A news recommendation method and system based on candidate-perceived fine-grained interactive information

Abstract

The present invention belongs to the field of news recommendation and provides a news recommendation method and system based on fine-grained interaction information of candidate perception. To address the problem that similar news is difficult to distinguish when forming vector representations, the system fully integrates the features of candidate news and historical news to better match the two, and proposes a method to solve the problem of similar vector representations of similar news to achieve accurate matching. Through a fine-grained interaction module, the text fragments of candidate news and each historical news are matched at each semantic granularity. This enables interaction between the two at the word level, helping similar historical news to form more specific and accurate representations. In addition, an activation unit module is used to establish a connection between user historical behavior and candidate news to construct a user interest representation that is less affected by category imbalance.

Description

Technical Field

[0001] The present invention belongs to the field of news recommendation, and in particular relates to a news recommendation method and system based on candidate-perceived fine-grained interactive information. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] News recommendation, as an important branch of the recommendation system research field, aims to help users find news that matches their interests as much as possible through news content and user information. Existing methods usually model candidate news based on their content, independently model user interests based on clicked news, and then match candidate news and user interests based on their relevance. For example, An et al. used GRU networks and ID embedding to learn user interest representations from clicked news. Wu et al. learned user interest representations by aggregating news clicked by users and applying an attention network. Yang et al. implemented hierarchical user interest modeling to capture users' multi-level interests.

[0004] These methods are extremely helpful for news recommendation, but they all model user interests in a candidate-agnostic manner, meaning click predictions are based solely on the similarity between user clicks and candidate news items. This can be affected by an imbalance in news categories within the dataset, resulting in biased representations of user interests.

[0005] Furthermore, simply incorporating candidate news into user interest modeling is not conducive to learning differentiated representations of similar news. In fact, the imbalance in news categories within the dataset may result in different exposure rates for different news categories. For example, among users' clicked news history, fewer health-related news items are associated with health, while more sports-related news items are associated with health. This suggests that, on the one hand, users tend to prefer sports news, but on the other hand, users may click on sports news due to its higher exposure. Because sports news contains more information, constructing user interest representations often prioritizes users' interest in sports news while ignoring their interest in health news, resulting in biased user representations. Summary of the Invention

[0006] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a news recommendation method and system based on candidate-perceived fine-grained interactive information, which obtains better interest matching by considering the differences of similar news. Based on the fine-grained interactive information, the candidate news and the text fragment of each historical news are matched at each semantic granularity to realize the interaction between the two at the word level, helping to form a more specific and accurate representation of similar historical news.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A first aspect of the present invention provides a news recommendation method based on candidate-perceived fine-grained interactive information, comprising the following steps:

[0009] Obtain historical click news data and candidate news data;

[0010] Based on historical click-through news data, candidate news data, and the trained news recommendation model, the probability of clicking on a candidate news item is obtained, and the candidate news item with the highest probability is recommended to the user.

[0011] The construction process of the news recommendation model is as follows:

[0012] Extract multi-scale feature maps of historical click news data and candidate news data respectively;

[0013] Based on this multi-scale feature map, a fine-grained interaction module was constructed. At each semantic level, the matching features between candidate news and each historical click news were captured at different granularity levels to distinguish the vector representations of similar historical click news and obtain the user's first interest feature representation.

[0014] Combining this multi-scale feature map, the core activation unit is used to learn the user's second interest feature representation from the candidate news;

[0015] The probability of clicking on the candidate news is predicted based on the user's first interest feature representation and the user's second interest feature representation.

[0016] A second aspect of the present invention provides a news recommendation system based on candidate-perceived fine-grained interactive information, comprising:

[0017] A data acquisition module is configured to: acquire historical click news data and candidate news data;

[0018] A news recommendation module is configured to: obtain the probability of clicking on a candidate news item based on historical click-through news data, candidate news data, and a trained news recommendation model, and recommend the candidate news item with the highest probability to the user;

[0019] The construction process of the news recommendation model is as follows:

[0020] Extract multi-scale feature maps of historical click news data and candidate news data respectively;

[0021] Based on this multi-scale feature map, a fine-grained interaction module was constructed. At each semantic level, the matching features between candidate news and each historical click news were captured at different granularity levels to distinguish the vector representations of similar historical click news and obtain the user's first interest feature representation.

