CN113420221B - Interpretable recommendation method integrating implicit article preference and explicit feature preference of user - Google Patents
Interpretable recommendation method integrating implicit article preference and explicit feature preference of user Download PDFInfo
- Publication number
- CN113420221B CN113420221B CN202110747613.1A CN202110747613A CN113420221B CN 113420221 B CN113420221 B CN 113420221B CN 202110747613 A CN202110747613 A CN 202110747613A CN 113420221 B CN113420221 B CN 113420221B
- Authority
- CN
- China
- Prior art keywords
- user
- article
- feature
- expression
- explicit
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Abstract
The invention discloses an interpretable recommendation method fusing implicit article preference and explicit feature preference of a user, which can greatly improve the accuracy and interpretability of rating prediction by mining the feature preference of the user and the feature information of an article capturing time sequence information through extracting feature words in comments and mining the implicit expressions of the user and the article through the historical behaviors of the user; then, by extracting feature words in the user comments, capturing time sequence information in the comments by using a GRU (group-based Unit), and obtaining explicit expressions of the user and the article by introducing an attention mechanism and a feature attraction mechanism; and finally, combining the implicit expression and the display expression to obtain the final expression of the user and the article. The method has the advantages of high rationality of the recommendation result and capability of providing recommendation explanation while ensuring the recommendation precision.
Description
Technical Field
The invention relates to the technical field of computer personalized recommendation methods, in particular to an interpretable recommendation method fusing implicit article preference and explicit feature preference of a user.
Background
With the rapid development of internet technology in recent years, the problem of information overload is becoming more serious. The classified catalogue and the search engine alleviate the problem of information overload to a certain extent, but both have certain limitations. Therefore, the recommendation system has come up, helps users to complete information screening by analyzing the historical behaviors of the users, and can find out the potential interest preference of the users. The recommendation system does not need the user to provide explicit needs, but actively finds information that can satisfy the user's interests and needs from a large amount of information for the user.
Nowadays, the recommendation system is applied to all aspects of our life, a plurality of new recommendation technologies are promoted, the accuracy of a recommendation algorithm is improved, and the recommendation effect is greatly improved. Conventional recommendation systems often lack an explanation of the recommendation result, and interpretable recommendation refers to giving an explanation of recommending a commodity while recommending the commodity to a user. By giving an explanation of the recommendation, the transparency of the system and the confidence level of the user and the use experience can be increased, and the user is helped to make a selection more quickly and accurately.
In addition to using the scoring information of users in recommendation systems, attempts have been made in recent years to add user comment data to improve the accuracy and interpretability of recommendation systems. The evaluation of the user on the article reflects the display preference of the user to a certain extent, so that the explanation of the recommendation by using the user comment data is one of the research hotspots of many existing recommendation works. Although many related scholars have conducted research in recent years that could explain the field of recommendations, there are still some problems with existing work. Existing approaches almost all model with all comments of a user or an item and deal with the comments of the user and the item in the same way. On one hand, the noise in the comment data is large, and the information amount transferred by a large part of text content is extremely small. On the other hand, the comments of the users and the articles are greatly different, the comments of the users are sent by one user and can represent the characteristic preference of the users, and the comments of the articles are sent by a plurality of users, so that the evaluation dimensions are different and the subjectivity is strong, and therefore the characteristic information of the articles cannot be objectively reflected by the comments of the articles.
Although the current recommendation system also has the concept of introducing attention weight based on item characteristics and attention weight based on user characteristic preference, the attention weight based on item characteristics only describes the intensity of a certain characteristic of the item and is irrelevant to the target user; the attention weight based on the user feature preference describes the preference strength of a target user for a certain feature, and an article to be recommended may not have the feature or the feature weight is small, so that it is not reasonable to make a recommendation explanation only by using the attention weight based on the user feature preference or the article feature. For example, the attention weights of the target user u to the feature 1, the feature 2 and the feature 3 are 0.8, 0.1 and 0.1 respectively, and the attention weights of the item to be recommended v to the feature 1, the feature 2 and the feature 3 are 0.1, 0.1 and 0.8 respectively. When recommending an item v to a target user u, if feature 1 is used as the recommendation interpretation, although the preference degree of the user u for feature 1 is the highest, the intensity of feature 1 of the item v is very small, obviously unreasonable; if feature 3 is used as the recommended interpretation, although feature 3 of item v is the most intense, user u has a very small preference for feature 3, and it is obviously not reasonable. Therefore, it is urgently needed to develop a more reasonable and more accurate interpretable recommendation method integrating implicit item preference and explicit feature preference of a user, so that a recommendation system is more perfect.
