CN112541128B - Personalized news recommendation method based on characteristic bidirectional dynamic cooperation - Google Patents
Personalized news recommendation method based on characteristic bidirectional dynamic cooperation Download PDFInfo
- Publication number
- CN112541128B CN112541128B CN202011365756.8A CN202011365756A CN112541128B CN 112541128 B CN112541128 B CN 112541128B CN 202011365756 A CN202011365756 A CN 202011365756A CN 112541128 B CN112541128 B CN 112541128B
- Authority
- CN
- China
- Prior art keywords
- news
- user
- vector
- sequence
- feature
- 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
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The personalized news recommendation method based on the characteristic bidirectional dynamic collaboration comprises a user-news bidirectional dynamic collaboration characterization module, a convolution layer, a pooling layer and a prediction layer. The user-news bidirectional dynamic collaborative representation module comprises a user dynamic collaborative representation network and a news dynamic collaborative representation network. For a user history browsing news sequence, a user characteristic matrix is obtained by virtue of a user dynamic collaborative representation network and an attention model, and the user characteristic matrix can reflect news content characteristics and browsing sequence characteristics of user history browsing and contains news-user interaction information. For candidate news, through a news dynamic collaborative representation network, the characteristics of historical news readers can be utilized on the basis of news content characteristics, user-news collaborative information is introduced, and a news characteristic matrix is obtained; the user browsing mode and the news audience characteristics can be mined simultaneously to obtain a user preference vector and a news collaborative feature vector; and finally, the prediction layer predicts the browsing probability of the candidate news by the user by using the vector.
Description
Technical Field
The invention belongs to the field of news recommendation, and particularly relates to a personalized news recommendation method based on feature bidirectional dynamic collaboration.
Background
Along with the rapid development of the mobile internet, people acquire information more and more conveniently, and prefer to read news information through an online news information platform. At present, millions of news articles are generated on an online information platform every day, and a lot of problems are caused while a large amount of fresh information is brought to people. The most serious problem is that the information is overloaded, and people are wrapped by massive information, so that the speed of obtaining effective information by people is reduced, and the time cost of inquiring the information is increased. Therefore, various online news information platforms often provide personalized news recommendation services for platform users by means of information recommendation technologies, so that the browsing experience of the users is improved, and adverse effects caused by information overload are reduced. The personalized news recommendation service recommends news contents meeting the interest and the like of the user according to the interest and the like of the user, and has the characteristics of initiative and personalization. Nowadays, a plurality of online news platforms including today's headlines all have personalized recommendation functions, and the personalized recommendation effect increasingly becomes important competitiveness of the platform, and has important influence on ensuring user stock and improving platform traffic. Personalized news recommendation improves the acquisition accuracy of the characteristics of users and news by utilizing the Internet technology, can better recommend interesting news for different users, and for the common people, the acquisition of interesting information by utilizing a personalized news recommendation service becomes an indispensable part in daily life.
Traditional information recommendation technologies such as content-based recommendation and collaborative filtering-based recommendation have been widely applied in other fields, but are not suitable for the recommendation requirements of the current online news platform. The news recommendation of the online news platform has the following characteristics: 1) the news magnitude is large, the updating speed is high, and the traditional recommendation technology cannot cope with the news data which changes rapidly; 2) the reading time of a user is sensitive, news is invalid quickly, the user has certain timeliness requirements on news, and the interest degree of historical news is reduced along with the change of time, so that a large amount of news is invalid at every moment.
The invention
Disclosure of Invention
In the face of dynamic news big data, how to design an effective personalized recommendation method for news recommendation is a key problem in the field of news recommendation. With the rapid development and application of deep learning theory, dynamic feature modeling for news and users based on deep learning theory is a main approach to solve the problem. The existing deep learning-based method starts from a news sequence and extracts abstract content information and sequence information to represent users and news, and the method enhances the model modeling capability, but depends on news content characteristics and ignores the bidirectional cooperative relationship between the users and the news. The two-way collaboration relationship, i.e., user and news, can be characterized with each other, reflecting the implicit characteristics of each other, i.e., similar users have similar reading interests and similar news have similar audiences. The bidirectional collaboration relationship contains rich interaction relationship, and is an important supplement for the representation of basic news content. The existing deep learning-based method does not always consider the relationship, so that the personalized news recommendation method based on the feature bidirectional dynamic cooperation is provided, and the recommendation effect is further improved.
