

[0085] Select top-N recommendation, combining similarity and current news popularity to give users the top N news. vec_rec , the formula is as follows:

[0086]

[0087] Heat is the heat of the news to be recommended. Represents the similarity between the news to be recommended and the user’s interests, where α is the heat adjustment parameter.

[0088] Further, step 2: comprehensive cold start hot information recommendation, column recommendation, similar user preference information recommendation, and recall information of content-based information recommendation are deduplicated and integrated, and then comprehensively sorted according to popularity to complete the overall recommendation, as follows:

[0089] Integrate cold start hot information recommendation, column recommendation, similar user preference information recommendation, content-based information recommendation and other multi-channel recall information to generate the final candidate recommended news set News all ;

[0090] News all =News heat_rec ∪News label_rec ∪News sim_rec ∪News con_rec ∪News vec_rec

[0091] Recommend news set News for multiple recall candidates all Perform deduplication fusion, perform comprehensive sorting according to heat, select TOPN according to the sorting, and perform comprehensive personalized news recommendation. The recommended list is News rec ;

[0092] News rec =Rank(Set(News all :Heat)) N .

[0093] Further, step 3: record the user's behavior log for recommended news, and update the recommended data according to the user's latest behavior record when the user refreshes the recommendation, as follows:

[0094] Record all user behavior logs for recommended news and save them to the database. When the user refreshes the recommendation list, the recommendation data is updated based on the user's latest behavior records.

[0095] A personalized news information recommendation system based on multi-channel recall, implementing the personalized news information recommendation method based on multi-channel recall, realizing personalized news information recommendation based on multi-channel recall, comprising:

[0096] Module 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with a function to select interest tags;

[0097] Module 2: Comprehensive cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation;

[0098] Module 3: Record the user's behavior log for recommended news, and update the recommendation data based on the user's latest behavior record when the user refreshes the recommendation.

[0099] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the personalized news information recommendation method based on multi-channel recall is implemented to achieve personalized news information recommendation based on multi-channel recall.

[0100] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the personalized news information recommendation method based on multi-channel recall to achieve personalized news information recommendation based on multi-channel recall.

[0101] Compared with the prior art, the present invention has the following significant advantages: (1) The present invention adopts a cold-start hot information recommendation method, which makes up for the limitation of being unable to model the interests of cold-start users when relying solely on the neural network information recommendation model, reduces the dependence on user historical behavior data, and can build a more complete information recommendation method; (2) The present invention integrates the time-based content recommendation and similarity-based user preference information recommendation methods. The former models user interests based on user historical behavior data and behavior time-series information, and recommends information of known user interest; the latter recommends preference information of similar users based on user personal information and user behavior characteristics, avoids homogenization of information recommendation, and explores other potential interests of users; (3) The present invention adopts a multi-channel recall recommendation method, and the recall strategies designed from different angles ensure the comprehensiveness and novelty of information recommendation, and improve the recommendation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 This is the overall flow chart of personalized news information recommendation. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0104] like Figure 1As shown in the figure, a personalized news information recommendation method based on multi-way recall is proposed, which divides the overall news recommendation into four parts: hot information recommendation without user historical behavior data, column information recommendation based on user interest tags, and information recommendation based on time series content and similar user preference information recommendation with user historical behavior data.

[0105] Step 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with an interest tag selection function. If the user selects an interest tag, recommend the tag column information according to the user's interest tag; if the user refuses to select an interest tag, perform cold start hot information recommendation. If the user is not a new user, determine whether there is user historical behavior data information. If there is no historical data, determine the interest tag according to the new user and perform cold start hot information recommendation or column recommendation; if there is historical behavior data, recommend comprehensive information based on similar user preference information and information recommendation based on time series content.

