An efficient recommendation method based on user session and social network

Abstract

The application provides a recommendation method based on user conversation and social network, comprising the following steps: constructing a knowledge graph embedding network: using social network information and historical behavior information of a user, a knowledge graph embedding network for obtaining user conversation and social network related information is constructed; training and learning of a heterogeneous knowledge graph network: a heterogeneous knowledge graph network is established by using the knowledge graph embedding network, and hidden information in the user and conversation in the heterogeneous knowledge graph network is learned; embedding of personalized conversation features: for the user and conversation information in the heterogeneous knowledge graph network, personalized conversation embedding information for each user is obtained by using a calculation module, personal preference information of the user and a context feature vector in the conversation are obtained, a final calculation function of user preference is obtained by using weighted summation of the personal preference and the context feature vector, and a final representation mode of user personal preference is obtained in combination with cross entropy.

Description

Technical Field

[0001] This invention relates to the field of information and Internet information recommendation technology, and in particular to an efficient recommendation method based on user conversations and social networks. Background Technology

[0002] Recommender systems are a new technology that has emerged with the booming development of information technology and the internet industry. Their main purpose is to solve the problem of information overload causing inconvenience for users in their daily internet use. For users, finding items, information, and corresponding social contacts that interest them from massive amounts of data and resources is a primary concern. For information platforms, trading platforms, and social networking platforms, recommending targeted products, information, and social contacts to different users is an effective solution to promote the prosperity and development of the entire internet platform and boost economic and trade growth. Serving users through effective recommendation algorithms, thereby saving users time, improving satisfaction, and creating more commercial value for companies and society, is a key research focus for many internet technology companies. Currently, recommender systems are widely used in various industries, including search, movies, music, news, books, food, and financial services, playing a vital role in each. Recommender systems typically generate recommendation lists in two ways: collaborative filtering and content-based recommendation, also known as user-based personalized recommendation. Collaborative filtering methods primarily build models by combining a user's historical behavior (such as the items or services they have purchased, selected, or rated) with similar decisions made by other users, thus linking the interests of different users. This model can be used to predict which items different users might be interested in (primarily item-based collaborative filtering). Content-based recommendation schemes, on the other hand, utilize discrete features corresponding to goods or services. By abstracting the objects, they obtain the corresponding feature information contained in different objects and recommend items or services that users are more interested in. These two traditional recommendation systems are suitable for relatively simple scenarios and systems. With the rapid development of the internet and the explosive increase in platforms and information, single recommendation systems can no longer meet current needs. Therefore, hybrid recommendation systems combining the two different recommendation schemes have emerged.

[0003] Machine learning (ML) is the science that studies how to use computers to simulate or implement human learning activities. It is one of the most intelligent and cutting-edge research areas in artificial intelligence. Since the 1980s, machine learning has attracted widespread interest from research institutions and high-tech companies as a means to achieve artificial intelligence. Especially in the last decade or so, research in machine learning has been applied to real-world production and service environments, greatly improving the user experience and economic benefits of related products. "Deep learning" was first proposed in 2006 by Hinton, a leading figure in machine learning. Deep learning refers to discovering distributed feature representations of data by combining low-level features to form more abstract high-level representations (attribute categories or features). In recent years, deep learning has been widely applied in signal processing, handling objects including images, speech, and text. While deep learning is booming, the integration and development of recommender systems is also one of the current research hotspots. How to combine the current optimized deep learning schemes with traditional recommender systems is a current research hotspot. The main research contents include: (1) improving the representation and learning ability of the network. Deep learning networks are composed of a large number of neurons, so they have a strong representation learning ability. They can learn from complex data content and obtain the effective factors hidden in it as the feature information needed by the subsequent recommender system. (2) Deeply integrated collaborative filtering scheme. Traditional collaborative filtering schemes recommend related items of interest by performing singular value decomposition (SVD) on the relation matrix. The operation of singular value decomposition can be further simplified to a simple linear neural network operation. Therefore, by introducing nonlinear units through complex neural networks, the network can be strengthened to extract subtle feature information hidden in the content. As neural networks continue to develop, deep learning technologies such as memory networks, attention networks, and convolutional neural networks are widely used in traditional recommender systems to further improve the corresponding recommendation accuracy and performance. (3) Fusion of multiple feature information. Traditional recommender systems often consider a single indicator of the object. However, in actual enterprise-level application systems, in order to improve the accuracy of the model as much as possible, it is often necessary to fuse all the information involved by the user. For users or recommended objects, the richness of information often depends on the thorough mining of the original heterogeneous data. Therefore, it's crucial to fully acquire the different features corresponding to the object and then weight the combinations of these features to achieve optimal performance. Traditional feature fusion often involves hard combinations through manual intervention, which to some extent highlights feature fusion but fails to fully uncover the deeper information hidden behind it. Furthermore, the processing of fusion between different features is often insufficient, and there is a lack of understanding of user interests and long-term evolution patterns, leading to poor performance after feature information fusion. Figure 1The diagram illustrates a common recommender system structure, primarily consisting of a database preprocessing module, a recall layer, a ranking layer, and a re-ranking layer. Different optimization schemes exist for each module. How to integrate deep learning into this system to further improve recommendation performance is a major concern for researchers. In practical applications, many online platforms have become an indispensable part of users' daily lives. On these platforms, users can create their own information and discover relevant information from others, establishing a platform for the interaction between users and information. In this context, recommending relevant information to users places higher demands on both feasibility and accuracy.