[0022] Combining this multi-scale feature map, the core activation unit is used to learn the user's second interest feature representation from the candidate news;

[0023] The probability of clicking on the candidate news is predicted based on the user's first interest feature representation and the user's second interest feature representation.

[0024] A third aspect of the present invention provides a computer-readable storage medium.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a news recommendation method based on candidate-perceived fine-grained interactive information as described in the first aspect.

[0026] A fourth aspect of the present invention provides a computer device.

[0027] 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 program, the steps of the news recommendation method based on candidate-perceived fine-grained interactive information as described in the first aspect are implemented.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. Similar news items are difficult to distinguish when forming vector representations. Simply incorporating candidate news into user interest modeling without considering word-level information fails to provide sufficient differentiation for recommending appropriate news items. To address this issue, we propose a news recommendation method with candidate-aware fine-grained interactive information. This method fully integrates the features of candidate and historical news to better match the two, addressing the problem of similar vector representations for similar news items and achieving precise matching.

[0030] 2. This invention uses a fine-grained interaction module to match candidate news items with text snippets of historical news at each semantic granularity, enabling word-level interaction between the two and helping to create more specific and accurate representations of similar historical news items. Furthermore, an activation unit module establishes a connection between user historical behavior and candidate news items, constructing a user interest representation that is less affected by category imbalance.

[0031] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0033] Figure 1 This is an example of a user reading a piece of news;

[0034] Figure 2 This is a flowchart of a news recommendation method provided by an embodiment of the present invention;

[0035] Figure 3 It is a structural diagram of an activation unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0038] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0039] Due to the imbalance of news categories in the dataset, the probability of different categories of news being exposed to users may be different.

[0040] like Figure 1 As shown, the representation of similar news should be more discriminative. For user U, they are only interested in news related to singer A's songs. D11 and D13 are both news related to A, but one is about a new song release, and the other is about some of his personal life experiences. If candidate news were simply added to the user's interest representation, although the resulting news representation would all be about A, it would not be specific as to whether it was news about his songs. Therefore, it is essential to interact candidate news with historical news at the word level. This is because by considering the interaction between words, the differences between similar news can be taken into account.

[0041] The present invention provides a fine-grained interaction module designed to capture the matching features between candidate news and each historical news from different granularity levels to distinguish similar historical news and help users form more accurate representations. In addition, in order to form a user interest representation with less deviation when facing the user's historical click news affected by category imbalance, an activation unit structure is designed to establish the relationship between the user's historical click news and candidate news, and fully integrate the features of candidate news and each historical news. During the model training process, negative sampling technology is applied to K+1 news to predict click scores, that is, the news consists of a positive sample of one user and a randomly selected negative sample of one user. Then, positive news is jointly predicted. and Negative news.

[0042] Example 1

[0043] like Figure 2 As shown, this embodiment provides a news recommendation method based on candidate-perceived fine-grained interactive information, including the following steps:

[0044] 1. Problem Definition

[0045] Given a user u, it is composed of a set of historical click news browsed by the user, denoted as D u =[D1,D2,…,D N ]; where N is the number of news items clicked by the user in history. The goal is to calculate the probability that user u clicks on a set of candidate news items. Recommend the best news by sorting the candidate news according to their click scores. C =[D c1 ,D c2 ,…,D cM ], where cM is the probability of the number of candidate news

[0046] 2. News Modeling Module

[0047] The news modeling module aims to learn deep semantic representations of candidate news.