[0106] (1) Cold start hotspot information recommendation

[0107] Cold start hot information recommendation, calculate the news popularity, sort according to the news popularity, and give priority to recommending the top N most popular data. News popularity is determined by the initial popularity and interactive popularity, and the popularity decays exponentially over time. The initial popularity is determined by the data source and content attributes, and the interactive popularity is determined by the user's behavior data and is also the most important popularity. The popularity time decay factor indicates that the popularity decays exponentially over time. The details are as follows:

[0108] Initial heat h ini Determined by the data source and content attributes. The calculation formula is as follows:

[0109] h ini =w c *x c +w s *x s +w l *x l +w r × r +b ini

[0110] Among them, x c represents the basic score of news category, w c is the initial heat weight value of the news category, x s represents the news source base score, w s is the initial heat weight value of the news source, x l represents the news text length fraction, w l is the initial heat weight value of the news length, xr represents the news content richness score (whether it contains images and videos), w r is the initial heat weight value of news content richness, b ini is the initial heat deviation. x c 、x s 、x l 、x r Can be defined according to the application area.

[0111] Interaction heat act Determined by the user's behavior data. The calculation formula is as follows:

[0112] h act =w i *x i +w k *x k +w m *x m +w h *x h +w e × e +b act

[0113] Among them, x i represents the number of news clicks, w i is the weight of news clicks, x k Indicates the number of news likes, w k is the weight of the news likes, x m represents the amount of news comments, w m is the weight of news comments, x h represents the news sharing amount, w h is the weight of news sharing, x e represents the amount of news collection, w e is the weight of news collection, b act Interaction heat deviation.

[0114] Heat time decay factor h ～ This means that the heat decays exponentially over time. The calculation formula is as follows:

[0115]

[0116] Among them, κ is the adjustment parameter, T p is the current time, and T0 is the initial time of news release.

[0117] The news heat is determined by the above three factors, and the calculation formula is as follows:

[0118]

[0119] Sort by popularity and recommend the top N most popular news in news set S. heat_rec For the current user, the formula is represented as follows:

[0120] News heat_rec =Rank(Set(S:Heat)) N

[0121] S is the news set, Heat is the news heat, Rank(Set(S:Heat)) N Indicates sorting by news popularity and taking the top N news as News heat_rec A collection of highly recommended news.

[0122] (2) Tag column recommendation

[0123] Tag column recommendation, match user interest tags with news categories by keywords, and recommend news information with the same category and tags. The details are as follows:

[0124] Tag column recommendation: In the news collection and storage phase, news categories are classified and keyword tags are annotated according to news titles, keywords, and content to form a news category and tag collection. User interest tags are matched with news categories and tags by keywords, and news information with consistent categories and tags is recommended. The recommended news is recorded as News. label_rec The specific classification algorithm adopts the existing algorithm, which is not within the scope of improvement covered by the present invention and will not be described in detail.

[0125] (3) Recommendation of similar user preference information

[0126] Similar user preference information recommendation, based on user interest tags, personal information, and historical behavior data, model the user's interests, calculate user similarity, and recommend news information preferred by other users with similar interests to the user. The details are as follows:

[0127] Similarity user preference information recommendation is based on user interest tags, personal information, and historical behavior data. It combines user-based collaborative filtering and content-based collaborative filtering algorithms to model user interests, calculate user similarities, and recommend news information preferred by other users with similar interests to users.

[0128] The user to be recommended information is u. User-based collaborative filtering constructs a user interest vector based on the onehot interest vector construction method according to the news categories, data sources, user interest tags, etc. that the user is concerned about.

[0129]

[0130] Content-based collaborative filtering uses the text vectors mapped to all news headlines browsed by the user as the user interest vector.

[0131] The titles of all news browsed by the user are ST = {ST1, ST2, ..., ST n}, for News headlines Perform word segmentation, stop word filtering, and word form restoration operations to obtain The word vector for words appearing in the word2vec model is recorded as For unregistered words that do not appear in word2vec, zero vectors are used for filling. The mapped text vector representation is as follows:

[0132]

[0133] User interest vector Characterization is as follows:

[0134]

[0135] User similarity calculation: Calculate two interest vectors of user u Two interest vectors with other users u' Similarity.

