News recommendation system and method based on artificial intelligence

Abstract

The present invention relates to the field of artificial intelligence and information recommendation technology, and in particular to a news recommendation system and method based on artificial intelligence. It includes: collecting and preprocessing user behavior data and news content data, generating interactive embedding vectors of user interests and news content based on the preprocessed user vector sequence and news vector sequence; introducing a dynamic two-way propagation network model to obtain a high-dimensional comprehensive representation vector, and further obtaining the predicted value of the dynamic two-way propagation network model; calculating the loss function based on the predicted value of the dynamic two-way propagation network model; dynamically adjusting the parameters of the dynamic two-way propagation network model, optimizing the dynamic two-way propagation network model, until the loss function reaches the minimum value, and obtaining the final dynamic two-way propagation network model. It solves the technical problems of one-sidedness in user interest modeling, limitations in news content analysis, low precision in personalized recommendation, poor adaptability of the model, and lack of flexibility and dynamism in recommendation results.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and information recommendation technology, and in particular to a news recommendation system and method based on artificial intelligence. Background Art

[0002] With the rapid development of Internet technology, the ways and channels of obtaining information have become more diverse, and news reading has become an indispensable part of people's daily lives. However, faced with a huge amount of news, users often face the problem of information overload when obtaining news content of their interest. In order to help users quickly find content that suits their personal interests from a large amount of news, news recommendation systems have emerged. These news recommendation systems analyze user behavior and preferences to provide users with personalized news recommendations, thereby improving users' experience of news reading and increasing user stickiness.

[0003] In recent years, the progress of artificial intelligence technology, especially deep learning technology, has provided new possibilities for the intelligent development of news recommendation systems. Through machine learning and deep learning technology, news recommendation systems can mine more complex and implicit interest patterns from users' historical behaviors, and adjust news recommendation strategies in real time to provide more accurate and personalized services. At the same time, with the development of social media and mobile Internet, users' behavior data on different platforms is increasing, and the diversity of news content and user needs have become more complex. How to effectively use user behavior data and news content to improve the effectiveness of news recommendation systems has become a hot topic of research.

[0004] In addition, the timeliness and popularity of the news content itself are also key factors that the news recommendation system needs to consider. As news events continue to develop, users' focus is also changing dynamically. How to effectively integrate real-time hot information in the news recommendation process and keep the news recommendation content consistent with the current hot spots is one of the keys to improving the value of the news recommendation system.

[0005] The existing news recommendation system has the following technical problems: user interest modeling is one-sided and it is difficult to fully capture user interests; news content analysis has limitations, resulting in low prediction accuracy; in terms of personalized recommendations, the accuracy needs to be improved and it is difficult to accurately match user preferences; the model has poor adaptability and is weak in adapting to new situations and new users; the real-time performance of the news recommendation system needs to be improved; and the recommendation results lack flexibility and dynamism, making it difficult to quickly respond to and meet the ever-changing needs of users. Summary of the invention

[0006] The present invention provides a news recommendation system and method based on artificial intelligence to solve the following technical problems: user interest modeling is one-sided and it is difficult to fully capture user interests; news content analysis has limitations, resulting in low prediction accuracy; in terms of personalized recommendation, the accuracy needs to be improved and it is difficult to accurately match user preferences; the model has poor adaptability and is weak in adapting to new situations and new users; the real-time performance of the news recommendation system needs to be improved; and the recommendation results lack flexibility and dynamism, making it difficult to quickly respond to and meet the ever-changing needs of users.

[0007] The news recommendation system and method based on artificial intelligence of the present invention specifically include the following technical solutions:

[0008] A news recommendation method based on artificial intelligence includes the following steps:

[0009] S1. Collect and preprocess user behavior data and news content data to obtain a preprocessed user vector sequence and a preprocessed news vector sequence; based on the preprocessed user vector sequence and the preprocessed news vector sequence, generate an interactive embedding vector of user interests and news content through a multi-level nonlinear adaptive attention network;

[0010] S2. Based on the interactive embedding vector of user interests and news content, a dynamic two-way communication network model is introduced to obtain a comprehensive representation vector; based on the comprehensive representation vector, the predicted value of the dynamic two-way communication network model is obtained; based on the predicted value of the dynamic two-way communication network model, the loss function is calculated; the parameters of the dynamic two-way communication network model are dynamically adjusted to optimize the dynamic two-way communication network model until the loss function reaches the minimum value, thereby obtaining the final dynamic two-way communication network model.