The invention aims to disclose a personalized news recommendation method based on feature bidirectional dynamic collaboration, which aims at defects existing in current personalized news recommendation and develops work around high-quality feature representation theories and methods of users and news. The achievement of the invention enriches and expands news recommendation methods and theories, improves news recommendation effect, improves the experience of users suitable for online news information platforms, and indirectly creates more commercial values for the platforms.
The technical scheme is as follows:
a personalized news recommendation method based on feature bidirectional dynamic collaboration is characterized in that a main module comprises a user-news bidirectional dynamic collaborative characterization module, a convolutional layer, a pooling layer and a prediction layer, wherein the user-news bidirectional dynamic collaborative characterization module comprises a user dynamic collaborative characterization network and a news dynamic collaborative characterization network; for a user history browsing news sequence, a user characteristic matrix can be obtained by virtue of a user dynamic collaborative representation network and an attention model, and can reflect news content characteristics and browsing sequence characteristics of user history browsing, including news-user interaction information; for candidate news, through the news dynamic collaborative representation network, the characteristics of historical news readers can be utilized on the basis of news content characteristics, user-news collaborative information is introduced, and a news characteristic matrix is obtained finally; deep features in the user and news feature matrix are learned through the convolutional layer and the pooling layer, user browsing modes and news audience features in the user and news feature matrix can be simultaneously mined, and user preference vectors and news collaborative feature vectors are obtained; and finally, the prediction layer predicts the browsing probability of the candidate news by the user by using the vector.
The invention carries out deep and systematic research on the news recommendation field, fully utilizes the user-news cooperative relationship, introduces an attention mechanism and a convolutional neural network, provides a new thought and method for the representation of news and users in recommendation, expands the theoretical method of news recommendation and can further improve the news recommendation effect.
Advantageous effects
1) The invention provides a method for bidirectional dynamic collaborative representation of a user and news, aiming at the problems of relatively single feature representation and the like in the conventional news recommendation, which comprises the following steps: for user characterization, performing feature modeling on a historical news sequence by using an attention mechanism; for news representation, the characteristics of a news history reader are introduced, and the characteristics are fused with news content characteristics by using a multilayer attention network, so that the bidirectional cooperation of users and news is realized, and the quality of the users and the news representation is improved.
2) The invention deeply extracts the user and news representations by using the convolutional neural network, can mine richer user-news interaction relation on the basis of bidirectional collaborative representation, and extracts abstract user browsing modes and news audience characteristics, thereby improving the recommendation effect.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and not to limit the invention. In the drawings:
FIG. 1 Overall framework of the invention
FIG. 2 is a schematic diagram of a network structure for dynamic collaborative representation of users
FIG. 3 is a schematic diagram of a news dynamic collaborative representation network structure
Detailed Description
The following detailed description of the embodiments of the present invention will be provided with reference to the accompanying drawings and examples, so that how to implement the embodiments of the present invention by using technical means to solve the technical problems and achieve the technical effects can be fully understood and implemented.