[0136]

[0137] Take the top ζ user sets U1 and U2 in similarity ranking respectively, merge the set U = U1 ∪ U2, and get the top k similar users. Extract all the news S browsed by the top k similar users k , filter the news S that the current user u has browsed u , the remaining news is sorted by popularity, and the top N news are recommended. sim_rec For the current user, the formula is represented as follows:

[0138] News sim_rec =Rank(Set(S k -S u :Heat)) N

[0139] (4) Information recommendation based on temporal content

[0140] Information recommendation based on time series content combines user historical behavior data and historical behavior time series information to model user interests, and learn vector representations of news data based on normalized word frequency models and word vector models, calculate the similarity between user interests and news, and recommend news that users are interested in. The details are as follows:

[0141] (41) Information recommendation based on time series content includes two recall methods. The first is to model user interests based on the normalized word frequency model, learn vector representations of news data, calculate the similarity between user interests and news, and recommend news that users are interested in.

[0142] According to the user behavior logs recorded in the system, the historical behavior data of a single user is constructed. Since different user behaviors have different importance, the behavior weight is set according to the importance of user behavior. Secondly, user interests change dynamically, and the temporal changes in user behavior represent the evolution of user interests. The closer the behavior time is, the more it can reflect the user's recent interests. Therefore, the temporal weight is set according to the user behavior time. The user behavior and its weight setting are shown in Table 1

[0143] Table 1 User historical behavior and weight

[0144] User behavior identification User Conduct Behavior Weight Timing Weight read Read and browse <![CDATA[W read ]]> <![CDATA[Q read ]]> favorite collect <![CDATA[W favorite ]]> <![CDATA[Q favorite ]]> like like <![CDATA[W like ]]> <![CDATA[Q like <!-- 9 -->]]> repost share <![CDATA[W repost ]]> <![CDATA[Q repost ]]> click_source View source link <![CDATA[W celick_source ]]> <![CDATA[Q click_source ]]> search search <![CDATA[W search ]]> <![CDATA[Q search ]]> unlike dislike <![CDATA[W unlike ]]> <![CDATA[Q unlike ]]> remove delete <![CDATA[W remove ]]> <![CDATA[Q remove ]]>

[0145] Consider the following points when setting weights:

[0146] (a) Positive behaviors of users are positive, including reading, browsing, collecting, liking, sharing, viewing source links, and searching. The larger the weight, the higher the value of the news to the user;

[0147] (b) Negative behaviors of users are negative, including dislike and deletion behaviors. The larger the absolute value of the weight, the higher the user's aversion to the news;

[0148] (c) The temporal weight of user behavior is related to time. The weight of short-term behavior is larger, while the weight of long-term behavior is smaller.

[0149] (d) The weight setting is placed in the configuration file and can be modified according to actual conditions;

[0150] User interests are modeled based on user behavior data, behavior timing information, and user behavior weights.

[0151] The i-th behavior of the user is H i , the behavior weight is W i , the time when the behavior occurs T i , the temporal weight Q of the behavior i , T p is the current time.

[0152] W i >0:H i Positive user behaviors, such as reading, liking, and collecting. The greater the weight, the higher the value of the news to the user.

[0153] W i <0:Hi For negative user behaviors, such as dislike and deletion, the greater the absolute value of the weight, the higher the user's aversion to the news;

[0154] The temporal weight of an action decays exponentially over time:

[0155]

[0156] ω is the behavior timing weight adjustment parameter.

[0157] User i-th behavior H i The news set for the following operation is S i ={S i1 ,S i2 ,…,S in}, set S i The number of news is |S i |, S ik Represents the set S i The kth news item.

[0158] For S i D is obtained by performing operations such as word segmentation, stop word filtering, and word form restoration on each news in i ={d i1 ,d i2 ,…,d im}, in the set S i In the statistics ij The frequency of occurrence of a word is recorded as:

[0159]

[0160] For S i The frequency of all words in the word multiplied by the weight W of the user's i-th behavior i and the behavior timing weight Q i , get word d ij The weight frequency of:

[0161]

[0162] There are two cases for fusing word weight frequencies of different behaviors:

[0163] (a) Words appear repeatedly in different behavior data. The weighted frequencies of the same words in different behavior data are added and combined with the repetition penalty factor Multiply, Related to the number of occurrences, we get the final weighted frequency of the word:

[0164]

[0165] (b) A word only appears in a certain type of behavior data. The fused word weight frequency is the unique weight frequency of the word δ'(d) = δ(d ij ).