[0011] Preferably, the S1 specifically includes:

[0012] In the multi-level nonlinear adaptive attention network, a multi-level attention mechanism and adaptive weight adjustment are introduced.

[0013] Preferably, the S1 specifically includes:

[0014] The preprocessed user vector sequence and the preprocessed news vector sequence are used as the input of a multi-level nonlinear adaptive attention network. The similarity between the user vector and the news vector is calculated in each layer of the nonlinear adaptive attention network, and the attention weight between the user vector and the news vector is obtained.

[0015] Preferably, the S1 specifically includes:

[0016] Based on the attention weights between the user vector and the news vector and the hidden state vector of the current layer of the multi-level nonlinear adaptive attention network, the hidden state vector of the next layer of the multi-level nonlinear adaptive attention network is obtained.

[0017] Preferably, the S1 specifically includes:

[0018] By weighted combination of hidden state vectors of each layer in a multi-level nonlinear adaptive attention network, an interactive embedding vector of user interests and news content is generated.

[0019] Preferably, the S2 specifically includes:

[0020] In the process of implementing the dynamic two-way propagation network model, through two-way information flow and dynamic path adjustment, forward and reverse embedding vectors are generated based on the interactive embedding vectors of user interests and news content; based on the forward and reverse embedding vectors, a comprehensive representation vector is obtained.

[0021] Preferably, the S2 specifically includes:

[0022] Based on the predicted value of the dynamic bidirectional propagation network model, a loss function is defined; the calculation formula of the loss function is:

[0023] ,

[0024] in, represents the loss function; represents the first The actual labels of the samples; Represents the dynamic bidirectional propagation network model for the The predicted value of samples; Indicates the total number of samples; is the regularization coefficient, which is used to control the weight of the regularization term; represents regularization; Indicates The weight matrix of the layer nonlinear adaptive attention network; Represents the total number of layers of the nonlinear adaptive attention network.

[0025] A news recommendation system based on artificial intelligence, including the following parts:

[0026] The data collection module collects user behavior data and news content data in real time and transmits the user behavior data and news content data to the data preprocessing module;

[0027] The data preprocessing module preprocesses the user behavior data and the news content data respectively to form a preprocessed user vector sequence and a preprocessed news vector sequence; and transmits the preprocessed user vector sequence and the preprocessed news vector sequence to the adaptive attention module;

[0028] The adaptive attention module introduces a multi-level nonlinear adaptive attention network based on the preprocessed user vector sequence and the preprocessed news vector sequence, calculates the similarity between the user vector and the news vector in each layer of the nonlinear adaptive attention network, and obtains the attention weight between the user vector and the news vector; dynamically adjusts the attention weight between the user vector and the news vector, and calculates the hidden state vector of each layer of the multi-level nonlinear adaptive attention network; performs weighted combination of the hidden state vectors of each layer of the multi-level nonlinear adaptive attention network to generate the interactive embedding vector of the user interest and the news content; and transmits the interactive embedding vector of the user interest and the news content to the dynamic two-way propagation module;

[0029] The dynamic two-way propagation module generates forward and reverse interactive embedding vectors based on the interactive embedding vectors of user interests and news content through the dynamic two-way propagation network model; generates a comprehensive representation vector based on the forward and reverse interactive embedding vectors; and passes the comprehensive representation vector to the prediction and loss calculation module;

[0030] The prediction and loss calculation module obtains the prediction value of the dynamic bidirectional propagation network model based on the comprehensive representation vector, calculates the prediction error using the loss function, and passes the loss function to the model optimization and iteration module;

[0031] The model optimization and iteration module adjusts the parameters of the dynamic bidirectional propagation network model and optimizes the dynamic bidirectional propagation network model until the loss function reaches the minimum value, thereby obtaining the final dynamic bidirectional propagation network model.