The invention discloses a personalized news recommendation method based on characteristic bidirectional dynamic collaboration, which predicts the probability of clicking new news (namely candidate news) by a user according to the record of the historical browsing news of the user. The invention defines the news sequence (in ascending order of time) recently browsed by the user u as Cu＝{cu，1，cu，2，...，cu，kObtaining the vector of each word in corresponding content by using a glove word vector tool, and calculating an average vector for all word vectors of each news to obtain a news content vector, wherein the news content vector of the example news cu, 1 is represented as a d-dimensional vector
For candidate news x, the invention defines the sequence of users who have recently browsed news x, i.e. the sequence of readers (sorted in ascending order of time), as Ux＝{ux，1，ux，2，...，ux，m}. For a given user u and candidate news x, the task of news recommendation is to predict news click probability
First part, user-news bidirectional dynamic collaborative representation module
1.1 dynamic collaborative characterization of networks
1) Constructing a user historical news characteristic sequence; for user u, take k news records C recently browsed by user uu＝{cu，1，cu，2，...，cu，kAnd obtaining a corresponding news content vector sequence:
2) constructing a user interest characteristic matrix; for news content sequence VuThe present invention utilizes a self-attention mechanism to capture content relevance among a user's historical news. For VuThe ith news content vector of
Defining the corresponding attention vector of the vector as
The calculation is performed using the following formula:
wherein
In order to take care of the weight of attention,
and is
The weight is calculated by:
wherein WaFor the weight matrix parameters to be trained, the tanh function is a hyperbolic tangent function, and is defined as follows:
3) Splicing the attention vector with the original vector to obtain an updated content vector:
such a vector contains not only news content information, but also news content correlations in the user's historical reading sequence. For the new sequences obtained
The invention regards the user as a feature matrix according to the sequence, which contains the content features of the user's historical reading and browsing sequence features, and is called as a user feature matrix.
1.2 News dynamic collaborative representation network
1) Constructing a news history user feature vector: for candidate news x, m users who have recently browsed the news are acquired: u shapex＝{ux，1，ux，2，...，ux，mFor each user, the method utilizes an attention mechanism to extract the content characteristics of historical news of the user, and for the historical user ux，iIts user feature vector uxiThe calculation method is as follows:
wherein Vb，vb，qTAll are model parameters, and the news contents browsed by the historical users can be processed by an attention mechanismAnd (5) extracting the characteristics to obtain historical user characteristics containing candidate news.
2) And (3) news characteristic sequence construction: in order to introduce the cooperative information, the news is represented by using the historical user characteristics and the news characteristics together, the interactive relation between the news and the user can be reflected by the historical user information of the news, and the characteristics of the news audience are constructed; in addition, news content characteristics can reflect news characteristics more comprehensively and fully. Corresponding to the user characteristic characterization method, the characterization method of the user-news bidirectional collaboration is realized together. For the obtained historical user feature vector sequence: { ux1，ux2，...，uxmAnd the invention further fuses the user characteristics with the news content characteristics by utilizing a multilayer attention network, and learns the correlation between the historical user characteristics of the news and the content characteristics of the news.
For historical user feature vector uxiThe invention uses a binary attention mechanism to calculate the candidate news content vector vxIs given to the correlation vector Attn _ uxi：
Wherein Vx，qTAre all the parameters of the model, and are,
is the attention weight coefficient.
In addition, the method fuses the historical user feature sequence and the news feature vector based on the attention mechanism, and the historical user features { u } ux1，ux2，...，uxm}，The news attention vector Attn _ v is obtained by calculation according to the following formulax：
Wherein WxModel parameters that need to be trained.
3) On the basis of the calculation, the attention vector and the original vector are spliced to obtain a new candidate news feature vector sequence containing historical user features and news and user associated information: o isx＝{ox，1，ox，2，...，ox，m+1Where for i < m +1, ox，i＝[Attn_uxi，uxi]And the last vector of the sequence is ox，m+1＝[Attn_vx，vx]. The present invention treats the sequence as a matrix OxSuch feature matrix takes into account the dynamics of news, including news content and news-user interaction information, and is referred to as a news feature matrix in the present invention.
Second part, convolutional layer and pooling layer
After the bidirectional collaborative characterization modeling of the user and the article is completed, the interactive relation extraction is carried out on the obtained feature matrix of the user and the news by utilizing the convolutional neural network, and deep user reading preference, news audience and content features are mined.
The invention adopts a shared weight design in the convolutional layer and inputs two matrixes into the same convolutional layer. In a specific convolution operation, the length of the convolution kernel is set to be consistent with the size of the row vector. The direction of convolution calculation is longitudinal. The invention adopts single-layer convolution, thereby avoiding the rapid loss of information caused by multi-layer convolution.