[0166] The fused word weight frequency set is used as the user's interest feature set UF:

[0167] UF=Set(d:δ′(di))

[0168] d represents a specific word, and δ'(d) is the sum of the weighted frequencies of word d in different behavior data. Collect them together to generate the user's characteristic behavior vocabulary and weights, and save the characteristic keyword set constructed by the user's historical behavior data. The characteristic keyword set constructed by all positive behavior data is recorded as D + , the set of characteristic keywords constructed by all negative behavior data is D - d∈{D + ∪D -}

[0169] In order to avoid the influence of excessive weight values ​​of some words on subsequent similarity calculations, the user weight frequency is normalized. The normalization formula is:

[0170]

[0171] The user's positive interest vector is:

[0172]

[0173] Among them, d i ∈D + .

[0174] The user's negative interest vector is:

[0175]

[0176] Among them, D j ∈D - .

[0177] News vectorization, based on the positive behavior data feature keyword set D + And the negative behavior data feature keyword set D - According to the above steps, the keywords and word frequencies appearing in the news data to be recommended are counted, and the word frequencies are normalized and converted into positive word frequency vectors DV + And the negative word frequency vector DV - .

[0178] Similarity calculation, calculate the user's positive interest vector UFV + and the news positive word frequency vector DV +Similarity, user negative interest vector UFV - and the news negative word frequency vector DV - Similarity.

[0179]

[0180] When sim(UFV + ,DV + )>sim(UFV - ,DV - ), it indicates that the news is close to the user's interests, and the user is more likely to be interested and click. + ,DV + )≤sim(UFV - ,DV - ), it indicates that the user is not interested in the content. Therefore, sim(UFV + ,DV + )>sim(UFV - ,DV - )'s news S is added to the recommendation candidate list.

[0181] Based on user interests and news popularity, candidate news are ranked and the top N news are recommended. con_rec For the current user, the formula is represented as follows:

[0182] News con_rec =Rank(Set(S:sim(UFV + ,DV + )+αHeat)) N

[0183] Here, α is the heat adjustment parameter, which can adjust the impact of heat on content recommendation according to actual conditions.

[0184] (42) The second recall method of information recommendation based on temporal content is to model user interests based on the word vector model, learn vector representations of news data, calculate the similarity between user interests and news, and recommend news that users are interested in.

[0185] At this time, the user's interest vector UWV can be expressed as:

[0186]

[0187] in Indicates news i Chinese word d ij The word vector, W i is the behavior weight, Q i is the behavior time series weight, user behavior H iThe behavior weight is W i , the corresponding behavior operation data set is

[0188] S i ={S i1 ,S i2 ,…,S in}

[0189] The news text to be recommended is Its vector It is expressed as:

[0190]

[0191] in Indicates recommended news text Chinese words The word vector of The news text to be recommended;

[0192] Calculate user interest vector UWV and news vector Similarity;

[0193]

[0194] Select top-N recommendation, combining similarity and current news popularity to give users the top N news. vec_rec , the formula is as follows:

[0195]

[0196] Heat is the heat of the news to be recommended. Represents the similarity between the news to be recommended and the user’s interests, where α is the heat adjustment parameter.

[0197] Step 2: Comprehensively collect cold start hot information recommendations, column recommendations, similar user preference information recommendations, content-based information recommendations and other multi-channel recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation. The details are as follows:

[0198] Integrate cold start hot information recommendation, column recommendation, similar user preference information recommendation, content-based information recommendation and other multi-channel recall information to generate the final candidate recommended news set News all .

[0199] News all =News heat_rec ∪News label_rec ∪News sim_rec ∪News con_rec ∪News vec_rec

[0200] Recommend news set News for multiple recall candidates all Perform deduplication fusion, perform comprehensive sorting according to heat, select TOPN according to the sorting, and perform comprehensive personalized news recommendation. The recommended list is News rec .

[0201] News rec =Rank(Set(News all :Heat)) N

[0202] Step 3: Record the user's behavior log for recommended news, and update the recommended data according to the user's latest behavior record when the user refreshes the recommendation. The details are as follows:

[0203] Record all user behavior logs for recommended news and save them to the database. When the user refreshes the recommendation list, update the recommendation data according to the user's latest behavior records according to the above steps 1-2.