[0032] The beneficial effects of the technical solution of the present invention are:

[0033] 1. The multi-level nonlinear adaptive attention network effectively captures the complex interactive relationship between user interests and news content through multi-level attention mechanisms and adaptive weight adjustment, and generates interactive embedding vectors of user interests and news content; by fusing multi-level interactive embedding vectors, a highly abstract user interest representation is formed, which can more accurately reflect the user's interest in different news content, thereby improving the accuracy of personalized news recommendations.

[0034] 2. The dynamic two-way communication network model not only realizes the two-way interactive modeling of user interests and news content, but also realizes the deep integration of user interests and news content through two-way information flow and dynamic path adjustment, thereby ensuring the accuracy and personalization of news recommendation results; through high-dimensional comprehensive representation vectors, the dynamic two-way communication network model can dynamically reflect changes in user interests and make real-time adjustments to news content, thereby enhancing the adaptive ability of the dynamic two-way communication network model.

[0035] 3. Through optimization algorithms such as back propagation algorithm and gradient descent, the news recommendation system can continuously adjust and optimize the parameters of the dynamic two-way propagation network model to minimize the loss function, making the predicted value of the dynamic two-way propagation network model closer to the actual situation; the combination of cross entropy loss function and regularization term not only effectively reduces the prediction error, but also avoids overfitting of the dynamic two-way propagation network model, ensuring the generalization ability and stability of the news recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a structural diagram of the news recommendation system based on artificial intelligence according to the present invention;

[0037] Figure 2 This is a flow chart of the news recommendation method based on artificial intelligence described in the present invention. DETAILED DESCRIPTION

[0038] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0040] The specific scheme of the news recommendation system and method based on artificial intelligence provided by the present invention is described in detail below with reference to the accompanying drawings.

[0041] See attached Figure 1 , which shows a structural diagram of a news recommendation system based on artificial intelligence provided by an embodiment of the present invention, the system includes the following parts:

[0042] Data acquisition module, data preprocessing module, adaptive attention module, dynamic two-way propagation module, prediction and loss calculation module, model optimization and iteration module;

[0043] The data collection module collects user behavior data on different platforms in real time, captures news content data from news platforms and content providers in real time, and passes the collected user behavior data and news content data to the data preprocessing module;

[0044] The data preprocessing module performs preprocessing on the collected user behavior data, such as feature extraction and embedding operations, to form a preprocessed user vector sequence; preprocesses the news content data through natural language processing techniques, such as word segmentation, word vector embedding and topic modeling, to form a preprocessed news vector sequence; and transmits the preprocessed user vector sequence and the preprocessed news vector sequence to the adaptive attention module;

[0045] The adaptive attention module introduces a multi-level nonlinear adaptive attention network based on the preprocessed user vector sequence and news vector sequence. Through the multi-level attention mechanism, it accurately captures the complex interactive relationship between user interests and news content. In each layer of the nonlinear adaptive attention network, the similarity between the user vector and the news vector is calculated, and the attention weight between the user vector and the news vector is obtained. The attention weight between the user vector and the news vector is dynamically adjusted according to the attention adjustment factor, and the hidden state vector of each layer of the multi-level nonlinear adaptive attention network is calculated. The hidden state vectors of each layer of the multi-level nonlinear adaptive attention network are weightedly combined to generate the interactive embedding vector of user interests and news content, and the interactive embedding vector of user interests and news content is passed to the dynamic two-way propagation module.

[0046] The dynamic two-way propagation module generates forward and reverse interactive embedding vectors based on the interactive embedding vectors of user interests and news content through the dynamic two-way propagation network model. Based on the forward and reverse interactive embedding vectors, the two-way interactive relationship between user interests and news content is captured by dynamically adjusting forward and reverse propagation, and finally a high-dimensional comprehensive representation vector is generated to predict the probability value of users clicking on news, and the high-dimensional comprehensive representation vector is passed to the prediction and loss calculation module.

[0047] The prediction and loss calculation module obtains the predicted probability value of users clicking on news based on the high-dimensional comprehensive representation vector, that is, the predicted value of the dynamic two-way propagation network model; based on the predicted value of the dynamic two-way propagation network model, the prediction error is calculated using the loss function, and the loss function is passed to the model optimization and iteration module;

[0048] The model optimization and iteration module updates the parameters of the dynamic two-way propagation network model through the back-propagation algorithm to minimize the loss function; gradually adjusts the parameters of the dynamic two-way propagation network model through optimization algorithms such as gradient descent, and gradually optimizes the dynamic two-way propagation network model after multiple iterations until the loss function converges and stops. Finally, the dynamic two-way propagation network model can adapt to the interests and behavioral characteristics of different users and provide accurate personalized news recommendations.

[0049] See attached Figure 2 , which shows a flow chart of a news recommendation method based on artificial intelligence provided by an embodiment of the present invention, the method comprising the following steps:

[0050] S1. Collect and preprocess user behavior data and news content data to obtain a preprocessed user vector sequence and a preprocessed news vector sequence; based on the preprocessed user vector sequence and the preprocessed news vector sequence, generate an interactive embedding vector of user interests and news content through a multi-level nonlinear adaptive attention network;

[0051] The data collection module collects user behavior data in real time, including user click records, browsing time, reading frequency, likes and comments on different platforms, and integrates user search history, social media interaction data, content preference tags and other information to comprehensively portray the user's interest profile; in terms of news content, news content data is captured from news platforms and content providers in real time. After preliminary cleaning and deduplication, the news content data is further extracted, including structural features such as news text content, release time, author information, and subject classification. In order to improve the richness and accuracy of news content data, web crawler technology is used to dynamically obtain external information such as news hotspots and user comments to ensure the comprehensiveness and timeliness of news content data.

[0052] The data preprocessing module preprocesses the collected user behavior data and news content data. Specifically, the user behavior data is converted into a high-dimensional vector representation after feature extraction and embedding operations to form a preprocessed user vector sequence. , represents the number of user vectors, It is The news content data is processed through natural language processing techniques, such as word segmentation, word vector embedding, and topic modeling, to generate a semantic vector representation of the news, forming a preprocessed news vector sequence. , represents the number of news vectors, It is News vector.

[0053] After preprocessing the user behavior data and news content data, the adaptive attention module introduces a multi-level nonlinear adaptive attention network, which accurately captures the complex interactive relationship between user interests and news content through multi-level attention mechanisms and adaptive weight adjustment. The preprocessed user vector sequence and news vector sequence are used as the input of the multi-level nonlinear adaptive attention network. The similarity between the user vector and the news vector is calculated in each layer of the nonlinear adaptive attention network, and the attention weight between the user vector and the news vector is obtained; the specific formula is as follows:

[0054] ,

[0055] in, Indicates The first layer of nonlinear adaptive attention network User vector and News vector The attention weight between It is user vectors; It is news vector; is the attention adjustment factor, which is used to control the sensitivity of the attention mechanism, that is, the degree of influence of different news vectors on the current user vector; It is The nonlinear matching function of the layer nonlinear adaptive attention network is used to measure the User vector and News vector The similarity between . The nonlinear matching function is defined as:

[0056] ,

[0057] in, Indicates The weight matrix of the layer nonlinear adaptive attention network; It is The bias vector of the layer nonlinear adaptive attention network; represents the element-wise product between the user vector and the news vector, i.e., the Hadamard product; It is a nonlinear activation function, such as ReLU, which is used to introduce nonlinear characteristics to enhance the expressive power of the multi-level nonlinear adaptive attention network. The attention weight between the user vector and the news vector in the current layer of the nonlinear adaptive attention network is weighted with the user vector to obtain the hidden state vector of the next layer of the nonlinear adaptive attention network:

[0058] ,

[0059] in, Indicates Hidden state vector of the layer nonlinear adaptive attention network; Indicates Hidden state vector of the layer nonlinear adaptive attention network; It is The weight matrix of the layer nonlinear adaptive attention network; It is The bias vector of the layer nonlinear adaptive attention network; It is a coefficient used to control the information transfer between layers of the nonlinear adaptive attention network to ensure that the multi-level nonlinear adaptive attention network can maintain the information coherence and consistency between different layers during the deep learning process. Through the multi-level attention mechanism, the user's interests can be dynamically captured and the relevance of news recommendations can be continuously adjusted. Finally, the multi-level hidden state vectors are weighted and combined to generate the interactive embedding vector of user interests and news content:

[0060] ,

[0061] in, It is the interactive embedding vector of user interests and news content, which represents the interactive information between user interests and news content, including the user's potential interest in news and the potential impact of news content on user behavior. It is the weighted combination result of multi-level hidden state vectors. represents the total number of layers of the nonlinear adaptive attention network; It is an adaptive weight dynamically adjusted through back propagation, reflecting the influence of each layer's hidden state vector on the interaction embedding vector. By fusing multiple levels of interaction embedding vectors, a highly abstract user interest representation is formed, reflecting the user's interest in different news content.

[0062] S2. Based on the interactive embedding vector of user interests and news content, a dynamic two-way communication network model is introduced to obtain a high-dimensional comprehensive representation vector; based on the high-dimensional comprehensive representation vector, the predicted value of the dynamic two-way communication network model is obtained; based on the predicted value of the dynamic two-way communication network model, the loss function is calculated; the parameters of the dynamic two-way communication network model are dynamically adjusted to optimize the dynamic two-way communication network model until the loss function reaches the minimum value, thereby obtaining the final dynamic two-way communication network model.

[0063] The dynamic two-way propagation module adopts a dynamic two-way propagation network model. Through two-way information flow and dynamic path adjustment, it realizes the deep integration of user interests and news content. According to the interactive embedding vectors of user interests and news content, it generates forward and reverse embedding vectors respectively:

[0064] ,

[0065] ,

[0066] in, Indicated in time steps from the user vector To news vector The forward embedding vector of is used to map how user interests affect news content recommendations; Indicated in time step from the news vector To user vector The reverse embedding vector of , which is used to map how news content in turn affects user interests; and are the weight matrices for forward mapping and reverse mapping respectively; and are the bias vectors of the forward mapping and reverse mapping, respectively. In the process of forward propagation, a mapping from user vector to news vector is generated, while in the process of reverse propagation, a mapping from news vector to user vector is generated, thus forming a two-way information interaction mechanism.

[0067] The dynamic two-way propagation network model ensures the multi-level expression of the complex interaction between user interests and news content through the nonlinear activation function embedded between layers. The interactive embedding vectors of forward propagation and backward propagation are transformed linearly and activated nonlinearly to generate a high-dimensional comprehensive representation vector between user interests and news content. The specific formula of the high-dimensional comprehensive representation vector is:

[0068] ,

[0069] Among them, the high-dimensional comprehensive representation vector It is a condensed expression of the complex interactive relationship between user interests and news content, and contains information about users’ potential preferences for various types of news; is the weight matrix used to embed the forward embedding vector generated by forward propagation and the reverse embedding vector generated by back-propagation Perform linear transformation; is the two-way information fusion coefficient; is the bias vector used in the weight matrix Adjustments are made based on the linear transformation results.

[0070] The prediction and loss calculation module converts the high-dimensional comprehensive representation vector into an actual prediction value, that is, the predicted probability value of a user clicking on a specific news item; it maps the high-dimensional comprehensive representation vector to a scalar score through linear transformation, which represents the "tendency" or "score" of the user to click on the news item. The calculation formula is as follows:

[0071] ,

[0072] in, It is the first The scalar fraction of samples; is a weight vector used to transform the high-dimensional comprehensive representation vector into a scalar score; Indicates A high-dimensional comprehensive representation vector; represents transpose; is the bias that allows the dynamic bidirectional propagation network model to produce a basic output value even without other inputs.

[0073] The sigmoid activation function is used to convert the scalar score into a probability value. represents the predicted probability value of the user clicking on the news, that is, the dynamic two-way propagation network model for the The predicted value of samples:

[0074] ,

[0075] Among them, the technical purpose of the sigmoid activation function is to convert any real number (here refers to ) is compressed to the range of [0,1] so that the conversion result of the scalar score can be interpreted as a probability value. , which is directly used to measure the difference between the predicted value of the dynamic bidirectional propagation network model and the actual label.

[0076] The error between the predicted value and the actual label of the dynamic bidirectional propagation network model, i.e., the prediction error, is quantified through the loss function, which provides a basis for optimizing the parameters of the dynamic bidirectional propagation network model and thus improves the prediction accuracy of the dynamic bidirectional propagation network model. The calculation formula of the loss function can be expressed as:

[0077] ,

[0078] in, represents the loss function; represents the first The actual labels of samples, in the news recommendation scenario, It may indicate whether the user clicks on the recommended news, i.e. It is a binary variable with a value of 0 or 1; Represents the dynamic bidirectional propagation network model for the The predicted value of a sample, that is, the predicted probability value of the user clicking on the news; represents the total number of samples. The first two terms of the loss function are the standard form of the cross entropy loss function, which is used to measure the error between the predicted value and the actual label of the dynamic bidirectional propagation network model in the binary classification task. Specifically: when When When it is close to 0, that is, the user actually clicks on the news; when When The case close to 1 means that the user did not click on the news. is the regularization coefficient, which is used to control the weight of the regularization term and determines the contribution of the regularization term to the overall loss. , which can control the complexity of the dynamic bidirectional propagation network model and avoid overfitting; Represents regularization.

[0079] For each input sample, the dynamic bidirectional propagation network model is calculated through forward propagation. The predicted value of samples , which is a probability value between 0 and 1, indicating the probability that a user will click on a piece of news. and dynamic bidirectional propagation network model for the The predicted value of samples , calculate the loss function. For each sample, the loss function measures the match between the predicted probability distribution output by the dynamic bidirectional propagation network model and the actual label.

[0080] After calculating the loss function, the model optimization and iteration module passes the partial derivative of the loss function for each parameter (including weights and biases) back to each layer of the dynamic two-way propagation network through the back-propagation algorithm, and updates the parameters to minimize the loss function. Through optimization algorithms such as gradient descent, the parameters of the dynamic two-way propagation network model are gradually adjusted so that the value of the loss function is smaller in the next iteration, so that the predicted value of the dynamic two-way propagation network model is closer to the actual label. Through multiple iterations, the dynamic two-way propagation network model is gradually optimized so that the dynamic two-way propagation network model converges on the training set, that is, the loss function reaches the minimum value. Finally, the dynamic two-way propagation network model can adapt to the interests and behavioral characteristics of different users, thereby providing accurate personalized news recommendations.

[0081] In summary, the news recommendation system and method based on artificial intelligence have been completed.