Disclosure of Invention
The invention aims to provide an interpretable recommendation method fusing implicit article preference and explicit feature preference of a user. The method has the advantages of high recommendation result reasonability and capability of providing recommendation explanation while ensuring recommendation precision.
The technical scheme of the invention is as follows: an interpretable recommendation method fusing implicit article preference and explicit feature preference of a user comprises the following steps:
step A, extracting the latest p articles from the historical records of the user to obtain the implicit expression of the user;
b, extracting the latest q users from the preference records of the articles to obtain the implicit expression of the articles;
c, extracting feature preferences of the user from the comments of the user, calculating the attention weight of the user to each feature, and obtaining the explicit expression of the user;
d, extracting feature information of the article from the comment of the article, calculating the attention weight of each feature to the article, calculating the initial attraction of the feature information of the article to the user according to the explicit expression result of the user in the step C, calculating the final attraction of each feature of the article to the user according to the attention weight of each feature to the article and the initial attraction of the feature information of the article to the user, and obtaining the explicit expression of the article;
Step E, obtaining the final expression of the user according to the implicit expression of the user in the step A and the explicit expression of the user in the step C; obtaining the final expression of the article according to the implicit expression of the article in the step B and the explicit expression of the article in the step D;
and F, generating a recommendation list and giving a recommendation explanation.
Compared with the prior art, the invention has the beneficial effects that: the attractiveness of each feature of the object to the target user is calculated according to the feature preference of the target user, and the explicit feature expression of the object is determined according to the attractiveness of the feature, so that the recommendation precision and reasonability are guaranteed. Specifically, when the explicit expression of the user is obtained, the attention mechanism is introduced based on the feature words, and the attention weight of the user to each feature is calculated, so that the explicit expression of the user is obtained; when the explicit expression of the article is obtained, introducing an attraction mechanism, firstly calculating attention weight of each feature to the article, then calculating initial attraction of feature information of the article to a user according to an explicit expression result of the user, and then calculating final attraction of each feature of the article to the user according to the attention weight and the initial attraction so as to obtain the most accurate and reasonable explicit expression of the article; and finally, fusing the implicit preference and the explicit characteristic preference of the user to obtain the final expression of the user, fusing the implicit expression and the explicit characteristic information of the article to obtain the final expression of the article, and effectively guaranteeing the recommendation precision and rationality.
In the aforementioned interpretable recommendation method combining the implicit item preference and the explicit feature preference of the user, the obtaining of the user explicit expression in step C specifically includes the following steps:
step c1, extracting feature words: extracting characteristic words from the comments of the user, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the user A word vector representing the ith feature word of user u;
step c2, capturing time sequence information: feature word sequenceSending the data into a GRU network according to time to obtain a new feature word sequence The vector representation containing the time sequence information is obtained after the ith characteristic word representing the user u passes through the GRU;
step c3, attention-drawing mechanism: and measuring the preference degree of the user to the features by using the attention weight based on the features, wherein the calculation formula of the attention weight of the user u to the ith feature is as follows:
wherein, W u ,b u Learnable parameters for the layer network;
in the aforementioned interpretable recommendation method fusing the implicit item preference and the explicit feature preference of the user, the obtaining of the explicit expression of the item in the step D specifically includes the following steps:
step d1, extracting feature words: extracting characteristic words from the comments of the article, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the article A word vector representing the ith feature word of item v;
step d2, capturing timing information: feature word sequenceSending the data into a GRU network according to time to obtain a new feature word sequence The vector representation containing time sequence information is obtained after the ith characteristic word representing the article v passes through the GRU;
step d3, attention-drawing mechanism: the expression strength of the feature to the item is measured by using the feature-based attention weight, and the calculation formula of the attention weight of the ith feature to the item v is as follows:
wherein, W v ,b v Learnable parameters for the layer network;
step d4, introducing an attraction mechanism: calculating the preliminary attraction of each feature of the item to the user according to the user features, wherein the preliminary attraction of the ith feature of the item v to the user u is calculated according to the formula:
wherein, W u,v ,b u,v Learnable parameters for the layer network;
step d5, calculating the final attraction of the ith characteristic of the item v to the user u, wherein the calculation formula is as follows:
in the aforementioned interpretable recommendation method for fusing implicit item preference and explicit feature preference of user, in the step a, v is set to { v ═ v 1 ,v 2 ,…,v k V is the set of items preferred by the user, v i Representing the ith item, and setting the item vector sequence v as { v } 1 ,v 2 ,…,v p Inputting into a multilayer perceptron to obtain an implicit expression of a userTo U y 。
In the interpretable recommendation method combining the implicit item preference and the explicit feature preference of the user, in the step B, u is set to { u ═ 1 ,u 2 ,…,u q U is the set of users who prefer the item i Representing the ith user, and setting the user vector sequence u as { u } 1 ,u 2 ,…,u q Inputting into a multilayer perceptron to obtain an implicit expression V of an article y 。
In the interpretable recommendation method fusing the implicit item preference and the explicit feature preference of the user, a prediction of a score is further included between the step E and the step F, so that the final expression of the user is U ═ U x :U y ]The final expression of the article is V ═ V x :V y ]The prediction score of user i for item j is:
in the aforementioned interpretable recommendation method fusing the implicit item preference and the explicit feature preference of the user, the parameters of the predictive scoring model can be optimized as follows:theta is the set of model parameters and lambda is the regular term coefficient.