[0204] The present invention also proposes a personalized news information recommendation system based on multi-channel recall, and implements the personalized news information recommendation method based on multi-channel recall to achieve personalized news information recommendation based on multi-channel recall, including:

[0205] Module 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with a function to select interest tags;

[0206] Module 2: Comprehensive cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation;

[0207] Module 3: Record the user's behavior log for recommended news, and update the recommendation data based on the user's latest behavior record when the user refreshes the recommendation.

[0208] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the personalized news information recommendation method based on multi-channel recall is implemented to achieve personalized news information recommendation based on multi-channel recall.

[0209] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the personalized news information recommendation method based on multi-channel recall to achieve personalized news information recommendation based on multi-channel recall.

[0210] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0211] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A personalized news information recommendation method based on multi-channel recall, characterized in that: The steps include: Step 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with an interest tag selection function. If the user selects an interest tag, recommend tag column information according to the user's interest tag; If the user refuses to select the interest tag, cold start hot information recommendation is performed; if the user is not a new user, determine whether there is user historical behavior data information. If there is no historical data, the interest tag is determined according to the new user and cold start hot information recommendation or column recommendation is performed; if there is historical behavior data, comprehensive information recommendation is performed based on similar user preference information recommendation and information recommendation based on time series content; Step 2: Comprehensively collect the cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation; Step 3: Record the user's behavior log for recommended news, and update the recommendation data based on the user's latest behavior record when the user refreshes the recommendation.

2. The personalized news information recommendation method based on multi-channel recall according to claim 1 is characterized in that: The cold start hot information recommendation calculates the news popularity, sorts the news according to the news popularity, and recommends the top N pieces of data with the highest popularity first. The news popularity is determined by the initial popularity and the interactive popularity. The news popularity decays exponentially over time. The details are as follows: Initial heat h ini Determined by the data source and content attributes, the calculation formula is as follows: h ini =w c *x c +w s *x s +w l *x l +w r ×x r +b ini Among them, x c represents the basic score of news category, w c is the initial heat weight value of the news category, x s represents the news source base score, w s is the initial heat weight value of the news source, x l represents the news text length fraction, w l is the initial heat weight value of the news length, x r represents the news content richness score, w r is the initial heat weight value of news content richness, b ini is the initial heat deviation; Interaction heat act It is determined by the user's behavior data and the calculation formula is as follows: h act =w i *x i +w k *x k +w m *x m +w h *x h +w e ×x e +b act Among them, x i represents the number of news clicks, w i is the weight of news clicks, x k Indicates the number of news likes, w k is the weight of the news likes, x m represents the amount of news comments, w m is the weight of news comments, x h represents the news sharing amount, w h is the weight of news sharing, x e represents the amount of news collection, w e is the weight of news collection, b act is the interaction heat deviation; Heat time decay factor h ～ This means that the news popularity is exponentially decreasing over time, and the calculation formula is as follows: Among them, κ is the adjustment parameter, T p is the current time, T0 is the initial time of news release; The news heat is determined by the above three factors, and the calculation formula is as follows: Sort by popularity and recommend the top N most popular news in news set S. heat_rec For the current user, the formula is represented as follows: News heat_rec =Rank(Set(S:Heat)) N S is the news set, Heat is the news heat, Rank(Set(S:Heat)) N Indicates sorting by news popularity and taking the top N news as News heat_rec A collection of highly recommended news.

3. The personalized news information recommendation method based on multi-channel recall according to claim 1 is characterized in that: The tag column recommendation is to classify news categories and annotate them with keyword tags according to news titles, keywords and content during the news collection and storage phase to form a news category and tag set. The user's interest tags are matched with news categories and tags by keyword, and news information with consistent categories and tags is recommended. The recommended news is recorded as News. label_rec .