Input news feature matrix OxThe size is (m +1) multiplied by 2d, and the user characteristic matrix NuSize k × 2d, for convenience of calculation, the two matrices are uniformly lengthened to be L multiplied by 2 d. The convolutional layer is operated from top to bottom with a sliding window of width n and length 2 d. Specifically, the convolution operation is calculated as follows
convi＝f(W·ci：n+b)
Wherein b represents a bias term, ci：nAnd (3) representing the splicing of the sequence vectors, namely splicing the ith sequence vector to the (i + n-1) th sequence vector. Setting the step length to be 1, sliding the whole matrix from top to bottom according to the setting of the sliding amplitude 1, and obtaining the result of [ conv ] through one complete convolution operation1，conv2，...，convL-n_1]And forming a characteristic diagram. The invention simultaneously adopts a plurality of convolution cores with different widths to process the matrix, and combines the corresponding characteristic graphs to obtain the output of the convolution layer.
After the convolutional layer processing is completed, the method uses the pooling layer to perform local extraction on the features, so that the feature dimension is reduced. Considering that the information loss is serious due to a calculation method of maximizing pooling, the calculation is performed by using mean pooling, and a final vector is obtained through a pooling layer.
Obtaining a user preference vector n 'through the convolution and pooling operation'uAnd candidate News collaborative feature vector o'x。
Third part, model prediction and training
Obtaining a feature vector n 'for the above processing'uAnd o'xThe invention uses the cosine similarity calculation method to measure the relation between the two to obtain the predicted value
And indicating the probability value of clicking candidate news x by the user u, and recommending by using the predicted value. Assuming that the vector size is M, the calculation is as follows.
In model training, the invention uses a training method based on a pairwise method, and a data set is
Therefore, it is necessary to construct a training triplet < v, i+，i->. correspond to the user, and their browsed news (positive examples) and their unviewed news (negative examples), respectively. In practice, a negative sample is formed by randomly drawing 5 news for each user that they have not browsed. The invention uses the following objective function and utilizes a random gradient descent algorithm to optimize the model parameters:
L＝∑(u，i，j)∈Dmax{0，m-[cos(u，i)-cos(u，j)]}+λΩ(Θ)
wherein m is a positive and negative sample difference factor, which needs to be adjusted and set according to the actual data set condition, λ is a model regularization hyper-parameter to prevent overfitting, Ω (Θ) is an L2 regularization function, and regularization processing is performed on model parameters such as weights in convolution.
The overall schematic diagram of this embodiment is shown in fig. 1.
While the foregoing specification shows and describes several preferred embodiments of this invention, it is to be understood, as noted above, that the invention is not limited to the forms disclosed herein, but is not to be construed as excluding other embodiments and is capable of use in various other combinations, modifications, and environments and is capable of changes within the scope of the inventive concept as expressed herein, commensurate with the above teachings, or the skill or knowledge of the relevant art. And that modifications and variations may be effected by those skilled in the art without departing from the spirit and scope of the invention as defined by the appended claims.
Innovation point
One of the innovations is as follows: depth feature fusion model
In the traditional recommendation, the user and the news are characterized relatively independently, the user is often represented by using a historical browsing record, and the news is only represented by inherent text information. According to the depth feature fusion model provided by the invention, the user features are utilized to represent news, so that the bidirectional fusion representation of the user-news features is realized, the representation of the news is more dynamic, the information utilization is more sufficient, and a new feature representation thought is provided for news recommendation.
The second innovation is that: convolutional neural network-based interactive relationship extraction
On the basis of deep feature fusion, the convolutional neural network is used for extracting features of the feature matrix, and compared with neural network models such as a multilayer perceptron and the like used in other existing models, the convolutional neural network used in the method is simple in structure, few in parameters and very efficient.
The deep extraction of the interactive relation in the feature matrix can be completed by utilizing the convolutional neural network, the realization method expands the model construction mode of the conventional news recommendation, and meanwhile, the recommendation effect can be effectively improved.