In the aforementioned interpretable recommendation method fusing implicit item preferences and explicit feature preferences of a user, the model evaluation uses MAE and RMSE.
Drawings
FIG. 1 is a schematic flow chart of the present invention.
Detailed Description
The invention is further illustrated by the following figures and examples, which are not to be construed as limiting the invention.
Example (b): an interpretable recommendation method fusing implicit article preference and explicit feature preference of a user is shown in a main flow model diagram of fig. 1, and comprises the following steps:
step A, extracting the latest p articles from the historical records of the user and orderingv={v 1 ,v 2 ,…,v k Is a collection of items, v i Representing the ith item, and setting the item vector sequence v as { v } 1 ,v 2 ,…,v p Inputting a multilayer sensing machine to obtain implicit article preference of a user, namely implicitly expressing U y , For general symbols, representing a vector U y Is d dimension.
Step B, extracting the latest q users from the preference records of the articles, and making u ═ { u ═ 1 ,u 2 ,…,u q Is the user set, u i Representing the ith user, and setting the user vector sequence u as { u } 1 ,u 2 ,…,u q Inputting into a multilayer perceptron to obtain audience information of the article, namely implicit expression V y ,
C, extracting feature preference of the user from the comments of the user, calculating attention weight of the user to each feature, and obtaining explicit expression of the user;
the step C of obtaining the user explicit expression specifically comprises the following steps:
step c1, extracting feature words: extracting characteristic words from the comments of the user, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the userWherein the content of the first and second substances,a word vector representing the ith feature word of user u,
Step c2, capturing timing sequenceInformation: feature word sequenceSending the data into a GRU network according to time to obtain a new feature word sequenceWherein, the first and the second end of the pipe are connected with each other,the ith characteristic word representing the user u passes through the GRU to obtain a vector representation containing time sequence information,
the calculation mode of the characteristic word sequence in the GRU is as follows:
wherein z is t To refresh the door, r t A reset gate, sigma (·) is a Logistic function, the output interval is (0, 1),is a candidate state for the current time instant,is an input for the current time of day,is the external state at the previous moment, W * ,U * ,b * E { z, r, f } is the learnable network parameter.
Step c3, attention-drawing mechanism: and measuring the preference degree of the user to the features by using the attention weight based on the features, wherein the calculation formula of the attention weight of the user u to the ith feature is as follows:
wherein, W u ,b u Learnable parameters for the layer network;
d, extracting feature information of the article from the comment of the article, calculating the attention weight of each feature to the article, calculating the initial attraction of the feature information of the article to the user according to the explicit expression result of the user in the step C, calculating the final attraction of each feature of the article to the user according to the attention weight of each feature to the article and the initial attraction of the feature information of the article to the user, and obtaining the explicit expression of the article;
The step D of obtaining the explicit expression of the article specifically comprises the following steps:
step d1, extracting feature words: extracting characteristic words from the comments of the article, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the articleWherein the content of the first and second substances,a word vector representing the ith feature word of item v,
step d2, capturing timing information: feature word sequenceSending the data into a GRU network according to time to obtain a new feature word sequenceWherein the content of the first and second substances,the ith characteristic word representing the item v is represented by a vector containing time sequence information obtained after passing through a GRU,
the calculation mode of the feature word sequence in the GRU is consistent with the calculation mode in the step c 2;
step d3, attention-drawing mechanism: the expression strength of the feature to the item is measured by using the feature-based attention weight, and the calculation formula of the attention weight of the ith feature to the item v is as follows:
wherein, W v ,b v Learnable parameters for the layer network;
step d4, introducing an attraction mechanism: calculating the preliminary attraction of each feature of the item to the user according to the user features, wherein the preliminary attraction of the ith feature of the item v to the user u is calculated according to the formula:
wherein, W u,v ,b u,v Learnable parameters for the layer network;
step d5, calculating The ith characteristic of the article v is finally attractive to the user u, and the calculation formula is as follows:
step E, obtaining the final expression U ═ U of the user according to the implicit expression of the user in the step A and the explicit expression of the user in the step C x :U y ],Obtaining a final expression V ═ V of the article according to the implicit expression of the article in the step B and the explicit expression of the article in the step D x :V y ],
the parameters of the model can be obtained by the following optimization problem: wherein r is ij And (4) the user i truly scores the item j, theta is a set of model parameters, and lambda is a regular term coefficient.