4. The personalized news information recommendation method based on multi-channel recall according to claim 2 is characterized in that: The similarity user preference information recommendation is based on user interest tags, personal information, and historical behavior data, and integrates user-based collaborative filtering and content-based collaborative filtering algorithms to model the user's interests, calculate user similarities, and recommend news information preferred by other users with similar interests to the user, as follows: The user to be recommended information is u. User-based collaborative filtering constructs the user interest vector based on the onehot interest vector construction method according to the news categories, data sources and user interest tags that the user is concerned about. Content-based collaborative filtering uses the text vectors mapped to all news titles browsed by the user as the user interest vector. Suppose all news titles browsed by the user are ST = {ST1, ST2, ..., ST n }, for News headlines Perform word segmentation, stop word filtering, and word form restoration operations to obtain The word vector for words appearing in the word2vec model is recorded as For unregistered words that do not appear in word2vec, zero vectors are used for filling. The mapped text vector representation is as follows: User interest vector Characterization is as follows: User similarity calculation: Calculate two interest vectors of user u and two interest vectors of other users u′ Similarity; Take the top ζ user sets U1 and U2 in similarity ranking respectively, merge the set U = U1 ∪ U2, get the top k similar users, and extract all the news S browsed by the top k similar users k , filter the news S that the current user u has browsed u , the remaining news is sorted by news popularity, and the top N news are recommended. sim_rec For the current user, the formula is represented as follows: News sim_rec =Rank(Set(S k -S u :Heat)) N 。 5. The personalized news information recommendation method based on multi-channel recall according to claim 1 is characterized in that: The information recommendation based on time series content combines the user's historical behavior data and historical behavior time series information, models the user's interests based on the normalized word frequency model and the word vector model, and performs vector representation learning on the news data, calculates the similarity between the user's interests and the news, and recommends the news that the user is interested in, as follows: (41) User interest modeling is performed based on the normalized word frequency model, and vector representation learning is performed on news data to calculate the similarity between user interests and news, and recommend news that users are interested in. The details are as follows: According to the recorded user behavior logs, the historical behavior data of a single user is constructed. Since different user behaviors have different importance, the behavior weight is set according to the importance of user behavior. Secondly, user interests change dynamically, and the temporal changes of user behaviors represent the evolution of user interests. The closer the behavior time is, the more it can reflect the user's recent interests. Therefore, the temporal weight is set according to the user behavior time. User behaviors and their weight settings are shown in Table 1; Table 1 User historical behavior and weight Based on user behavior data, behavior time series information and user behavior weight, user interest is modeled. Let user i-th behavior H i , the behavior weight is W i , the time when the behavior occurs T i , the temporal weight Q of the behavior i , T p is the current time, where: Positive behaviors of users are positive, including reading, browsing, collecting, liking, sharing, viewing source links, and searching. The larger the weight, the higher the value of the news to the user. Negative behaviors of users are negative, including dislike and deletion. The larger the absolute value of the weight, the higher the user's aversion to the news; The temporal weight of behavior is related to time. The weight of short-term behavior is larger, and the weight of long-term behavior is smaller. The temporal weight of behavior decays exponentially over time: ω is the behavior temporal weight adjustment parameter; Assume that user i-th behavior H i The news set for the following operation is S i ={S i1 ,S i2 ,…,S in }, set S i The number of news is |S i |,S ik Represents the set S i The kth news in S i Each news in the D is segmented, stop word filtered, and word form restored to obtain D i ={d i1 , d i2 , ..., d im }, in S i In the statistics of words d ij The frequency of occurrence is recorded as: For S i The frequency of all words in the word multiplied by the weight W of the user's i-th behavior i and the behavior timing weight Q i , get word d ij The weight frequency of: There are two cases for fusing word weight frequencies of different behaviors: (a) Words appear repeatedly in different behavior data. The weighted frequencies of the same words in different behavior data are added and combined with the repetition penalty factor Multiply, Related to the number of occurrences, we get the final weighted frequency of the word: (b) A word only appears in a certain type of behavior data. The fused word weight frequency is the unique weight frequency of the word δ′(d) = δ(d ij ); The fused word weight frequency set is used as the user's interest feature set UF: UF=Set(d:δ′(d)) d represents a specific word, δ′(d) is the sum of the weighted frequencies of word d in different behavior data, which are aggregated to generate the user's characteristic behavior vocabulary and weights, and the characteristic keyword set constructed by the user's historical behavior data is saved. The characteristic keyword set constructed by all positive behavior data is recorded as D + , the set of characteristic keywords constructed by all negative behavior data is D - ; In order to avoid the influence of excessive weight values ​​of some words on subsequent similarity calculations, the user weight frequency is normalized. The normalization formula is: The user's positive interest vector is: Among them, d i ∈D + ; The user's negative interest vector is: Among them, d j ∈D - ; Based on the positive behavior data feature keyword set D + And the negative behavior data feature keyword set D - , count the keywords and word frequencies in the news data to be recommended, and normalize the word frequencies to convert them into news positive word frequency vectors DV + and news negative word frequency vector DV - ; Calculate the user's positive interest vector UFV respectively + and the news positive word frequency vector DV + Similarity, user negative interest vector UFV - and the news negative word frequency vector DV - Similarity; When sim(UFV + , DV + )>sim(UFV - , DV - ), it indicates that the news is close to the user's interests, and the user is more likely to be interested and click. + , DV + )≤sim(UFV - , DV - ), it indicates that the user is not interested in the content. Therefore, sim(UFV + , DV + )>sim(UFV - , DV - )'s news S is added to the recommendation candidate list; Based on user interests and news popularity, candidate news are ranked and the top N news are recommended. con_rec For the current user, the formula is represented as follows: News con_rec =Rank(Set(S:yes(UFV + ,DV + )+αHeat)) N Where α is the heat adjustment parameter; (42) Modeling user interests based on word vector models, learning vector representations of news data, calculating the similarity between user interests and news, and recommending news that users are interested in; The user's interest vector UWV is expressed as: in Indicates news i Chinese word d ij The word vector, W i is the behavior weight, Q i is the behavior time series weight, user behavior H i The behavior weight is W i , the corresponding behavior operation data set is S i ={S i1 ,S i2 ,...,S in } The news text to be recommended is Its vector It is expressed as: in Indicates recommended news text Chinese words The word vector of The news text to be recommended; Calculate user interest vector UWV and news vector Similarity; Select top-N recommendation, combining similarity and current news popularity to give users the top N news. vec_rec , the formula is as follows: Heat is the heat of the news to be recommended. Represents the similarity between the news to be recommended and the user’s interests, where α is the heat adjustment parameter.