Claims (3)
1. A personalized news recommendation method based on feature bidirectional dynamic collaboration is characterized in that a main module comprises a user-news bidirectional dynamic collaborative characterization module, a convolutional layer, a pooling layer and a prediction layer, wherein the user-news bidirectional dynamic collaborative characterization module comprises a user dynamic collaborative characterization network and a news dynamic collaborative characterization network; for a user history browsing news sequence, a user characteristic matrix can be obtained by virtue of a user dynamic collaborative representation network and an attention model, and can reflect news content characteristics and browsing sequence characteristics of user history browsing, including news-user interaction information; for candidate news, through the news dynamic collaborative representation network, the characteristics of historical news readers can be utilized on the basis of news content characteristics, user-news collaborative information is introduced, and a news characteristic matrix is obtained finally; deep features in the user and news feature matrix are learned through the convolutional layer and the pooling layer, user browsing modes and news audience features in the user and news feature matrix can be simultaneously mined, and user preference vectors and news collaborative feature vectors are obtained; finally, the prediction layer predicts the browsing probability of the candidate news by the user by using the vector;
the specific implementation method comprises the following steps:
defining the news sequence which is browsed by the user u recently as Cu＝{cu,1,cu,2,…,cu,kAnd each news comprises news title and text content, a glove word vector tool is used for obtaining the vector of each word in corresponding content, an average vector is calculated for all word vectors of each news to obtain a news content vector, and news cu,1Is expressed as a d-dimensional vector
For candidate news x, the reader sequence, which is the sequence of users who have recently browsed news x, is defined as Ux＝{ux,1,ux,2,…,ux,m}；
For a given user u and candidate news x, the task of news recommendation is to predict news click probability
Step 1, designing a user-news bidirectional dynamic collaborative representation module
Step 1.1 dynamic collaborative representation of the network by the user
1) Constructing a user historical news characteristic sequence; for user u, take k news records C recently browsed by user uu＝{cu,1,cu,2,…,cu,kAnd obtaining a corresponding news content vector sequence:
2) constructing a user interest characteristic matrix; for news content sequence VuCapturing content correlation among historical news of the user by using a self-attention mechanism; for VuThe ith news content vector of
Defining the corresponding attention vector of the vector as
The calculation is performed using the following formula:
wherein
In order to take care of the weight of attention,
and is provided with
The weight is calculated by:
wherein WaFor the weight matrix parameters to be trained, the tanh function is a hyperbolic tangent function, and is defined as follows:
3) Splicing the attention vector with the original vector to obtain an updated content vector:
for the new sequences obtained
Regarding the content characteristics and browsing sequence characteristics as a characteristic matrix according to the sequence, wherein the characteristic matrix comprises the content characteristics and browsing sequence characteristics of the historical reading of the user and is called a user characteristic matrix;
step 1.2 News dynamic collaborative representation network
1) Constructing a news history user feature vector: for candidate news x, m users U which have recently browsed the candidate news x are obtainedx＝{ux,1,ux,2,…,ux,mFor each user, extracting content characteristics of historical news of the user by using an attention mechanism, and for historical users ux,iIts user feature vector uxiThe calculation method is as follows:
wherein Vb,vb,qTAll the parameters are model parameters, and the feature extraction can be carried out on news contents browsed by a historical user through an attention mechanism to obtain the historical user features containing candidate news;