The evaluation of the recommendation model uses MAE and RMSE, which are calculated as follows:
wherein r is ij The user i is given a true score for item j.
Step F, scoring according to the predictionAnd recommending the Top-K item to the user. At the same time according toAndgenerating a recommended interpretation in the form of: you may be interested in [ featured word ], and the item is more relevant to [ featured word ].
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned examples, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to those skilled in the art without departing from the principles of the present invention should also be considered as within the scope of the present invention.
Claims (7)
1. An interpretable recommendation method fusing implicit article preference and explicit feature preference of a user is characterized in that: the method comprises the following steps: step A, extracting the latest data from the user's historyAn item, obtaining an implicit expression of a user;
step B, extracting the latest item from the preference records of the itemsThe user obtains the implicit expression of the article;
c, extracting feature preferences of the user from the comments of the user, calculating the attention weight of the user to each feature, and obtaining the explicit expression of the user;
d, extracting feature information of the article from the comment of the article, calculating the attention weight of each feature to the article, calculating the initial attraction of the feature information of the article to the user according to the explicit expression result of the user in the step C, calculating the final attraction of each feature of the article to the user according to the attention weight of each feature to the article and the initial attraction of the feature information of the article to the user, and obtaining the explicit expression of the article;
step E, obtaining the final expression of the user according to the implicit expression of the user in the step A and the explicit expression of the user in the step C; obtaining the final expression of the article according to the implicit expression of the article in the step B and the explicit expression of the article in the step D;
Step F, generating a recommendation list and giving a recommendation explanation;
the step D of obtaining the explicit expression of the article specifically comprises the following steps:
step d1, extracting feature words: extracting characteristic words from the comments of the article, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the article,Representing an articleTo (1)A word vector of the individual feature words;
step d2, capturing time sequence information: feature word sequenceSending the data into GRU network according to time to obtain new characteristicsToken sequence,Representing an articleTo (1) aVector representation containing time sequence information is obtained after the characteristic words pass through GRU;
step d3, attention-drawing mechanism: using feature-based attention weights to measure the expressive intensity of features on an item, the firstCharacteristic pair articleThe formula for calculating the attention weight of (1) is:
wherein, the first and the second end of the pipe are connected with each other,,learnable parameters for the layer network;
step d4, introducing an attraction mechanism: calculating a preliminary appeal of each feature of the item to the user based on the user features, the itemTo (1) aIndividual characteristics to the userPreliminary attractive force of
The calculation formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,,learnable parameters for the layer network;
step d5, calculating itemTo (1) aIndividual characteristics to the userAnd final attraction force is calculated according to the formula: ;
2. the method of claim 1, wherein the method comprises: the step C of obtaining the user explicit expression specifically comprises the following steps:
step c1, extracting feature words: extracting characteristic words from the comments of the user, converting the characteristic words into word vectors, and obtaining a characteristic word sequence of the user,Representing a userTo (1) aA word vector of the individual feature words;
step c2, capturing timing information: feature word sequenceSending the data into a GRU network according to time to obtain a new feature word sequence,Representing a userTo (1) aVector representation containing time sequence information is obtained after the characteristic words pass through GRU;
step c3, attention-drawing mechanism: using feature-based attention weights to measure user preference for featuresTo the firstThe attention weight of each feature is calculated by the formula:
3. the interpretable recommendation method fusing user implicit item preferences with explicit feature preferences according to claim 1, wherein: in the step A, order For the set of items preferred by the user,is shown asAn article, vector-sequencing the articleInputting a multi-layer perceptron to obtain implicit expression of a user。