[0082] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0083] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0084] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A news recommendation method based on artificial intelligence, characterized in that: The following steps are involved: S1. Collect and preprocess user behavior data and news content data to obtain preprocessed user vector sequences and preprocessed news vector sequences; based on the preprocessed user vector sequences and preprocessed news vector sequences, introduce a multi-level nonlinear adaptive attention network, and introduce a multi-level attention mechanism and adaptive weight adjustment in the multi-level nonlinear adaptive attention network; use the preprocessed user vector sequence and the preprocessed news vector sequence as inputs of the multi-level nonlinear adaptive attention network, calculate the similarity between the user vector and the news vector in each layer of the nonlinear adaptive attention network, and obtain the attention weight between the user vector and the news vector. The specific formula is as follows: ， in, Indicates The first layer of nonlinear adaptive attention network User vector and News vector The attention weight between It is user vectors; It is news vectors; represents the number of news vectors; is the attention regulator; It is Nonlinear matching function of layer nonlinear adaptive attention network; Based on the attention weights between the user vector and the news vector and the hidden state vector of the current layer of the multi-level nonlinear adaptive attention network, the hidden state vector of the next layer of the multi-level nonlinear adaptive attention network is obtained: ， in, Indicates Hidden state vector of the layer nonlinear adaptive attention network; Indicates Hidden state vector of the layer nonlinear adaptive attention network; is a nonlinear activation function; It is The weight matrix of the layer nonlinear adaptive attention network; represents the number of user vectors; It is Bias vector of the layer nonlinear adaptive attention network; is the coefficient used to control the information transfer between layers of the nonlinear adaptive attention network; By weighted combination of the hidden state vectors of each layer in the multi-level nonlinear adaptive attention network, the interactive embedding vector of user interests and news content is generated: ， in, is the interactive embedding vector of user interests and news content; represents the total number of layers of the nonlinear adaptive attention network; are adaptive weights that are dynamically adjusted through back-propagation; S2. Based on the interactive embedding vector of user interests and news content, a dynamic two-way propagation network model is introduced. In the implementation process of the dynamic two-way propagation network model, the forward and reverse embedding vectors are generated based on the interactive embedding vector of user interests and news content through two-way information flow and dynamic path adjustment: ， ， in, Indicated in time steps from the user vector To news vector The forward embedding vector of ; Indicated in time step from the news vector To user vector The reverse embedding vector of and are the weight matrices for forward mapping and reverse mapping respectively; and are the bias vectors for forward mapping and reverse mapping respectively; Based on the forward and reverse embedding vectors, the comprehensive representation vector is obtained. The specific formula is: ， in, is the comprehensive representation vector; is the weight matrix; is the two-way information fusion coefficient; is the bias vector; Based on the comprehensive representation vector, the prediction value of the dynamic two-way propagation network model is obtained; based on the prediction value of the dynamic two-way propagation network model, the loss function is calculated; the parameters of the dynamic two-way propagation network model are dynamically adjusted to optimize the dynamic two-way propagation network model until the loss function reaches a minimum value, thereby obtaining the final dynamic two-way propagation network model.

2. The news recommendation method based on artificial intelligence according to claim 1, characterized in that: The S2 specifically includes: Based on the predicted value of the dynamic bidirectional propagation network model, a loss function is defined; the calculation formula of the loss function is: ， in, represents the loss function; represents the first The actual labels of the samples; Represents the dynamic bidirectional propagation network model for the The predicted value of samples; Indicates the total number of samples; is the regularization coefficient, which is used to control the weight of the regularization term; represents regularization; Indicates The weight matrix of the layer nonlinear adaptive attention network; Represents the total number of layers of the nonlinear adaptive attention network.

3. A news recommendation system based on artificial intelligence, applied to the news recommendation method based on artificial intelligence as claimed in claim 1, characterized in that: Includes the following parts: The data collection module collects user behavior data and news content data in real time and transmits the user behavior data and news content data to the data preprocessing module; The data preprocessing module preprocesses the user behavior data and the news content data respectively to form a preprocessed user vector sequence and a preprocessed news vector sequence; and transmits the preprocessed user vector sequence and the preprocessed news vector sequence to the adaptive attention module; The adaptive attention module introduces a multi-level nonlinear adaptive attention network based on the preprocessed user vector sequence and the preprocessed news vector sequence, calculates the similarity between the user vector and the news vector in each layer of the nonlinear adaptive attention network, and thus obtains the attention weight between the user vector and the news vector; Dynamically adjust the attention weights between the user vector and the news vector, and calculate the hidden state vectors of each layer of the multi-level nonlinear adaptive attention network; perform weighted combination of the hidden state vectors of each layer of the multi-level nonlinear adaptive attention network to generate the interactive embedding vectors of user interests and news content; pass the interactive embedding vectors of user interests and news content to the dynamic two-way propagation module; The dynamic two-way propagation module generates forward and reverse interactive embedding vectors based on the interactive embedding vectors of user interests and news content through the dynamic two-way propagation network model; generates a comprehensive representation vector based on the forward and reverse interactive embedding vectors; and passes the comprehensive representation vector to the prediction and loss calculation module; The prediction and loss calculation module obtains the prediction value of the dynamic bidirectional propagation network model based on the comprehensive representation vector, calculates the prediction error using the loss function, and passes the loss function to the model optimization and iteration module; The model optimization and iteration module adjusts the parameters of the dynamic bidirectional propagation network model and optimizes the dynamic bidirectional propagation network model until the loss function reaches the minimum value, thereby obtaining the final dynamic bidirectional propagation network model.

Images