[0048] Each news title D has M words, namely: D i =[w1,w2,...,w M ];

[0049] The specific process of learning the deep semantic representation of candidate news includes:

[0050] Step 201: Input the historical click news and candidate news titles into the BERT model, and obtain the tokens representation of the hidden layer corresponding to each news item, which is recorded as E w =[e1,e2,...,eM ];

[0051] Step 202: The obtained low-dimensional vector representation of the news headline is fed into a dilated convolutional layer (HDC) and an Att-Bi-LSTM layer to obtain a reinforced semantic representation and a multi-granular semantic representation. The reinforced semantic representation and the multi-granular semantic representation are then fused to obtain a deep semantic representation of the historical clickbait news and a deep semantic representation of the candidate news.

[0052] Specifically, in order to further strengthen the semantic relationship, the hidden layer representation E w =[e1,e2,...,e M ] is sent to Bi-LSTM, and an attention network is used to aggregate the hidden representations output by Bi-LSTM to obtain news representations rich in contextual semantic information. Bl ; Among them, the representations of historical click news and candidate news are respectively recorded as and

[0053] At the same time, a dilated convolutional layer (HDC) is applied to obtain multi-granularity semantic features in news text.

[0054] Assuming there are L dilated convolutional layers, the multi-scale news representation of the i-th news is: Since the importance of representations at different scales is different, the multi-scale representations of each news item are simply concatenated and fed into the attention mechanism to obtain r H ; Among them, the representations of historical click news and candidate news are respectively recorded as and

[0055] Step 203: Part of the news Bl and global information r H Fusion to the final news representation r: r = Concat (r Bl +r H ), to enrich the semantic representation of news.

[0056] 3. User modeling module

[0057] The user modeling module aims to learn the representation of user interests from the news browsed by the user. It consists of two parts, such as Figure 2 As shown in the two modules B and C.

[0058] The core activation unit of the first step is as follows Figure 2 As shown in the C module, it is able to learn the representation of candidate news that is aware of user interests.

[0059] The specific implementation is as follows:

[0060] Step 301: In order to better integrate the information of candidate news and historical news and enhance the semantic representation, the deep semantic representation of candidate news is transformed into c Make a vector copy and compare it with the historical news i The enhanced matrix V is obtained by splicing the three matrices and the difference between the two i =Concat(r i ,r c ,r i -r c ), input three-layer fully connected neural network, such as Figure 3 , the purpose of this structure is to learn the relationship between each historical news and candidate news. i represents the deep semantic features of the i-th historical click news, r c Deep semantic representations for candidate news.

[0061] Step 302: To fully interact with candidate news and historical news information, i Send it into the three-layer fully connected neural network and get V i 1 =W1V i +b1,V i 2 =W2V i 1 +b2 and u i =W3V i 2 +b3, where W1, b1, W2, b2, W3, b3 are all learnable parameters. i Represents the final user interest feature representation.

[0062] The second part Figure 2 As shown in Module B in Figure 1, word-level interactions between candidate news and historical news are equally important, enabling better differentiation between similar historical clicked news. Therefore, capturing the features of each interactive news item at a finer granularity and inferring user interests from them has become a pressing issue.

[0063] Step 303: After the dilated convolutional layer (HDC), the k-th historical click news d k The multi-scale feature map can be expressed as:

[0064]

[0065] in, Represents the semantic representation of the k-th browsed news on the l-th layer.

[0066] The multi-scale feature map of candidate news can be expressed as:

[0067]

[0068] At the lth semantic level, kth browsed news d k The semantic matching matrix of the multi-scale feature map of and the multi-feature map of the candidate news is:

[0069]

[0070] in, Chinese elements It can be understood as the correlation between the i-th field of the k-th news and the j-th field of the candidate news c at the l-th semantic level. represents the specific granular news features of the K-th historical news at the l-th semantic level of HDC, C l represents the specific granular news features of the candidate news at the lth semantic level of HDC, fs represents the number of convolutional filters in each layer, N dk Indicates the length of the k-th historical news, N c Indicates the length of the candidate news.