5. The interpretable recommendation method fusing user implicit item preferences with explicit feature preferences according to claim 1, wherein: the step E and the step F also comprise the prediction of scores, and the final expression of the user is asThe final expression of the article isUser ofTo the articleThe prediction score of (a) is:。
7. The interpretable recommendation method fusing user implicit item preferences with explicit feature preferences according to claim 6, wherein: the model was evaluated using MAE and RMSE.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110747613.1A CN113420221B (en) | 2021-07-01 | 2021-07-01 | Interpretable recommendation method integrating implicit article preference and explicit feature preference of user |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110747613.1A CN113420221B (en) | 2021-07-01 | 2021-07-01 | Interpretable recommendation method integrating implicit article preference and explicit feature preference of user |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113420221A CN113420221A (en) | 2021-09-21 |
CN113420221B true CN113420221B (en) | 2022-09-09 |
Family
ID=77720060
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110747613.1A Active CN113420221B (en) | 2021-07-01 | 2021-07-01 | Interpretable recommendation method integrating implicit article preference and explicit feature preference of user |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113420221B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113850656B (en) * | 2021-11-15 | 2022-08-23 | 内蒙古工业大学 | Personalized clothing recommendation method and system based on attention perception and integrating multi-mode data |
CN114519097B (en) * | 2022-04-21 | 2022-07-19 | 宁波大学 | Academic paper recommendation method for heterogeneous information network enhancement |
WO2023225987A1 (en) * | 2022-05-27 | 2023-11-30 | 京东方科技集团股份有限公司 | Correlation degree prediction method and apparatus, and machine learning model training method and apparatus |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8977654B1 (en) * | 2012-09-21 | 2015-03-10 | Google Inc. | Assigning classes to users of an online community |
CN109902229B (en) * | 2019-02-01 | 2019-12-24 | 中森云链(成都)科技有限责任公司 | Comment-based interpretable recommendation method |
CN111488524B (en) * | 2020-04-08 | 2022-08-16 | 吉林大学 | Attention-oriented semantic-sensitive label recommendation method |
-
2021
- 2021-07-01 CN CN202110747613.1A patent/CN113420221B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN113420221A (en) | 2021-09-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113420221B (en) | Interpretable recommendation method integrating implicit article preference and explicit feature preference of user | |
CN111339415B (en) | Click rate prediction method and device based on multi-interactive attention network | |
CN108959603B (en) | Personalized recommendation system and method based on deep neural network | |
CN111222332B (en) | Commodity recommendation method combining attention network and user emotion | |
CN104935963A (en) | Video recommendation method based on timing sequence data mining | |
CN113158033A (en) | Collaborative recommendation model construction method based on knowledge graph preference propagation | |
CN111242729A (en) | Serialization recommendation method based on long-term and short-term interests | |
CN115917535A (en) | Recommendation model training method, recommendation device and computer readable medium | |
CN115048586B (en) | Multi-feature-fused news recommendation method and system | |
CN112288554B (en) | Commodity recommendation method and device, storage medium and electronic device | |
CN107016122A (en) | Knowledge recommendation method based on time-shift | |
Fakhfakh et al. | Deep learning-based recommendation: Current issues and challenges | |
Yakhchi et al. | Towards a deep attention-based sequential recommender system | |
Dai et al. | BTR: a feature-based Bayesian task recommendation scheme for crowdsourcing system | |
Khan et al. | Comparative analysis on Facebook post interaction using DNN, ELM and LSTM | |
CN113190751B (en) | Recommendation method fusing keyword generation | |
CN112364245B (en) | Top-K movie recommendation method based on heterogeneous information network embedding | |
CN111966888A (en) | External data fused interpretable recommendation method and system based on aspect categories | |
CN110851694A (en) | Personalized recommendation system based on user memory network and tree structure depth model | |
CN114817692A (en) | Method, device and equipment for determining recommended object and computer storage medium | |
CN115712780A (en) | Information pushing method and device based on cloud computing and big data | |
CN114358364A (en) | Attention mechanism-based short video frequency click rate big data estimation method | |
Mao et al. | TCR: Temporal-CNN for reviews based recommendation system | |
Li et al. | Research on recommendation algorithm based on e-commerce user behavior sequence | |
Ma et al. | In-depth Recommendation Model Based on Self-Attention Factorization. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |