A news recommendation method and system integrating multi-granularity information

Abstract

This paper proposes a news recommendation method and system that integrates multi-granularity information, belonging to the field of personalized news recommendation. First, interactive attention is applied between candidate news and historical news browsed by the user to capture key semantic information in the news, and weighted aggregation is performed to obtain the key parts of the text. A feature granularity network is then constructed to extract multi-granularity representations of the news. A Bi-LSTM and representation attention are used to learn the representation of each piece of news. The multi-granularity user interest information after feature expression is concatenated with the original information to obtain a final user interest representation vector, which is then used to match candidate news and calculate click probability. This paper utilizes interactive attention networks and representation attention to fuse interaction-based and representation-based models. Furthermore, it fully considers the multiple granularities of information present in news text and models the representations of different levels of user interest, thereby improving the accuracy of news recommendations.

Description

Technical Field

[0001] The present invention belongs to the field of personalized news recommendation, and in particular relates to a news recommendation method and system integrating multi-granularity 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] In today's information explosion, personalized news recommendation technology has become a key technology for many digital news platforms. The rise of digital news platforms such as MSN News and Google News has provided users with a large amount of news. However, it is relatively difficult for users to select the news they are interested in from this large amount of news. The application of personalized news recommendation technology can not only alleviate the situation of information overload, but also enhance the user's news reading experience, and it is particularly important to improve user stickiness.

[0004] Accurately matching user interests with candidate news is a prerequisite for personalized news recommendation. Many existing methods are based on representation matching strategies. They typically perform word embedding on each piece of news to obtain a news representation. They then learn a comprehensive interest representation for each user by aggregating historical news that the user has browsed. Finally, they match this comprehensive vector representation with the candidate news vectors to perform recommendations, achieving good results.

[0005] However, these methods also have shortcomings. They do not consider candidate news in the user modeling process. Since the learned user interests are matched with candidate news in the last step, this may weaken the fine-grained matching signal (word-level relationship) between the user and the candidate news. Recent research has addressed this issue by modeling recommendation methods based on fine-grained interactive matching between candidate news and each news article viewed by the user. Interaction-based news recommendation methods can better grasp the semantic focus of news and focus on word-level behavioral interactions, but may not be able to abstract high-level user interest representations from the news viewed by the user, and cannot grasp the user's interest bias.

[0006] Therefore, it is a problem worth solving to make full use of the above two architectures in an effective way so that the model can not only discover rich detailed clues of user interests from fine-grained behavioral interactions, but also abstract high-level user interest representations from users. Summary of the Invention

[0007] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a news recommendation method and system that integrates multi-granularity information. It uses interactive attention networks and representation attention to integrate interaction-based and representation-based models, fully considers the multiple granularity information existing in news texts, models the representation of users' interests at different levels, and improves the accuracy of news recommendations.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect of the present invention provides a news recommendation method integrating multi-granularity information;

[0010] A news recommendation method integrating multi-granularity information, comprising:

[0011] Obtain candidate news and user-hit news, perform news encoding on each, and obtain candidate news representation and user-hit news representation; perform weighted aggregation on the candidate news representation to obtain a weighted candidate news representation;

[0012] Based on the weighted candidate news representation and the user's historical click news representation, different convolution windows are used to group the news representations and extract multi-granularity features respectively;

[0013] The multi-granular features of candidate news and historical click news are input into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation;

[0014] Based on the vector representation of users' historical clicked news, dilated convolutions with different expansion rates are performed in a hierarchical manner to obtain multi-granularity representation of user interests;

[0015] Combined with multi-granularity user interest representation and candidate news vector representation, the probability of users browsing each candidate news is predicted, and candidate news is recommended to users based on the predicted probability.

[0016] Furthermore, the specific steps of the news coding are:

[0017] Get the headlines of news;

[0018] Use pre-trained Glove embedding to embed the title;

[0019] The words in the news headlines are encoded through a word embedding lookup table to obtain a news representation in the form of word vectors.

[0020] Furthermore, the specific steps of the weighted aggregation are:

[0021] Generate a similarity matrix between candidate news representations and user historical clicked news representations;

[0022] Based on the similarity matrix, calculate the attention weight of each word in the candidate news representation;

[0023] Perform attention-weighted aggregation on the candidate news representations to obtain weighted candidate news representations.