6. The personalized news information recommendation method based on multi-channel recall according to claim 1 is characterized in that: Step 2: Comprehensively deduce and integrate the cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, and then sort them comprehensively according to popularity to complete the overall recommendation. The details are as follows: Integrate cold start hot information recommendation, column recommendation, similar user preference information recommendation, content-based information recommendation and other multi-channel recall information to generate the final candidate recommended news set News all ; News all =NewS heat_rec ∪News label_rec ∪News sim_rec ∪News con_rec ∪News vec_rec Recommend news set News for multiple recall candidates all Perform deduplication fusion, perform comprehensive sorting according to heat, and select TOP N for comprehensive personalized news recommendation according to the sorting. The recommended list is News rec ; News rec =Rank(Set(News all :Heat)) N 。 7. The personalized news information recommendation method based on multi-channel recall according to claim 1 is characterized in that: Step 3: Record the user's behavior log for recommended news, and update the recommended data according to the user's latest behavior record when the user refreshes the recommendation, as follows: Record all user behavior logs for recommended news and save them to the database. When the user refreshes the recommendation list, the recommendation data is updated based on the user's latest behavior records.

8. A personalized news information recommendation system based on multi-channel recall, characterized in that: Implementing the personalized news information recommendation method based on multi-channel recall described in any one of claims 1 to 7 to realize personalized news information recommendation based on multi-channel recall, comprising: Module 1: For the user to whom information is to be recommended, determine whether the user is a new user. If the user is a new user, provide the user with a function to select interest tags; Module 2: Comprehensive cold start hot information recommendation, column recommendation, similar user preference information recommendation, and content-based information recommendation recall information, perform deduplication fusion, and perform comprehensive sorting according to popularity to complete the overall recommendation; Module 3: Record the user's behavior log for recommended news, and update the recommendation data based on the user's latest behavior record when the user refreshes the recommendation.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the personalized news information recommendation method based on multi-channel recall described in any one of claims 1 to 7 is implemented to realize personalized news information recommendation based on multi-channel recall.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the personalized news information recommendation method based on multi-channel recall described in any one of claims 1 to 7 is implemented to achieve personalized news information recommendation based on multi-channel recall.

Images