[0004] With increasingly clear requirements and performance needs, current recommender systems, while capable of associating users with related objects and making recommendations, still have some shortcomings and limitations. Session-based recommendation, based on a series of previous actions within the same session, predicts the user's next action. A session generally refers to a series of user behaviors that are very close in time. The advantage of session-based recommendation lies in its ability to further mine useful information from anonymous sessions. If the recommender system can track user IDs and has sufficient session data for most users to learn from, then session-based recommendation systems can still function normally. This is because the same user in the same session has consistent characteristics, while the characteristics of the same user in different sessions show some correlation. Therefore, session-based recommendation has strong robustness. In addition to capturing user interests, session-based recommender systems can also utilize user IDs to provide customized recommendation services for users with different functionalities. Although the above recommender systems have achieved the fusion of most mixed information, the following shortcomings and limitations still exist:

[0005] 1) Users' personal preferences and interests are dynamic, not static and unchanging;

[0006] 2) There is a mutual influence between a user's historical behavior and the historical behavior of others in their associated social networks;

[0007] 3) Currently, session-based recommendation networks all suffer from extremely high computational complexity and computational cost, making efficient deployment and application impossible;

[0008] In conclusion, how to achieve an efficient recommendation solution based on user conversations and social networks is a problem that major platforms urgently need to solve. Summary of the Invention

[0009] The purpose of this invention is to provide an efficient recommendation method based on user conversations and social networks, which combines dynamic user conversation content with relevant social network information for efficient recommendation.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions.

[0011] An efficient recommendation method based on user conversations and social networks includes the following steps:

[0012] Construction of Knowledge Graph Embedding Network: Utilizing users' social network information and historical behavior information, a knowledge graph embedding network is constructed to obtain in-depth information related to user sessions and social networks;

[0013] Training and learning of heterogeneous knowledge graph networks: Building heterogeneous knowledge graph networks using knowledge graph embedding networks and learning the hidden information in users and sessions within the heterogeneous knowledge graph networks;

[0014] Embedding of personalized conversation features: Extract user and conversation information from the learned heterogeneous knowledge graph network, use the computing module to calculate the characteristic conversation embedding information extraction for each user, obtain the user's personal preferences and the context feature vector in the conversation, use the personal preferences and context feature vector to perform weighted summation to obtain the final calculation function of user preferences, and combine it with cross-entropy to obtain the final expression of user preferences.

[0015] In this technical solution, firstly, the user's historical behavior and social network information are integrated. At this point, the information source not only contains historical information but also social network information, which makes it easier to mine more information from the dynamically changing social environment, so as to obtain new interests or hobbies of customers in a timely manner and update them promptly.

[0016] Secondly, this technical solution integrates deep learning capabilities for relevant information, particularly the ability to learn hidden information in users and conversations within heterogeneous knowledge graph networks, enabling the discovery of hidden customer-related information and thereby uncovering more information.

[0017] In this technical solution, the final step of embedding personalized conversation features involves integrating the feature vectors of user preferences and conversations, performing weighted calculations, and combining them with cross-entropy to achieve an efficient recommendation scheme based on user conversations and social networks.

[0018] As a further improvement of the present invention, the construction of the knowledge graph embedding network is specifically as follows:

[0019] 1) Establishment of the input layer: The user's social network information and / or historical behavior information outside the user's social network information are formatted and input to form the user's input layer, which mainly includes social network layer information as edge information and user's historical session information as node information.

[0020] 2) Establishment of knowledge graph neural network: For each user, establish a corresponding convolutional neural network. Specifically, combine the user input layer and the convolutional neural network to form a convolutional neural network with at least 2 layers.

[0021] 3) Output of the graph network layer: The graph neural network extracts and processes the user's historical conversation information and social relationship network information to obtain the user's personal preference feature information, and then calculates the user's personal preference evaluation based on the personal preference feature information.

[0022] In this technical solution, through the above processing methods and processes, the user's personal preferences jointly embedded in conversation information and social network information can be maximized. At the same time, this solution is compatible with traditional conversation-based recommendation systems, and can perform offline calculations on user conversation information and social network information to ensure that the computing framework can complete efficient recommendation calculations.

[0023] In this technical solution, the common features output by the convolutional network layer and the node layer are combined to synthesize the feature information of each user. The main purpose is to complete the global classification. Specifically, the user is distinguished by the comprehensive feature information. Here, global classification means that the network extracts the user's social network features and the user's historical conversation features and then performs joint recommendation and calculation. It is similar to using two different measurement methods to comprehensively evaluate and recommend users. In this way, the combined processing of multiple methods achieves better results.

[0024] As a further improvement of the present invention, in step 2) the establishment of the graph knowledge neural network, nonlinear activation is achieved between the two layers of convolutional neural networks and between adjacent convolutional neural networks through the ReLU activation function.

[0025] In this technical solution, the ReLU activation function is used to complete the nonlinear activation. Its purpose is mainly to avoid the overfitting problem in the learning process of the network. This is a nonlinear operation involved in convolutional neural networks, which is mainly to avoid network overfitting and to ensure efficient computation and fast convergence of the model.

[0026] As a further improvement of the present invention, the construction of the heterogeneous knowledge graph network is specifically as follows: based on the different types of edge information in the heterogeneous knowledge graph network, it is determined whether the user has an interactive relationship with the session. If there is an interactive relationship, the edge information is retained; otherwise, it is deleted.

[0027] In this technical solution, the relationship between a session and a user is determined by various types of edge information (including user-user, user-session, session-user, session-session, etc.), thereby filtering data in advance for later deep learning and improving data accuracy.

[0028] As a further improvement of the present invention, the learning of the heterogeneous knowledge graph network specifically includes learning user preferences, remapping user and session relationships, and representing new nodes.

[0029] In this technical solution, by learning user preferences, we can understand the preferences directly expressed by users, and remap the relationship between users and conversations, as well as the expression of new nodes. The purpose is to conduct further learning, find hidden customer preferences and needs, and update customer needs in a timely manner so as to provide customers with the latest recommendation solutions later.

[0030] As a further improvement of the present invention, the learning of user preferences specifically involves: obtaining user preferences through a comprehensive weighted calculation based on information that the user is interested in and conversation information related to the user.

[0031] In this technical solution, the user's attention information directly reflects the user's preferences, while the relevant conversation information indirectly reflects the customer's current preferences and whether there have been any changes from previous preferences through conversation information analysis.

[0032] As a further improvement of the present invention, the remapping of user and session relationships specifically involves: finding users who follow each other, and using information from their following process, employing learning parameters, information evaluation, aggregation, and importance to obtain a linear mapping of the feature vectors corresponding to all nodes and sessions, so that all feature information is mapped to a new feature space.

[0033] In this technical solution, through mutual monitoring, parameter and information evaluation, relevant information can be used for analysis and learning to further explore customer preference-related information and avoid the failure to discover hidden information in a timely manner.

[0034] As a further improvement of the present invention, the expression of the new node is specifically as follows: the expression of the new node information is completed by comprehensively processing the aggregated hierarchical information and the old node information. At the same time, after completing the linear transformation of the corresponding feature information, nonlinear processing is performed through the ReLU activation function to obtain the result after fully learning the user and related session influence.

[0035] In this technical solution, the main purpose of using the ReLU function for nonlinear processing is to avoid overfitting during the network learning process, so that the network can efficiently and quickly complete convergence learning during training, while reducing the phenomenon of overfitting.

[0036] As a further improvement of the present invention, in the embedding of the personalized conversation features, obtaining the user's personal preferences specifically involves: utilizing information about personal preferences S... KG and U KGThe computation module uses a graph neural network to obtain the dynamic personal preferences of the current user in the current session. Per .

[0037] The two additional computing modules in this technical solution are used to calculate the user's personal social relationship network and embedded information in the user's conversation, respectively, to further obtain the user's dynamic preferences.

[0038] As a further improvement of the present invention, the embedding of personalized session features also includes the processing of context feature vectors in the session, specifically: by acquiring input and output information, the feature vector relationship expressing the context is obtained, and by merging context feature information and combining the feature vector relationship, the context feature vector in the session is obtained.

[0039] This technical solution processes the contextual feature vectors within a conversation to depict the user's migration and change trends during the conversation, providing a richer description of the user's changes. This allows for better understanding of how the feature vectors reflect user preferences and, specifically, reduces overfitting. Attached Figure Description

[0040] Figure 1 A flowchart of the recommended method in the prior art;

[0041] Figure 2 A flowchart illustrating an efficient recommendation method based on user conversations and social networks provided by this invention;

[0042] Figure 3 A flowchart illustrating the construction of a knowledge graph embedding network provided by this invention;

[0043] Figure 4 The flowchart in Embodiment 3 provided by the present invention;

[0044] Figure 5 A statistical graph illustrating the features of the dataset in Embodiment 4 of this invention;

[0045] Figure 6 A statistical chart showing the recommended accuracy in Embodiment 4 of the present invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent changes or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0047] Example 1

[0048] Specifically, refer to the appendix Figure 2 As shown in the figure, an efficient recommendation method based on user sessions and social networks in this embodiment includes the following steps:

[0049] Construction of Knowledge Graph Embedding Network: Utilizing users' social network information and historical behavior information, a knowledge graph embedding network is constructed to obtain in-depth information related to user sessions and social networks;

[0050] Training and learning of heterogeneous knowledge graph networks: Building heterogeneous knowledge graph networks using knowledge graph embedding networks and learning the hidden information in users and sessions within the heterogeneous knowledge graph networks;

[0051] Embedding of personalized conversation features: Extract user and conversation information from the learned heterogeneous knowledge graph network, use the computing module to calculate the characteristic conversation embedding information extraction for each user, obtain the user's personal preferences and the context feature vector in the conversation, use the personal preferences and context feature vector to perform weighted summation to obtain the final calculation function of user preferences, and combine it with cross-entropy to obtain the final expression of user preferences.

[0052] In this embodiment, the user's historical behavior and social network information are first integrated. At this point, the information source not only contains historical information but also social network information, which makes it easier to mine more information from the dynamically changing social environment, so as to obtain new interests or hobbies of customers in a timely manner and update them promptly.

[0053] In this embodiment, secondly, it has the ability to learn complex heterogeneous fusion information with deep neural networks, especially the ability to learn the information hidden in users and sessions in heterogeneous knowledge graph networks, so that the hidden association information of customers can be mined, and more information can be found.

[0054] In this embodiment, finally, for the embedding of personalized conversation features, the feature vectors of user preferences and conversations are fused together, weighted, and combined with cross-entropy, which can realize an efficient recommendation scheme based on user conversations and social networks.

[0055] Example 2

[0056] In this embodiment, the construction of a knowledge graph embedding network is introduced.

[0057] Furthermore, the construction of the knowledge graph embedding network specifically involves:

[0058] 1) Establishment of the input layer: The user's social network information and / or historical behavior information outside the user's social network information are formatted and input to form the user's input layer, which mainly includes social network layer information as edge information and user's historical session information as node information.

[0059] 2) Establishment of knowledge graph neural network: For each user, establish a corresponding convolutional neural network. Specifically, combine the user input layer and the convolutional neural network to form a convolutional neural network with at least 2 layers.

[0060] 3) Output of the graph network layer: The graph neural network extracts and processes the user's historical conversation information and social relationship network information to obtain the user's personal preference feature information, and then calculates the user's personal preference evaluation based on this.

[0061] At this point, for example, the features of a convolutional neural network are A+B, and the features of the node layer will not be the same as those of the social network layer, thus avoiding information duplication.

[0062] In this embodiment, through the above processing method and process, the user's personal preferences jointly embedded in the conversation information and social network information can be maximized. At the same time, the solution is compatible with traditional conversation-based recommendation systems and can perform offline calculations on the user's conversation information and social network information, enabling the computing framework to complete efficient recommendation calculations.