[0024] Furthermore, the news representation is grouped and convolved using different convolution windows. Specifically, the following steps are performed:

[0025] Perform information extraction and dimensionality reduction on news representation;

[0026] Perform granular information extraction and dimension expansion;

[0027] All the extracted granularity information is combined to obtain multi-granularity features.

[0028] Furthermore, through the hierarchical interest tree RTCN, multi-granularity user interest representation is obtained. The specific steps are as follows:

[0029] Extract vector representations of user click history news at different levels;

[0030] Concatenate the vector representations of user click history news at different levels;

[0031] Apply attention mechanism to the concatenated vector.

[0032] Furthermore, the probability of the user browsing each candidate news is predicted, and a matching score between the candidate news and the user is calculated.

[0033] Furthermore, the matching score between the candidate news and the user is the inner product of the multi-granularity user interest representation and the candidate news vector representation.

[0034] A second aspect of the present invention provides a news recommendation system that integrates multi-granularity information.

[0035] A news recommendation system integrating multi-granularity information, comprising a news encoding module, a multi-granularity extraction module, a news vector construction module, an interest layer extraction module and a probability prediction module;

[0036] The news encoding module is configured to: obtain candidate news and user-hit news, perform news encoding on each of them to obtain candidate news representation and user-hit news representation; perform weighted aggregation on the candidate news representation to obtain a weighted candidate news representation;

[0037] The multi-granularity extraction module is configured to: group and convolve the news representations using different convolution windows based on the weighted candidate news representations and the user's historical click news representations, and extract multi-granularity features respectively;

[0038] The news vector construction module is configured to: input the multi-granular features of candidate news and historical click news into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation;

[0039] The interest layer extraction module is configured to perform dilated convolution with different expansion rates in a hierarchical manner based on the vector representation of the user's historical click news to obtain multi-granular user interest representation;

[0040] The probability prediction module is configured to: combine multi-granular user interest representation and candidate news vector representation, predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

[0041] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a news recommendation method for integrating multi-granularity information as described in the first aspect of the present invention.

[0042] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the news recommendation method integrating multi-granularity information as described in the first aspect of the present invention are implemented.

[0043] One or more of the above technical solutions have the following beneficial effects:

[0044] In view of the fact that the existing news recommendation model based on representation will lose fine-grained matching signals, and the interaction-based method will not be able to grasp the global semantic information of the news, this paper proposes a news recommendation method that integrates multi-granularity information. It not only uses the interactive attention network and representation attention to integrate the interaction-based and representation-based models, but also fully considers the multiple granularity information in the news text and models the representation of users' interests at different levels to improve the accuracy of news recommendation.

[0045] The present invention first constructs a weighted news representation through interactive attention, and then extracts multiple granular information from the news text by building a granular network for feature extraction; uses Bi-LSTM to learn the representation of each news, and introduces an attention mechanism in the aggregation process to characterize the importance of each word to the news representation; designs an RTCN module to construct a hierarchical interest tree of users based on their historical news browsing, with different levels corresponding to user interests of different granularities; finally, the historical browsing news representations of users at different levels are spliced ​​and aggregated into the final multi-granular user interests to match candidate news and calculate the click probability, and a comprehensive empirical evaluation is carried out. Experiments on the real MIND news dataset verify the effectiveness of the scheme.

[0046] 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

[0047] 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.

[0048] Figure 1 Examples of news reading behavior for two users.

[0049] Figure 2 This is a flow chart of the method of the first embodiment.

[0050] Figure 3 This is a system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0051] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present disclosure. 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 disclosure belongs.

[0052] 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 disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0053] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0054] In fact, word-level interactions between clicked news of the same user contain rich detailed clues for understanding user interests. However, most existing methods model user interests only in the last step and match them with candidate news without interacting with the candidate news before that, thus ignoring word-level behavioral interactions, which may lead to suboptimal user interest modeling.

[0055] Figure 1 The following is an example of two users’ news reading behaviors. d1-d3 are the three historical news articles clicked by the users, and c is the news article that was actually clicked later.

[0056] Based on user 1's reading history, we can see that the candidate news item "Colin Kaepernick" matches the word "Justin Reid" in the news item that user 1 first clicked (a fine-grained interaction occurred), which prompted the user's current click. Furthermore, we can infer from the terms "NFL Game" and "Dallas Cowboys" that the user is an NFL fan, which is important because they refer to the game and the team. Unfortunately, the user's aggregate vector obtained by adopting a representation-based matching strategy mixes all terms in d1, d2, and d3, and also includes noise such as Christmas and travel, which are irrelevant to the current click. This blurs these fine-grained interests and reduces the ability to model user interests. To capture the fine-grained matching signals between candidate news items and users, researchers have proposed interactive matching strategies. For example, FIM extracts multi-level representations for each news item and performs fine-grained matching through convolution, while AMM utilizes complementary information from different fields (such as title, summary, and body) to obtain multi-field matching representations.

[0057] Interaction-based news recommendation methods have greatly explored semantic clues related to user interests. However, these methods ignore the relationships between different clicked news from the same user, and the relationships between these different clicked news contain rich detailed clues to infer user interests to some extent.

[0058] For example, based on the analysis of user 2's reading history, a comprehensive analysis of "dementia," "anxiety symptom," and "Nutritionally" in d1, d2, and d3 indicates that the user is browsing health-related news. However, an interaction-based matching strategy cannot encourage user 2 to click on candidate news. On the one hand, the keyword Weight-Loss in the candidate news is far away from dementia, anxiety symptom, and Nutritionally in the historical news in semantic space. On the other hand, there is no interaction between different clicked news by the same user, making it impossible to abstract the user's interest information at a high level.

[0059] The two different matching strategies mentioned above are not simply inclusive, but complementary. Therefore, the problem to be solved is to effectively enable the model to achieve both fine-grained word-level behavioral interaction between candidate news and user-browsed news, and high-level abstract expression of user interests.

[0060] To solve this problem, the present invention provides a news recommendation method that integrates multi-granularity information and proposes a model MnRec that integrates multi-granularity information for news recommendation. In the model, the two matching methods mentioned above are integrated using interactive attention and representation attention to fully utilize the advantages of the two frameworks and compensate for the problems exposed by a single matching method. In addition, in order to better realize news modeling, a granular network module is designed to realize multi-granularity information extraction of news. Finally, a multi-level interest modeling framework for users is proposed to enhance user interest representation.

[0061] Example 1

[0062] like Figure 2 As shown, this embodiment discloses a news recommendation method that integrates multi-granularity information, which includes the following main steps: news encoding, multi-granularity extraction, news vector construction, interest layer extraction, and calculation of the probability of users clicking on candidate news.

[0063] Step 1: Obtain candidate news and user-hit news, encode them separately to obtain candidate news representation and user-hit news representation; perform weighted aggregation on the candidate news representation to obtain a weighted candidate news representation;

[0064] Step 101: Obtain the user's clicked news and candidate news titles, encode the news titles, and obtain a news title word vector sequence as the news representation. The specific steps of news encoding are as follows:

[0065] (1) Using pre-trained Glove embedding, candidate news title word sequence D c =[w1,w2,…,w T ], the word sequence of the kth news title clicked by the user in history is D k =[w1,w2,…,w J ]. Where T represents the number of words in the candidate news title, and J represents the number of words in the k-th news title that the user has clicked on historically.

[0066] (2) Through a word embedding lookup table Convert the news headline word sequence into a word vector sequence, where V and D are the vocabulary size and word embedding dimension respectively.

[0067] The candidate news mapped to the low-dimensional vector space is represented as c = [e1, e2, ..., e T ], the user’s historical click news is represented by d=[e1,e2,…,e J ]. The kth news clicked by the user in history is represented as d k .

[0068] Step 102: Perform weighted aggregation on the candidate news representations to obtain weighted candidate news representations;

[0069] The specific steps of weighted aggregation are:

[0070] (1) Generate the similarity matrix M between the candidate news representation and the user's historical click news representation c,k =Linear(c·d k + bias), where Linear is a linear layer and bias is a trainable parameter. represents the similarity between the t-th word in the candidate news and the j-th word in the k-th historical news, T and J represent news c and d respectively k The word length of .