2) and (3) news characteristic sequence construction: in order to introduce collaborative information, news is characterized by using news history user characteristics and news characteristics together, and the news history user information can reflect news and usageBuilding characteristics of news audiences by the interaction relation of users; in addition, news content characteristics can reflect news characteristics more comprehensively and fully; corresponding to the characterization method of the user dynamic collaborative characterization network in the step 1.1, the characterization method of user-news two-way collaboration is realized together; for the obtained historical user feature vector sequence: { ux1,ux2,…,uxmMerging the user characteristics with news content characteristics by utilizing a multilayer attention network, and learning the correlation between the historical user characteristics of news and the content characteristics of the news;
for historical user feature vector uxiComputing the candidate news content vector v by using a binary attention mechanismxIs given to the correlation vector Attn _ uxi：
Wherein Vx,qTAre all the parameters of the model and are all the parameters of the model,
is an attention weight coefficient;
fusing a historical user feature sequence and a news feature vector based on an attention mechanism, and aiming at the historical user features { u }x1,ux2,…,uxmThe news attention vector Attn _ v is obtained by calculation according to the following formulax：
Wherein WxModel parameters needing to be trained;
2) and splicing the attention vector with the original vector to obtain a new candidate news feature vector sequence containing historical user features and news and user associated information: o isx＝{ox,1,ox,2,…,ox,m+1Wherein for i<m+1,Ox,i＝[Attn_uxi,uxi]And the last vector of the sequence is Ox,m+1＝[Attn_vx,vx](ii) a Treat the sequence as a matrix OxThe matrix is called a news characteristic matrix;
step 2, designing the convolution layer and the pooling layer
After the bidirectional collaborative characterization modeling of the user and the article is completed, extracting the interactive relation of the obtained feature matrix of the user and the news by using a convolutional neural network, and mining the reading preference of deep users and the characteristics of news audiences and contents;
step 3, model prediction and training
Deriving a feature vector n 'for the processing'uAnd o'xMeasuring the relation between the two by using a cosine similarity calculation method to obtain a predicted value
Representing the probability value of clicking candidate news x by the user u, and recommending by using the predicted value;
specifically, the convolution operation is calculated as follows
convi＝f(W·ci:n+b)
Wherein b represents a bias term, ci:nRepresenting concatenation of sequence vectors, i.e. i-th sequence vector to i + n-1Splicing sequence vectors; setting the step length to be 1, sliding the whole matrix from top to bottom according to the setting of the sliding amplitude 1, and obtaining the result of [ conv ] through one complete convolution operation1,conv2,…,convL-n_1]Forming a characteristic diagram; simultaneously, processing the matrixes by adopting a plurality of convolution kernels with different widths, and merging corresponding characteristic graphs to obtain the output of the convolution layer;
after the convolutional layer processing is finished, the pooling layer is used for carrying out local extraction on the features, so that feature dimensions are reduced; obtaining a final vector through a pooling layer;
obtaining a user preference vector n 'through the convolution and pooling operations'uAnd candidate News collaborative feature vector o'x；
The step 3:
assuming that the vector size is M, the calculation is as follows:
in model training, a pair-by-pair-based training method is used, and a data set is
Requires the construction of a training triplet < u, i+,i->. correspond to the user, his browsed news positive sample and his unviewed news negative sample, respectively; the model parameters were optimized using a stochastic gradient descent algorithm using the following objective function:
L＝∑(u,i,j)∈Dmax{0,m-[cos(u,i)-cos(u,j)]}+λΩ(Θ)
wherein m is a positive and negative sample difference factor, which needs to be adjusted and set according to the actual data set condition, λ is a model regularization hyper-parameter to prevent overfitting, Ω (Θ) is an L2 regularization function, and regularization processing is performed on model parameters such as weights in convolution.
2. The method of claim 1, wherein the step 2:
adopting a shared weight design in the convolutional layer, and inputting two matrixes into the same convolutional layer; during the specific convolution operation, the length of the convolution kernel is set to be consistent with the size of the row vector, and the direction of the convolution calculation is longitudinal.