[0071] So for a news pair <d k ,c>, there will be L+1 matching matrices, and the interaction matrix between the kth historical news and candidate news can be expressed as

[0072] Step 304: The interaction matrix Q of k historical news and candidate news k,c Feed it into a 2D multi-channel convolutional CNN to identify significant matching signals:

[0073]

[0074] in, is the concatenation operator, Represents the vector from q i-2w',j-w' to q i,j+w' , and b f ∈R kn' are the kernel and bias parameters of the CNN filter, kn' and 2w'+1 are the number of filters and window size.

[0075] In this embodiment, kn' is set to be the same as L+1.

[0076] Step 305: Finally, the fine-grained interaction score between the i-th historical click news and the candidate news is obtained:

[0077]

[0078] In order to further distinguish similar historical news, the obtained score is multiplied with the original one to form a more accurate historical news representation based on the participation of candidate news:

[0079]

[0080] 4. Click prediction module

[0081] The final representation of user interest is the weighted sum of the browsed news representations:

[0082]

[0083]

[0084] Among them, W u ,b u and q u is a learnable parameter and N is the number of news viewed.

[0085] In obtaining the candidate news representation r c After obtaining the final representation u of user interest, the probability score of the user clicking on the candidate news can be predicted by the inner product

[0086] In this embodiment, negative sampling technology is used for the following training. In the same session, each news browsed by the user is regarded as a positive sample, and then K exposed news that are not clicked by the user are randomly sampled as negative samples. The click probability score of the i-th positive news in the same period is expressed as The click-through rate score of the jth piece of K negative news is expressed as The news click prediction problem can be considered as a pseudo K+1-way classification task.

[0087] Apply the softmax function to normalize these click probability scores to the posterior click probability of the positive sample;

[0088]

[0089] The loss function L is the negative log-likelihood of all positive samples S:

[0090]

[0091] Recommend news based on predicted click probability scores.

[0092] Example 2

[0093] This embodiment provides a news recommendation system based on candidate-perceived fine-grained interactive information, including:

[0094] A data acquisition module is configured to: acquire historical click news data and candidate news data;

[0095] A news recommendation module is configured to: obtain the probability of clicking on a candidate news item based on historical click-through news data, candidate news data, and a trained news recommendation model, and recommend the candidate news item with the highest probability to the user;

[0096] The construction process of the news recommendation model is as follows:

[0097] Extract multi-scale feature maps of historical click news data and candidate news data respectively;

[0098] Based on this multi-scale feature map, a fine-grained interaction module was constructed. At each semantic level, the matching features between candidate news and each historical click news were captured at different granularity levels to distinguish the vector representations of similar historical click news and obtain the user's first interest feature representation.

[0099] Combining this multi-scale feature map, the core activation unit is used to learn the user's second interest feature representation from the candidate news;

[0100] The probability of clicking on the candidate news is predicted based on the user's first interest feature representation and the user's second interest feature representation.

[0101] Example 3

[0102] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the news recommendation method based on candidate-perceived fine-grained interactive information as described in the first embodiment are implemented.

[0103] Example 4

[0104] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the news recommendation method based on candidate-perceived fine-grained interactive information as described in Example 1 are implemented.

[0105] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0106] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0107] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0109] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0110] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A news recommendation method based on candidate-aware fine-grained interactive information, characterized in that: The steps include: Obtain historical click news data and candidate news data; Based on historical click-through news data, candidate news data, and the trained news recommendation model, the probability of clicking on a candidate news item is obtained, and the candidate news item with the highest probability is recommended to the user. The construction process of the news recommendation model is as follows: Extract multi-scale feature maps of historical click news data and candidate news data respectively; Based on this multi-scale feature map, a fine-grained interaction module was constructed. At each semantic level, the matching features between candidate news and each historical click news were captured at different granularity levels to distinguish the vector representations of similar historical click news and obtain the user's first interest feature representation. Combining this multi-scale feature map, the core activation unit is used to learn the user's second interest feature representation from the candidate news; The probability of clicking on the candidate news is predicted based on the user's first interest feature representation and the user's second interest feature representation; wherein, Based on the multi-scale feature map, a fine-grained interaction module is constructed. At each semantic level, the matching features between the candidate news and each historical clicked news are captured at different granularity levels, including: Get the Multi-scale feature maps of historical click news and multi-scale feature maps of candidate news; In the On the semantic level, based on The multi-scale feature maps of historical click news and the multi-scale feature maps of candidate news are used to obtain the corresponding semantic matching matrix; Based on the semantic matching matrix, it represents An interaction matrix of historical click news and candidate news; Combined with The interaction matrix of historical click news and candidate news is used to obtain matching features using 2D multi-channel convolutional CNN; The expression of the semantic matching matrix is: Where, represents the Kth historical news in the atrous convolutional layer l The specific granularity news features obtained at the semantic level, Indicates the candidate news in the hole convolution layer l The specific granularity news features obtained at the semantic level, Indicates the number of convolution filters in each layer, N dk represents the length of the k-th historical news, Indicates the length of the candidate news.

