An interest activation news recommendation method and system based on multi-level matching

Abstract

The present invention provides an interest activation news recommendation method and system based on multi-level matching, including: obtaining candidate news and user historical click news, respectively performing news encoding to obtain candidate news representations and user historical click news representations; obtaining user global interest features based on the user historical click news representations; obtaining user dynamic interest features based on the candidate news representations and the user historical click news representations; obtaining local interest matching ranking features at different similarities between each candidate news and the user historical click news based on the candidate news representations and the user historical click news representations; jointly using the user global interest features, dynamic interest features and local interest matching ranking features to predict the probability of the user browsing each candidate news, and recommending the candidate news to the user based on the predicted probability. It provides an improved trade-off between accuracy and speed, enriches the user's interest representation, and enhances the robustness and interpretability of predicting the probability of the user browsing each candidate news.

Description

Technical Field

[0001] The present invention belongs to the technical field of personalized news recommendation, and particularly relates to a news recommendation method and system for interest activation based on multi-level matching. Background Art

[0002] The statements in this part only provide background technical information related to the present invention, and do not necessarily constitute prior art.

[0003] As an important branch in the research field of recommendation systems, news recommendation aims to help users find news that matches their interest preferences as much as possible through news content and user information. Personalized news recommendation is crucial for helping users find news they are interested in and alleviating information overload. With the great progress of deep learning based on artificial neural networks, people have begun to try to use classical neural network architectures to handle news recommendation problems. They usually perform word embedding on each news to obtain the representation of the news, then learn the comprehensive interest representation of each user by integrating the historical news browsed by the user. Finally, they perform recommendation by matching the comprehensive vector representation with the candidate news vector.

[0004] Although the deep learning-based method has achieved certain success in news recommendation, there are still some problems. Since the important semantic features of news are hidden in text fragments of different granularities, however, in some existing deep learning methods, the learned fixed user feature representation can only be matched with candidate news in the last step, lacking the interaction between text pairs of words, and the resulting information loss is difficult to measure. In addition, the fixed vector cannot fully express the different interests of users. Summary of the Invention

[0005] To solve the technical problems existing in the above background art, the present invention provides a news recommendation method and system for interest activation based on multi-level matching. Considering the candidate news, the dynamic feature vector of the user interest is deduced according to the difference of the candidate news, further strengthening the user's interest representation and enhancing the robustness and interpretability of predicting the probability of the user browsing each candidate news.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The first aspect of the present invention provides a news recommendation method for interest activation based on multi-level matching, which includes:

[0008] Obtain candidate news and user historical click news, and perform news encoding respectively to obtain candidate news representation and user historical click news representation;

[0009] Based on the user historical click news representation, obtain the user global interest feature;

[0010] Based on the candidate news representation and the representation of the user's historical clicked news, an object attention mechanism is adopted to calculate the attention scores between each user's historical clicked news and each candidate news, and the attention scores are used to weight and calculate the user's historical clicked news to obtain the user's dynamic interest features;

[0011] Based on the candidate news representation and the representation of the user's historical clicked news, capture the semantic relevance between the clicked news and the candidate news at multiple different similarity levels, and obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different levels;

[0012] Combine the user's global interest features, dynamic interest features and local interest matching ranking features to predict the probability that the user browses each candidate news, and recommend candidate news to the user based on the predicted probability.

[0013] Furthermore, the specific steps of the news encoding are as follows:

[0014] Obtain the title of the news;

[0015] Use the news title to construct a vector representation to obtain the word vector sequence of the news title;

[0016] Input the news title word vector sequence into a convolutional neural network to obtain a sequence of context word representation vectors;

[0017] Input the sequence of context word representation vectors into a word-level attention network to obtain the final representation of the news.

[0018] Furthermore, the steps for obtaining the local interest matching ranking features are as follows:

[0019] Calculate the association matrix based on the candidate news representation and the representation of the user's historical clicked news;

[0020] Pool the association matrix using a Gaussian kernel function to convert the word interactions in the association matrix into the preliminary matching ranking features of the candidate news;

[0021] Based on the preliminary matching ranking features of the candidate news, apply a convolutional neural network and a max pooling function to obtain the local interest matching ranking features of each candidate news and the user's historical clicked news.

[0022] Furthermore, the specific steps for predicting the probability that the user browses each candidate news are as follows:

[0023] Based on the user's global interest features, dynamic interest features and candidate news representation, calculate the matching score based on the user's global interest and the matching score based on the dynamic interest;

[0024] Based on the local interest matching ranking features, calculate the local interest matching score;

[0025] Combine the matching scores based on the user's global interests, the matching scores based on dynamic interests, and the local interest matching scores to calculate the probability of the user browsing each candidate news.

[0026] Furthermore, the matching score based on the user's global interests is the inner product of the user's global interest features and the candidate news representation.

[0027] Furthermore, the matching score based on dynamic interests is the inner product of the dynamic interest features and the candidate news representation.

[0028] Furthermore, the user's global interest features are the weighted sum of the user's historical clicked news representations.

[0029] The second aspect of the present invention provides an interest-activated news recommendation system based on multi-level matching, which includes:

[0030] A news encoding module, which is configured to: obtain candidate news and the user's historical clicked news, perform news encoding respectively, and obtain the candidate news representation and the user's historical clicked news representation;

[0031] A user encoding module, which is configured to: obtain the user's global interest features based on the user's historical clicked news representation;

[0032] A target-aware interest activation module, which is configured to: based on the candidate news representation and the user's historical clicked news representation, use the target attention mechanism to calculate the attention scores between each user's historical clicked news and each candidate news, and use the attention scores to perform weighted calculation on the user's historical clicked news to obtain the user's dynamic interest features;

[0033] A multi-level matching interaction module, which is configured to: based on the candidate news representation and the user's historical clicked news representation, capture the semantic relevance between the clicked news and the candidate news at multiple different similarity levels, and obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different levels;

[0034] A probability prediction module, which is configured to: combine the user's global interest features, dynamic interest features, and local interest matching ranking features to predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

[0035] The third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in an interest-activated news recommendation method based on multi-level matching as described above.

[0036] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in a news recommendation method based on multi-level matching for interest activation as described above are implemented.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] The present invention provides a news recommendation method based on multi-level matching for interest activation, which retains the fine-grained information of the text through multi-level matching interaction. The text word pairs are modeled at multiple different similarity levels, making the encoded user interest information richer.

[0039] The present invention provides a news recommendation method based on multi-level matching for interest activation, which proposes an attention network mechanism based on candidate news, can adaptively activate the interests of different users in different candidate news, deduce the dynamic feature vectors of user interests according to different candidate news, further strengthen the user interest representation, and greatly improve the expressiveness of the model.

[0040] The present invention provides a news recommendation method based on multi-level matching for interest activation, which effectively assigns the weights of the scores of each module, effectively fuses the matching scores based on the global user interests, the matching scores based on dynamic interests, and the interest matching scores at different similarity levels, enabling the model to better fuse, and thus providing an improved trade-off between accuracy and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0042] Figure 1 is a flowchart of a news recommendation method based on multi-level matching for interest activation according to Embodiment 1 of the present invention;

[0043] Figure 2 is an example diagram of a user's news reading behavior according to Embodiment 1 of the present invention;

[0044] Figure 3 is a schematic diagram of global user interest encoding according to Embodiment 1 of the present invention;

[0045] Figure 4 is a schematic diagram of target-aware dynamic interest activation according to Embodiment 1 of the present invention;

[0046] Figure 5 is a diagram showing the change of the Gaussian radial basis function kernel according to Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0049] 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 also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] Embodiment 1

[0051] As Figure 1 shown, this embodiment provides an interest-activated news recommendation method based on multi-level matching, and proposes an interest-activated news recommendation model MIAR based on multi-level matching. By inputting the obtained candidate news and the user's historical clicked news into this model, the probability of the user browsing each candidate news can be obtained to solve the news recommendation click prediction problem. Different from previous work that models the user's interest as a single fixed vector to match candidate news, the present invention realizes that the user's interest is usually diverse and multi-granular, and a static vector cannot fully express the user's interest; in addition, the fused user interest vector will lose the matching information of text details; therefore, through the target-aware attention network mechanism, considering the candidate news, the dynamic feature vector of the user's interest is deduced according to the differences of the candidate news to further enrich the user's interest representation; in addition, through the multi-level matching interaction module, the text word pairs are modeled at multiple different similarity levels to obtain the interest matching information at different similarity levels; then, the high-order significant information is identified through a convolutional neural network; finally, the weights of the scores of different modules are effectively allocated to enable the model to better fuse, thereby providing an improved trade-off between accuracy and speed. The present invention has conducted extensive experiments on the real news dataset of MIND. The results show that compared with the baseline model, the MIAR model has achieved substantial improvement in performance indicators, verifying the effectiveness of the model in news recommendation.

[0052] As Figure 2As shown, it is an example of a user's news reading behavior. The user has various interests, including weight loss, pets, electrical appliance safety, and sports issues. Among them, the text fragments shown in bold are the key semantic clues. Different historical browsing news can reveal the user's interests in different topics or events. The themes of historical news D1 and D2 are related to "pet dog" and "weight loss" respectively. Naturally, they provide key clues to select candidate news C1 and C2. However, the granularity of the text matching fragments between them is different, such as "Dog" - "Golden Retriever", "Lose Weight" - "Lose 100 Pounds"; mining information related to the user's interests is crucial for improving the diversity characteristics of the user's interests. From the perspective of relevance matching, the historical clicked news D3 should match the candidate news C3, and they are highly relevant, both reflecting the safe use of devices in winter. If the different historical news clicked by the user is aggregated into a user vector for matching, this method may cause the loss of soft matching signals between text word pairs. That is, the matching information of text word pairs with different levels of similarity is lost. Such as the matching between "hand warmers" - "The little sun". In addition, the historical news clicked by the user is usually very rich, and the user's interests are diverse. If a fixed-size embedding vector is used to represent the user's comprehensive interest characteristics, although the fixed-size vector represents all the interests of a single user, it cannot dynamically express the user's different interests. This will limit the expressive ability of the recommendation model to a certain extent. If the dimension of this fixed-length vector is expanded to improve the interest characteristics, it may cause the risk of overfitting and reduce the performance of the model. However, the present invention observes that it is not necessary to embed all the user's interests into a vector. For example, there are news related to the following semantic clues (Messi, pet dog, weight loss) in the historical news read by the user. If the clue of golden retriever appears in the news recommended to the user, the present invention will focus on the user's interest in pet dogs rather than weight loss. Therefore, the user's different interests can be specifically activated when candidate items are given.

[0053] As Figure 1 As described above, an interest-activated news recommendation method based on multi-level matching of the present invention mainly consists of two parts: a local representation matching part (left) and a distributed matching part (right). The two parts included in this method include the following main steps: news encoding, user encoding, target-aware interest activation, multi-level matching interaction, and prediction of calculating the probability of the user clicking on the candidate news.

[0054] Step 1, as Figure 1 shown in the news encoding module, obtain the candidate news and the user's clicked news, and perform news encoding.

[0055] Step 101, Title Encoding: Obtain the titles of the news clicked by the user and the candidate news, and use the news titles to construct a vector representation to obtain the news title word vector sequence.

[0056] (1) Use the pre-trained Glove embedding to represent the news title with M words to obtain the news title word sequence D i =[w1, w2, …, w M , where M represents the number of words in the news title;

[0057] (2) Through a word embedding lookup table Convert the news title word sequence into a word vector sequence E w =[e1, e2, …, e M , where V and D are the vocabulary size and the word embedding dimension respectively.

[0058] Step 102, Input the news title word vector sequence into the Convolutional Neural Network (CNN) to obtain a sequence of context word representation vectors.

[0059] The present invention uses CNN to apply to the news title word vector sequence to obtain the corresponding embedding, and uses the max-pooling technique to retain the most significant features. CNN is an effective neural architecture for capturing local information, and the context in the news is very important for understanding the news semantics. For example, "A man spent $2,100 on an apple, but it was identified as a fake". The combination of the two words "apple" and "fake" is very important for understanding that the apple here is an iPhone rather than a fruit. Therefore, the present invention uses CNN to learn the context word representation by capturing the local context of the words. The i-th word representation after applying CNN is c i , and its calculation formula is:

[0060] c i = ReLU(F w × e (i-K):(i+K) + b w ) (1)

[0061] where e(i - K):(i + K) is the concatenation of the word embeddings from position i - k to i + k, and represent the parameters of the CNN filter, N f is the number of CNN filters, 2k + 1 is the window size of the filter, ReLU is used as the activation non-linear function, and the output of the CNN layer is a sequence of context word representation vectors, denoted as [c1, c2, …, c M .

[0062] Step 103: Input the sequence of context word representation vectors into the word-level attention network to obtain the final representation of the news.

[0063] Different users will have different interests, and at the same time, each user often has multiple interests. Therefore, different users may click on this news due to different aspects of the news. Different words in a news title often have different impacts on users. Therefore, learning the attention of users to each word helps to strengthen the representation of the news. Through the personalized attention mechanism of the present invention, the embedding weight α of each word in the news title is obtained i , and these weights can reflect the attention of users to each word. Then, all words are weighted and averaged to obtain the news title, that is, the final representation of the news.

[0064]

[0065]

[0066] Among them, W p and b p are projection parameters, and q w uses the embedding of the user identifier as the word attention query vector.

[0067] Apply the news encoder to all the clicked news of the user to obtain the representation [r1, r2,..., r N of the user's clicked news, where N is the number of clicked news.

[0068] Apply the news encoder to the candidate news to obtain the representation [r′0, r′1,..., r′ K of the user's candidate news, where K + 1 is the number of candidate news.

[0069] Step 2: User encoding: Based on the representation of the user's historical clicked news, use the personalized attention mechanism to obtain the global interest feature of the user.

[0070] As Figure 3 shown, user encoding is to mine the interest representation of the user from the news browsed by the user. However, not all the news browsed by a user can reflect his preferences, and the representation of the same news for different users is also different. Based on this, the personalized attention mechanism is also applied to the historical news clicked by the user.

[0071]

[0072]

[0073] Among them, q d 、b d 、W dis a parameter in the news attention network, α′ i is the attention weight representation of the i-th news.

[0074] The global interest feature of the user is S u , and the global interest feature of the user is the weighted sum of the user's historical clicked news representations, weighted by their attention weights, i.e.,

[0075]

[0076] Step 3, Target-aware Interest Activation: Based on the representations of the candidate news and the representations of the user's historical clicked news, use the target attention mechanism to calculate the attention scores between each user's historical clicked news and each candidate news, and use the attention scores to perform weighted calculation on the user's historical clicked news to obtain the user's dynamic interest feature.

[0077] Compressing the historical news browsed by the user into a fixed representation vector will, to a certain extent, limit the user's interest feature. Simply put, the user's interest is usually multi-granular, and a fixed-size vector representing all the interests of a single user cannot dynamically express different user interests. In addition, previous work only uses the historical sequence of user clicks to capture the user's interest and does not consider the connection with the candidate news. Accordingly, the present invention designs an attention mechanism based on candidate news, which can adaptively activate the interests of different users for different candidate news, deduce the dynamic feature vector of the user's interest according to different candidate news, further strengthen the user's interest representation, and greatly improve the expressiveness of the model.

[0078] As Figure 4 shown, when the embedding vector of each candidate news is obtained, the candidate target embedding is started to be constructed to adaptively consider the user's interest related to the candidate news. Here, the target embedding is defined as all the candidate news to be predicted. Usually, the given recommended candidate news only matches a part of the user's interests. To simulate this process, a new target attention mechanism is designed to calculate the soft attention scores of the user's historical clicked news with respect to each candidate news. Therefore, a target attention mechanism is introduced to calculate all the historical clicked news r i and each candidate news r′ t between the attention scores.

[0079] First, apply a weight matrix with shared parameters and a non-linear transformation to each historical click-candidate news pair;

[0080] Then, use the softmax function to normalize the self-attention scores:

[0081]

[0082] Finally, based on the user's historical clicked news sequence, the user's dynamic interest feature is represented by as follows:

[0083]

[0084] The obtained user dynamic interest feature varies with different candidate news embeddings.

[0085] Step 4: Multi-level matching interaction: Based on the candidate news representation and the user's historical clicked news representation, obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different similarity levels. Specifically, based on the candidate news representation and the user's historical clicked news representation, capture the semantic correlation between the clicked news and the candidate news at multiple different similarity levels, and obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different levels.

[0086] Due to the multi-granularity semantic information of the text and the diversity of user interests, single interest vector matching of candidate news will lose a large amount of semantic information contained in the text. When the user clicks on the news "Beware! The uneven quality of hand warmers poses a safety hazard", this news reflects the safe use of equipment in winter. If we integrate the historical news browsed by the user according to some existing methods to learn the comprehensive representation of the user, and then match this representation with the candidate news vector to calculate the prediction score, this will lose the detailed matching information. For example, the candidate news "The small sun is simple and portable, but the potential danger cannot be ignored" also reflects the safe use of equipment in winter, but gets a lower matching score. On the one hand, since the user features learned can only be matched with the candidate news in the last step, this results in a lack of information interaction between text word pairs and causes the loss of fine-grained matching information. On the other hand, matching word pairs like "small sun" and "hand warmer" do not contribute much to the final score. Their word similarities are at different levels. In the matching, only the accurately matched words will have a greater contribution to the result, which is not reasonable when dealing with news texts that are becoming more and more colloquial. The multi-level matching interaction of the present invention will solve this problem in the following three steps:

[0087] Step 401: Construct an association matrix: Based on the candidate news representation and the user's historical clicked news representation, calculate the association matrix.

[0088] Instead of performing attention-weighted aggregation on the word vectors embedded after CNN feature extraction, the present invention constructs an association matrix M, and each element in M is the similarity between title word embeddings, and the formula is as follows:

[0089]

[0090] where, is the i-th word embedding of the k-th candidate news title, is the j-th word embedding of the n-th historical click news title.

[0091] Step 402, Kernel function pooling and obtaining soft matching features: Pool the obtained association matrix M using the Gaussian kernel function, and convert the word interaction in the association matrix into the preliminary matching ranking feature φ(M) of the candidate news:

[0092] Obtain soft matching features:

[0093]

[0094] Kernel function pooling:

[0095]

[0096] Among them, Apply k kernels to each row corresponding to each word of the candidate news in the association matrix, and summarize them into a k-dimensional feature vector. The logarithmic sum of the feature vectors of each candidate news word forms the candidate news ranking feature vector φ.

[0097] The effect of [] depends on the use of the kernel function, and it works using the Gaussian radial basis function kernel (RBF)

[0098]

[0099] Among them, μ k is the kernel function center, is for M ij N,K and μ k is the squared Euclidean distance between [] and μ. As the distance between the two vectors increases, the Gaussian kernel function decreases monotonically. Because the value of the RBF kernel function decreases with the increase of the distance and is between 0 (limit) and 1 (when M ij K,N = μ k ). K k The closer it is to its average μ by calculating the similarity of word pairs k , the larger its value. Kernel pooling using the RBF kernel is a generalization of existing pooling techniques. As Figure 5 shown, it can be seen that when σ is relatively large, at this time the vector M ij K,N - μ k the change of the distance between [] has a smaller impact on the overall value of the exponent, and the curve is relatively smooth. When σ → ∞, the kernel pooling function degenerates into the mean pooling function. Similarly, when σ is relatively small, the vector M ij N,K - μ kIf the impact of the change in the distance between them on the overall value of the exponent becomes larger, then the Gaussian kernel function will be very sensitive to the distance between two points. That is, the kernel generated by μ = 1 and σ → 0 only responds to exact matches. Otherwise, different kernels focus on different similarity levels. For example, the kernel with μ = 0.8 calculates the number of document words in historical clicked news that are similar to the news candidate with a similarity close to 0.8. For example, the similarity matching between the news candidate "hot water bag" and words such as "hand warmer" and "small sun". That is, at different similarity levels, the soft term frequency (Soft-TF) of each word in the news title is calculated separately, and then the soft term frequencies of each word are summed to obtain the feature for ranking.

[0100] Step 403, Learn the significant feature representation: Based on the preliminary matching ranking features of the candidate news, apply a convolutional neural network and a max pooling function to obtain the local interest matching ranking features of each candidate news and the user's historical clicked news.

[0101] Based on the gradient received in backpropagation, the kernel pulls the similarity of the word pair closer to its μ to increase its Soft-TF count, or moves the word pair similarity away from μ to decrease the Soft-TF count. In the local representation, the word embeddings are initialized by Glove and trained by the kernel to provide an effective soft matching pattern. The word vector embeddings are updated based on the received gradient. Each kernel focuses on a group of word pairs belonging to a certain similarity range, and the importance of this similarity range is learned by the model. New matching patterns are discovered, moving some word pairs from near-zero similarity to important kernels. Intuitively, the learned word embedding matches form a multi-level Soft-TF pattern to find relevant news. The learned embedding parameters are used to memorize this information. Then, 1D-CNN and a max pooling neural network are used to identify and combine the stronger significant signals that match the candidate news in the entire historical click sequence as the final matching vector between the user and the candidate news. The local interest matching ranking features of the k-th candidate news and the historical clicked news at different similarities after convolution are expressed as Its calculation formula is:

[0102]

[0103] Where, and b h represent the parameters of the CNN filter, and ReLU is used as the activation non-linear function.

[0104] Step 5, Click prediction: Predict the probability that the user browses the candidate news based on the representation of the candidate news. Combine the user's global interest features, dynamic interest features, and local interest matching ranking features at different similarities to predict the probability that the user browses each candidate news.

[0105] In distributed matching, the global interest matching score of candidate news is calculated by the inner product of the candidate news and the user's global interest features. The matching score based on dynamic interest is the inner product of the dynamic interest features and the representation of the candidate news. The user interest representation obtained by encoding the user and the target perception interest mechanism is respectively matched with the candidate news embedding to obtain the matching score based on the user's global interest and the dynamic interest matching score activated by the candidate news The formula is as follows:

[0106]

[0107]

[0108] Based on the local interest matching ranking feature, the local interest matching scores derived from different similarity levels are calculated; in the local representation matching, the recommendation is made based on ranking the candidate news articles according to the probability that they are clicked by the user in the impression. The local interest matching score is

[0109]

[0110] where, W l and b l are learning parameters, is the final ranking feature representation of the i-th candidate news matching the historical clicked news

[0111] Finally, the scores of different methods are aggregated, and the matching scores of the three different methods are combined into the overall matching score (i.e., the probability that the user browses each candidate news):

[0112]

[0113] where, λ and η are hyperparameters that control the relative importance of the interest scores of different methods, and λ + η < 1

[0114] Since the number of positive and negative news samples is highly unbalanced. Therefore, negative sampling technology is applied in model training to jointly predict the click scores of k + 1 news. The k + 1 news consists of one positive sample of a user and one negative sample randomly selected from a user. Jointly predict the positive news and the prediction scores of k negative news . In this way, the news click prediction problem is transformed into a pseudo k + 1 classification task. The softmax is used to normalize these matching scores to calculate the matching scores of the positive samples. The loss function in the model training method is the negative log-likelihood of all positive samples, as follows:

[0115]

[0116] where y + represents the overall matching score of the i-th positive news, and y j - represents the overall matching score of the j-th negative news in the same time period as the i-th positive news, and S is the set of positive training samples.

[0117] Step 6: Recommend candidate news to the user based on the prediction probability, select the candidate news with a higher prediction probability, and recommend it to the user.

[0118] The present invention consists of two independent feature learning parts. One uses the local representation part to capture the multi-level soft matching signals between words in the text pair to achieve the matching between news at different similarity levels; the other uses the distributed representation part to dynamically activate the news interest features. In the local representation part, the representation of each news article is constructed, and then the user's candidate news is matched with the historical clicked news to form an association matrix between words. The multi-level matching interaction is applied to the association matrix, and multiple different kernel functions are used to capture the soft matching signals at multiple similarity levels, obtaining the matching scores of each news under different channels. Its channel represents the degree of relevance between the user and the candidate news at different similarity levels. Finally, a convolutional neural network is used to identify higher-level significant signals to predict the probability that the user clicks on the candidate news; in the distributed representation part, the context representation of the news title is learned by combining the CNN network and the attention mechanism. Instead of only using a fixed-size embedding vector to represent all the interests of the user, a new target attention network is added to dynamically activate the user's interests according to the candidate news, thereby strengthening the user's interest representation. These two parts are jointly trained as part of MIAR. In addition, a common observation in news recommendation is that most users usually only click on a few news shown in the exposure sequence. Therefore, the number of positive and negative news samples is highly unbalanced. An important feature of the MIND dataset used in the present invention is class imbalance, and the positive sample rate is very low, only 4%. Therefore, in the process of model training, the present invention applies the negative sampling technique to jointly predict the click score for K + 1 news. That is, the news consists of a positive sample of one user and a negative sample randomly selected from another user. Then, jointly predict the positive news and a negative news.

[0119] Considering the wide range of candidate news and the diversity of user interests, it is difficult to accurately model with a single fixed vector, which limits the representation ability of the recommendation model. Different from previous work that abstracts the historical news clicked by users into a fixed user interest vector to match candidate news, the model of the present invention combines multi-level matching interactions and retains the fine-grained information of the text. Modeling text word pairs at multiple different similarity levels makes the encoded user interest information richer.

[0120] In addition, the present invention realizes that the browsing news sequence of users at different stages can reflect the change of user interests, and static vectors cannot fully express user interests. Accordingly, an attention network based on candidate news is designed, which can adaptively activate the interests of different users in different candidate news, deduce the dynamic feature vectors of user interests according to different candidate news, further strengthen the user interest representation, and greatly improve the expressiveness of the model. Finally, the weights of the scores of different modules are effectively fused, thus providing an improved trade-off between accuracy and speed. Experimental results show that experiments are carried out on the real news dataset of MIND, verifying the effectiveness of the model of the present invention. In addition, through case analysis, the target attention network module also enhances the robustness and interpretability of click prediction.

[0121] Embodiment 2

[0122] This embodiment provides an interest-activated news recommendation system based on multi-level matching, which specifically includes the following modules:

[0123] An interest-activated news recommendation system based on multi-level matching of the present invention mainly consists of two parts: a local representation matching module and a distributed matching module. The two core modules of this system include the following main components: a news encoding module, a user encoding module, a target-aware interest activation module, a multi-level matching interaction module, and a prediction module for calculating the probability of a user clicking on a candidate news.

[0124] The news encoding module is configured to: obtain candidate news and user historical clicked news, and perform news encoding respectively to obtain candidate news representation and user historical clicked news representation;

[0125] The user encoding module is configured to: obtain the user's global interest feature based on the user historical clicked news representation;

[0126] The target-aware interest activation module is configured to: based on the candidate news representation and the user historical clicked news representation, use the target attention mechanism to calculate the attention scores between each user historical clicked news and each candidate news, and use the attention scores to perform weighted calculation on the user historical clicked news to obtain the user's dynamic interest feature;

[0127] A multi-level matching interaction module, which is configured to: capture the semantic correlation between the clicked news and the candidate news at multiple different similarity levels based on the candidate news representation and the user's historical clicked news representation, and obtain the local interest matching ranking features of each candidate news at different levels with the user's historical clicked news;

[0128] A probability prediction module, which is configured to: jointly use the user's global interest feature, dynamic interest feature, and local interest matching ranking feature to predict the probability of the user browsing each candidate news, and recommend candidate news to the user based on the predicted probability.

[0129] It should be noted here that each module in this embodiment corresponds one by one to each step in Embodiment 1, and its specific implementation process is the same, so it will not be repeated here.

[0130] Embodiment 3

[0131] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in a method for interest-activated news recommendation based on multi-level matching as described in Embodiment 1 above.

[0132] Embodiment 4

[0133] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in a method for interest-activated news recommendation based on multi-level matching as described in Embodiment 1 above.

[0134] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0135] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and / or blocksFigure 1 means for the functions specified in one or more boxes.

[0136] These computer program instructions can 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, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction means that implements the functions specified in one Figure 1 process or multiple processes and / or boxes Figure 1 or more boxes.

[0137] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 process or multiple processes and / or boxes Figure 1 or more boxes.