[0071] (2) Based on the similarity matrix, calculate the attention weight of each word in the candidate news representation;

[0072] Let a t ∈R J represents the attention weight of the t-th word of the candidate news to each word of the user's historical clicked news, ∑a tj =1, the attention weight is calculated as

[0073] (3) Perform attention-weighted aggregation on the candidate news representations to obtain weighted candidate news representations;

[0074] The weighted candidate news is expressed as therefore, It is a candidate news vector that includes historical news information.

[0075] Since the user browses more than one historical news, each candidate news participating in the calculation will get the corresponding k weighted representations of news. The attention mechanism is used to aggregate k weighted news items to get the news representation c agg .

[0076] In order to enrich the semantic relationship of news, the aggregated vector c agg Combined with the original vector c, a weighted candidate news representation c containing the user's news browsing history information is obtained init , so that it is easier to capture important key information in news headlines during information extraction and news representation.

[0077] Step 2: Based on the weighted candidate news representations and the user's historical click news representations, different convolution windows are used to group the news representations and convolve them to extract multi-granular features.

[0078] The ResNeXt network used for feature extraction can not only solve the degradation problem of deep neural networks, but also promote convergence. Therefore, the present invention introduces residual networks into recommendation tasks to realize multi-granularity feature extraction of text information. The triple information in each layer of the network represents the input feature dimension, convolution kernel size and output feature dimension. The MnRec model uses one-dimensional convolution with different convolution windows instead of two-dimensional convolution in the residual network to realize multi-granularity feature extraction of news text. The present invention calls this feature extraction network a granular network Granet. The network uses three different convolution windows to perform group convolution on the original text. Each group of convolution windows extracts feature representations of different granularity to capture more text feature information. The specific formula of the multi-granular network is as follows:

[0079]

[0080]

[0081]

[0082] Where, i,j∈{1,2,3};

[0083] The first layer: extract information and reduce dimension of news representation. Represents the tensor representation of the original text in the first layer of the Granet network after information extraction and dimensionality reduction;

[0084] The second layer: granular information extraction and dimension expansion, represents the result of granular information extraction and dimension expansion in the second layer; W mul represents the granularity sliding window. As the window slides, the granularity information in the text sequence is gradually extracted; σ represents the ReLU activation function.

[0085] The third layer: The model combines text information of all granularities to obtain a multi-granularity news representation tensor.

[0086] Finally, the residual splicing operation in ResNeXt is used to combine the multi-granularity information c contained in the feature-extracted candidate news fin With the original information c init Splicing to obtain the multi-granularity feature vector representation c of the candidate news mult The same operation is performed on the user's historical click news representation to obtain the multi-granularity feature vector representation d of the historical click news. mult .

[0087] Step 3: Input the multi-granular features of candidate news and historical click news into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation containing deep semantic information;

[0088] Because important news information can appear anywhere in a headline, this paper uses an attention-based bidirectional long short-term memory (Att-Bi-LSTM) network to capture the most important semantic information in a sentence. For news matching in the prediction phase, the importance of phrases extracted by the granular network varies, and low-importance phrases may be noise. Therefore, the MnRec model incorporates a representation attention mechanism after the bi-LSTM, which performs weighted aggregation on each word in the news to obtain a news vector representation. The formula for calculating the candidate news vector representation is as follows:

[0089] c bl =Bi-LSTM(c mult )

[0090] α=softmax(w T tanh(c bl ))

[0091] c rep =c bl α T

[0092] Similarly, for d mult Perform the same operation to get the user's historical click news vector representation d rep .

[0093] Step 4: Based on the vector representation of the user's historical clicked news, dilated convolutions with different expansion rates are performed in a hierarchical manner to obtain user interest representations with different granularity sizes;

[0094] User interests are multi-level and multi-granular. Hierarchical modeling of user interests helps to improve the final recommendation prediction. Therefore, the present invention constructs a hierarchical interest tree RTCN to capture user interests at multiple granularities in a hierarchical manner.