2. The news recommendation method based on candidate-perceived fine-grained interactive information according to claim 1, characterized in that: The step of extracting multi-scale feature maps of historical click news data and candidate news data specifically includes: Based on historical click news and candidate news headlines and the BERT model, we obtain the tokens representation of the hidden layer corresponding to each news item. The low-dimensional vector representation of the obtained news headline is fed into the dilated convolutional layer and the Att-Bi-LSTM layer respectively to obtain enhanced semantic representation and multi-granularity semantic representation; After fusing the corresponding enhanced semantic representation and multi-granularity semantic representation, we obtain the multi-scale feature maps of historical click news data and candidate news data.

3. The news recommendation method based on candidate-perceived fine-grained interactive information according to claim 1, characterized in that: The method of combining the multi-scale feature map and learning the user's second interest feature representation from the candidate news through the core activation unit includes: The multi-scale feature map of the candidate news is vectorized and concatenated with the multi-scale feature map of the historical click news and the difference between the two to obtain a concatenated matrix. The concatenated matrix is combined with a three-layer fully connected neural network to learn the relationship between each historical click news and the candidate news. The user's second interest feature representation is obtained through the full interaction between the multi-scale feature maps of the candidate news and the multi-scale feature maps of the historical click news.

4. The news recommendation method based on candidate-perceived fine-grained interactive information according to claim 1, characterized in that: When training the news recommendation model, negative sampling technology is applied to K+1 news items to obtain a loss function, and the click probability is predicted based on the loss function.

5. The news recommendation method based on candidate-perceived fine-grained interactive information according to claim 4, characterized in that: The loss function is obtained by applying negative sampling technology to the combined K+1 news, including: treating each historical click news as a positive sample; random sampling K The exposed news that has not been clicked by the user is taken as a negative sample; i The click probability score of the sample and K The negative sample j The click rate scores of the samples are normalized to the posterior click probability of the positive samples; the loss function is obtained based on the posterior click probability of the positive samples.

6. A news recommendation system based on candidate-aware fine-grained interactive information, characterized in that: A news recommendation method based on candidate-perceived fine-grained interactive information according to any one of claims 1 to 5 is adopted, comprising: A data acquisition module is configured to: acquire historical click news data and candidate news data; A news recommendation module is configured to: obtain the probability of clicking on a candidate news item based on historical click-through news data, candidate news data, and a trained news recommendation model, and recommend the candidate news item with the highest probability to the user; The construction process of the news recommendation model is as follows: Extract multi-scale feature maps of historical click news data and candidate news data respectively; Based on this multi-scale feature map, a fine-grained interaction module was constructed. At each semantic level, the matching features between candidate news and each historical click news were captured at different granularity levels to distinguish the vector representations of similar historical click news and obtain the user's first interest feature representation. Combining this multi-scale feature map, the core activation unit is used to learn the user's second interest feature representation from the candidate news; The probability of clicking on the candidate news is predicted based on the user's first interest feature representation and the user's second interest feature representation.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of a news recommendation method based on candidate-perceived fine-grained interactive information as described in any one of claims 1 to 5 are implemented.

8. 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 program, the steps of the news recommendation method based on candidate-perceived fine-grained interactive information are implemented as described in any one of claims 1 to 5.

Images