[0138] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The said program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the said storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0139] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An interest activation news recommendation method based on multi-level matching, characterized in that, Including: Obtain candidate news and the user's historical clicked news, perform news encoding on them respectively to obtain the candidate news representation and the user's historical clicked news representation; Based on the user's historical clicked news representation, adopt a personalized attention mechanism to obtain the user's global interest feature; The user's global interest feature is the weighted sum of the user's historical clicked news representation; Based on the candidate news representation and the user's historical clicked news representation, adopt a target attention mechanism to calculate the attention scores between each user's historical clicked news and each candidate news, and use the attention scores to perform weighted calculation on the user's historical clicked news to obtain the user's dynamic interest feature; Based on the candidate news representation and the user's historical clicked news representation, capture the semantic relevance between the clicked news and the candidate news at multiple different similarity levels to obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different levels; The steps for obtaining the local interest matching ranking features are as follows: Based on the candidate news representation and the user's historical clicked news representation, calculate the association matrix; pool the association matrix using the Gaussian kernel function to convert the word interaction in the association matrix into the preliminary matching ranking features of the candidate news; Based on the preliminary matching ranking features of the candidate news, apply a convolutional neural network and a max pooling function to obtain the local interest matching ranking features of each candidate news and the user's historical clicked news; Combine the user's global interest feature, dynamic interest feature, and local interest matching ranking features to predict the probability that the user browses each candidate news, and recommend candidate news to the user based on the predicted probability.

2. The method for interest-activated news recommendation based on multi-level matching according to claim 1, wherein, The specific steps of the news encoding are as follows: Obtain the title of the news; Use the news title to construct a vector representation to obtain the word vector sequence of the news title; Input the news title word vector sequence into a convolutional neural network to obtain a sequence of context word representation vectors; Input the sequence of context word representation vectors into a word-level attention network to obtain the final representation of the news.

3. The interest activation news recommendation method based on multi-level matching according to claim 1, characterized in that, The specific steps for predicting the probability that the user browses each candidate news are as follows: Based on the user's global interest feature, dynamic interest feature, and candidate news representation, calculate the matching score based on the user's global interest and the matching score based on the dynamic interest; Based on the local interest matching ranking features, calculate the local interest matching score; Combine the matching score based on the user's global interest, the matching score based on the dynamic interest, and the local interest matching score to calculate the probability that the user browses each candidate news.

4. The interest activation news recommendation method based on multi-level matching according to claim 3, characterized in that, The matching score based on the user's global interest is the inner product of the user's global interest feature and the candidate news representation.

5. The interest activation news recommendation method based on multi-level matching according to claim 3, characterized in that, The matching score based on the dynamic interest is the inner product of the dynamic interest feature and the candidate news representation.

6. An interest activation news recommendation system based on multi-level matching, characterized in that, Including: A news encoding module, which is configured to: Obtain candidate news and the user's historical clicked news, perform news encoding on them respectively to obtain the candidate news representation and the user's historical clicked news representation; A user encoding module, which is configured to: Based on the user's historical clicked news representation, adopt a personalized attention mechanism to obtain the user's global interest feature; The user's global interest feature is the weighted sum of the user's historical clicked news representation; A target perception interest activation module, which is configured to: based on the candidate news representation and the user's historical clicked news representation, adopt a target attention mechanism to calculate the attention scores between each user's historical clicked news and each candidate news, and use the attention scores to perform weighted calculation on the user's historical clicked news to obtain the user's dynamic interest feature; A multi-level matching interaction module, which is configured to: based on the candidate news representation and the user's historical clicked news representation, capture the semantic correlation between the clicked news and the candidate news at multiple different similarity levels, and obtain the local interest matching ranking features of each candidate news and the user's historical clicked news at different levels; The step of obtaining the local interest matching ranking feature is: based on the candidate news representation and the user's historical clicked news representation, calculate the association matrix; pool the association matrix using a Gaussian kernel function, and convert the word interaction in the association matrix into the preliminary matching ranking feature of the candidate news; Based on the preliminary matching ranking feature of the candidate news, apply a convolutional neural network and a max pooling function to obtain the local interest matching ranking feature of each candidate news and the user's historical clicked news; A probability prediction module, which is configured to: combine the user's global interest feature, dynamic interest feature and local interest matching ranking feature, predict the probability that the user browses each candidate news, and recommend the candidate news to the user based on the predicted probability.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps in a news recommendation method based on multi-level matching for interest activation as described in any one of claims 1-5.

8. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in a news recommendation method based on multi-level matching for interest activation as described in any one of claims 1-5.

Images