[0095] RTCN uses a three-layer structure and applies dilated convolution with different dilation rates to the news sequence browsed by users. Due to the different dilation rates, the news features extracted by each group of convolution windows have different granularity. This method enables the model to capture user interests at different granularities. 0 Represents the news collection of the user's historical clicks, where r 0 =[d rep1 ,…,d repN ](N is the length of the news sequence that the user has clicked on historically), d repN Represents the Nth news item clicked by the user in history. The calculation method of RTCN hierarchical interest tree is as follows:

[0096]

[0097] Where δ is the expansion rate, F w represents the convolution kernel, 2w+1 is its window size, and b is a learnable parameter. ReLU is the activation function. The input of each convolutional layer is the output of the previous layer.

[0098] Therefore, through the hierarchical interest tree RTCN, we can obtain multi-granularity user news vector representation. The specific steps are as follows:

[0099] The user news click vector set obtained at each layer is represented as r l =[d l rep1 ,…,d l repN ], where l represents the number of layers.

[0100] In order to enrich user interest information, the user news vector sets at different levels are concatenated to obtain r = [r 0 ,…,r l ];

[0101] The attention mechanism is applied to the concatenated vector to obtain the final multi-granularity user interest representation u. The formula is as follows:

[0102]

[0103]

[0104] Where W and b are parameters, and q represents the attention query vector.

[0105] Step 5: Combine the multi-granular user interest representation and the candidate news vector representation to predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

[0106] The probability of a user browsing a candidate news item is predicted based on its matching score. The matching score of a candidate news item is first calculated by the inner product of the news item and the user's representation vector using the following formula:

[0107]

[0108] A common observation in news recommendation is that most users usually click only a few news items displayed in the exposure sequence. Therefore, the number of positive and negative news samples (news that users did not click in the same exposure sequence) is highly unbalanced. During model training, negative sampling is applied by jointly predicting the click-through scores of k+1 news items. The k+1 news items consist of a positive sample of one user and a randomly selected negative sample of one user, and then jointly predict the positive news items. and K pieces of negative news The news click prediction problem is formulated as a pseudo K+1-way classification task; these click probability scores are normalized using softmax to calculate the backward click probability of the positive sample as follows:

[0109]

[0110] Among them, y + represents the click probability score of the i-th positive news, y j - represents the click probability score of the jth negative news in the same period as the ith positive news; the loss function L in the model training method is the negative log-likelihood of all positive samples, as follows:

[0111]

[0112] Where s is the set of positive training samples.

[0113] Example 2

[0114] This embodiment discloses a news recommendation system that integrates multi-granularity information;

[0115] like Figure 3 As shown, a news recommendation system integrating multi-granularity information includes a news encoding module, a multi-granularity extraction module, a news vector construction module, an interest layer extraction module and a probability prediction module;

[0116] The news encoding module is configured to: obtain candidate news and user-hit news, perform news encoding on each of them to obtain candidate news representation and user-hit news representation; perform weighted aggregation on the candidate news representation to obtain a weighted candidate news representation;

[0117] The multi-granularity extraction module is configured to: group and convolve the news representations using different convolution windows based on the weighted candidate news representations and the user's historical click news representations, and extract multi-granularity features respectively;

[0118] The news vector construction module is configured to: input the multi-granular features of candidate news and historical click news into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation;

[0119] The interest layer extraction module is configured to perform dilated convolution with different expansion rates in a hierarchical manner based on the vector representation of the user's historical click news to obtain multi-granular user interest representation;

[0120] The probability prediction module is configured to: combine multi-granular user interest representation and candidate news vector representation, predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

[0121] Example 3

[0122] The purpose of this embodiment is to provide a computer-readable storage medium.

[0123] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a news recommendation method for integrating multi-granularity information as described in Example 1 of the present disclosure.

[0124] Example 4

[0125] The purpose of this embodiment is to provide an electronic device.

[0126] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of a news recommendation method integrating multi-granularity information as described in Example 1 of the present disclosure are implemented.

[0127] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure 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, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, 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 flowchart and / or block diagram. 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.

[0129] 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.

[0130] 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 A step that specifies a function in one or more boxes.