3. The method of claim 2, wherein the input news feature matrix OxThe size is (m +1) multiplied by 2d, and the user characteristic matrix NuThe size is k multiplied by 2d, and the two matrixes are uniformly lengthened to be changed into the size of L multiplied by 2 d; the convolutional layer is operated from top to bottom with a sliding window of width n and length 2 d.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010926346X | 2020-09-07 | ||
| CN202010926346 | 2020-09-07 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112541128A CN112541128A (en) | 2021-03-23 |
| CN112541128B true CN112541128B (en) | 2022-05-13 |
Family
ID=75015327
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011365756.8A Active CN112541128B (en) | 2020-09-07 | 2020-11-29 | Personalized news recommendation method based on characteristic bidirectional dynamic cooperation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112541128B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113254503B (en) * | 2021-06-08 | 2021-11-02 | 腾讯科技（深圳）有限公司 | Content mining method and device and related products |
| CN114741608A (en) * | 2022-05-10 | 2022-07-12 | 中国平安财产保险股份有限公司 | News recommendation method, device, equipment and storage medium based on user portrait |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106446195A (en) * | 2016-09-29 | 2017-02-22 | 北京百度网讯科技有限公司 | News recommending method and device based on artificial intelligence |
| CN109874053A (en) * | 2019-02-21 | 2019-06-11 | 南京航空航天大学 | The short video recommendation method with user's dynamic interest is understood based on video content |
| CN110032679A (en) * | 2019-04-16 | 2019-07-19 | 北京航空航天大学 | A method of the dynamic news based on level attention network is recommended |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10127563B2 (en) * | 2011-09-15 | 2018-11-13 | Stephan HEATH | System and method for providing sports and sporting events related social/geo/promo link promotional data sets for end user display of interactive ad links, promotions and sale of products, goods, gambling and/or services integrated with 3D spatial geomapping, company and local information for selected worldwide locations and social networking |
-
2020
- 2020-11-29 CN CN202011365756.8A patent/CN112541128B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106446195A (en) * | 2016-09-29 | 2017-02-22 | 北京百度网讯科技有限公司 | News recommending method and device based on artificial intelligence |
| CN109874053A (en) * | 2019-02-21 | 2019-06-11 | 南京航空航天大学 | The short video recommendation method with user's dynamic interest is understood based on video content |
| CN110032679A (en) * | 2019-04-16 | 2019-07-19 | 北京航空航天大学 | A method of the dynamic news based on level attention network is recommended |
Non-Patent Citations (2)
| Title |
|---|
| "基于用户向量化表示和注意力 机制的深度神经网络推荐模型";郭旭;《计算机科学》;20190831;第46卷(第8期);第111-115页 * |
| "融合动态协同过滤和深度学习的推荐算法";邓存彬;《计算机科学》;20190831;第46卷(第8期);第28-33页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112541128A (en) | 2021-03-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110046304B (en) | User recommendation method and device | |
| Reddy et al. | Content-based movie recommendation system using genre correlation | |
| CN110309427B (en) | Object recommendation method and device and storage medium | |
| WO2022161202A1 (en) | Multimedia resource classification model training method and multimedia resource recommendation method | |
| Mohamed et al. | Content-based image retrieval using convolutional neural networks | |
| CN114817663B (en) | Service modeling and recommendation method based on class perception graph neural network | |
| CN112541128B (en) | Personalized news recommendation method based on characteristic bidirectional dynamic cooperation | |
| CN113590970B (en) | Personalized digital book recommendation system and method based on reader preference, computer and storage medium | |
| CN112800344B (en) | Deep neural network-based movie recommendation method | |
| CN114201516B (en) | User portrait construction method, information recommendation method and related devices | |
| CN118119954A (en) | Hint adjustment using one or more machine learning models | |
| Vilakone et al. | Movie recommendation system based on users’ personal information and movies rated using the method of k-clique and normalized discounted cumulative gain | |
| CN115964568A (en) | Personalized recommendation method based on edge cache | |
| Amara et al. | Cross-network representation learning for anchor users on multiplex heterogeneous social network | |
| CN115795044A (en) | Knowledge injection-based user relationship mining method and device | |
| CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system | |
| CN116776003A (en) | Knowledge graph recommendation method, system and equipment based on comparison learning and collaborative signals | |
| Alabdulrahman et al. | Active learning and deep learning for the cold-start problem in recommendation system: A comparative study | |
| Zhang et al. | Probabilistic matrix factorization recommendation of self-attention mechanism convolutional neural networks with item auxiliary information | |
| Zeng et al. | RACMF: robust attention convolutional matrix factorization for rating prediction | |
| Batchanaboyina et al. | Design and evaluation of outlier detection based on semantic condensed nearest neighbor | |
| Wu et al. | Applying a Probabilistic Network Method to Solve Business‐Related Few‐Shot Classification Problems | |
| Samih et al. | Deep graph embeddings for content based-book recommendations | |
| CN117786234B (en) | Multimode resource recommendation method based on two-stage comparison learning | |
| CN115098767B (en) | News recommendation method based on interest perception and user similarity |
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 |