[0131] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

[0132] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A news recommendation method integrating multi-granularity information, characterized in that: include: Obtain candidate news and user-hit news, encode them separately, and obtain candidate news representation and user-hit news representation; Perform weighted aggregation on the candidate news representations to obtain weighted candidate news representations; Based on the weighted candidate news representation and the user's historical click news representation, different convolution windows are used to group the news representations and extract multi-granularity features respectively; The multi-granular features of candidate news and historical click news are input into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation; Based on the vector representation of users' historical clicked news, dilated convolutions with different expansion rates are performed in a hierarchical manner to obtain multi-granularity representation of user interests; Combine multi-granular user interest representation and candidate news vector representation to predict the probability of users browsing each candidate news, and recommend candidate news to users based on the predicted probability; Among them, through the hierarchical interest tree RTCN, multi-granularity user interest representation is obtained. The specific steps are: Extract vector representations of user click history news at different levels; Concatenate the vector representations of user click history news at different levels; Apply attention mechanism to the concatenated vectors; RTCN uses a three-layer structure and applies different dilation rates to the news sequence browsed by the user for dilated convolution. Due to the different dilation rates, the news feature representation granularity extracted by each group of convolution windows is different. r 0 Represents the news collection of the user's historical clicks, where , N is the length of the news sequence that the user has clicked on historically, d repN The Nth news that a user has clicked on in history is represented by the RTCN hierarchical interest tree. The calculation method is as follows: Where δ is the expansion rate, represents the convolution kernel, is its window size, b is a learnable parameter, ReLU is the activation function, and the input of each convolutional layer is the output of the previous layer.

2. The news recommendation method integrating multi-granularity information according to claim 1, characterized in that: The specific steps of the news coding are: Get the headlines of news; Use pre-trained Glove embedding to embed the title; The words in the news headlines are encoded through a word embedding lookup table to obtain a news representation in the form of word vectors.

3. The news recommendation method integrating multi-granularity information according to claim 1, characterized in that: The specific steps of weighted aggregation are: Generate a similarity matrix between candidate news representations and user historical clicked news representations; Based on the similarity matrix, calculate the attention weight of each word in the candidate news representation; Perform attention-weighted aggregation on the candidate news representations to obtain weighted candidate news representations.

4. The news recommendation method integrating multi-granularity information according to claim 1, characterized in that: The news representation is grouped and convolved using different convolution windows. The specific steps are: Perform information extraction and dimensionality reduction on news representation; Perform granular information extraction and dimension expansion; All the extracted granularity information is combined to obtain multi-granularity features.

5. The news recommendation method integrating multi-granularity information according to claim 1, characterized in that: The method predicts the probability of the user browsing each candidate news and calculates the matching score between the candidate news and the user.

6. The news recommendation method integrating multi-granularity information according to claim 5, characterized in that: The matching score between the candidate news and the user is the inner product of the multi-granularity user interest representation and the candidate news vector representation.

7. A news recommendation system integrating multi-granularity information, characterized in that: A news recommendation method integrating multi-granularity information according to any one of claims 1 to 6 is adopted, comprising a news encoding module, a multi-granularity extraction module, a news vector construction module, an interest layer extraction module, and a probability prediction module; The news encoding module is configured to: obtain candidate news and user-hit news, perform news encoding on each of them to obtain candidate news representation and user-hit news representation; perform weighted aggregation on the candidate news representation to obtain a weighted candidate news representation; The multi-granularity extraction module is configured to: group and convolve the news representations using different convolution windows based on the weighted candidate news representations and the user's historical click news representations, and extract multi-granularity features respectively; The news vector construction module is configured to: input the multi-granular features of candidate news and historical click news into the attention-based bidirectional long short-term memory network to obtain the candidate news vector representation and the user's historical click news vector representation; The interest layer extraction module is configured to perform dilated convolution with different expansion rates in a hierarchical manner based on the vector representation of the user's historical click news to obtain multi-granular user interest representation; The probability prediction module is configured to: combine multi-granular user interest representation and candidate news vector representation, predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

8. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the news recommendation method integrating multi-granularity information as described in any one of claims 1 to 6 are implemented.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the news recommendation method integrating multi-granularity information are implemented as described in any one of claims 1 to 6.

Images