[0063] In this embodiment, the common features output from the convolutional network layer and the node layer are combined to synthesize the feature information of each user. The main purpose is to complete the global classification. Specifically, the user is distinguished by the comprehensive feature information. Here, global classification means that the network extracts the user's social network features and the user's historical conversation features and then performs joint recommendation and analysis. It is similar to using two different measurement methods to comprehensively evaluate and recommend users. In this way, the combined processing of multiple methods achieves better results.

[0064] A heterogeneous knowledge graph is constructed from social networks and all user historical behavior. Then, the heterogeneous knowledge graph network is used to learn fused knowledge about users and items, primarily including social relationships, user-session interactions, and transitions. Given a user and their current session, relevant user and session representations are retrieved and fed into the PSR model. Because user and session representations are socially aware—meaning their interests can be influenced by the behavior of others—the SSR model can leverage knowledge from social networks to provide corresponding, accurate recommendations.

[0065] In step 2) of establishing the convolutional neural network, nonlinear activation is achieved between two layers of convolutional neural networks and between adjacent convolutional neural networks through the ReLU activation function.

[0066] In this embodiment, the ReLU activation function is used to perform non-linear activation. Its purpose is mainly to avoid overfitting during the learning process of the network. This is a non-linear operation involved in convolutional neural networks, mainly to avoid network overfitting.

[0067] Furthermore, the construction of the heterogeneous knowledge graph network specifically involves: determining whether a user has an interactive relationship with the session based on different types of edge information in the heterogeneous knowledge graph network; if an interactive relationship exists, the edge information is retained; otherwise, it is deleted.

[0068] In this embodiment, the relationship between a session and a user is determined by various types of edge information (including user-user, user-session, session-user, session-session, etc.), thereby filtering the data in advance for later deep learning and improving the accuracy of the data.

[0069] Furthermore, the learning of the heterogeneous knowledge graph network specifically includes learning user preferences, remapping user and session relationships, and representing new nodes.

[0070] In this embodiment, by learning user preferences, we can understand the preferences directly expressed by users, and the remapping of user and conversation relationships, as well as the expression of new nodes, are aimed at further learning to find hidden customer preferences and needs, and to update customer needs in a timely manner so as to provide the latest recommendation solutions for customers in the future.

[0071] Furthermore, the learning of user preferences specifically involves: calculating user preferences by comprehensively weighting the information that the user is interested in and the conversation information related to the user.

[0072] In this embodiment, the user's attention information directly reflects the user's preferences, while the relevant conversation information indirectly reflects the customer's current preferences and whether there have been any changes from previous preferences through conversation information analysis.

[0073] Specifically, the remapping of user and session relationships involves: finding users who follow each other, and using information from their following process, employing learning parameters, information evaluation, aggregation, and importance to obtain a linear mapping of the feature vectors corresponding to all nodes and sessions, so that all feature information is mapped to a new feature space.

[0074] In this embodiment, through mutual following, parameter and information evaluation, etc., relevant information can be used for analysis and learning to further explore customer preference-related information and avoid the failure to discover hidden information in a timely manner.

[0075] Furthermore, the expression of the new node is specifically as follows: the expression of the new node information is completed by comprehensively processing the aggregated hierarchical information and the old node information. At the same time, after completing the linear transformation of the corresponding feature information, nonlinear processing is performed through the ReLU activation function to obtain the result after fully learning the user and related session influences.

[0076] In this embodiment, the main purpose of using the ReLU function for nonlinear processing is to avoid overfitting during the network learning process, so that the network can quickly complete convergence learning during training, while reducing the phenomenon of overfitting.

[0077] Specifically, in the embedding of the personalized conversation features, obtaining the user's personal preferences specifically involves: utilizing information about personal preferences S... KG and U KG The computation module uses a graph neural network to obtain the dynamic personal preferences of the current user in the current session. Per .

[0078] The two additional computing modules in this embodiment are used to calculate the user's personal social relationship network and embedded information in the user's conversation, respectively, to further obtain the user's dynamic preferences.

[0079] In this embodiment, the embedding of personalized session features also includes the processing of context feature vectors in the session. Specifically, by acquiring input and output information, the feature vector relationship expressing the context is obtained. By merging context feature information and combining the feature vector relationship, the context feature vector in the session is obtained.

[0080] Furthermore, by processing the contextual feature vectors within the conversation, we can characterize the user's shifting trends during the conversation, providing a richer description of the user's changes. This allows us to better reflect how the feature vectors reflect customer preferences and, specifically, reduce overfitting.

[0081] Example 3

[0082] See attached document Figure 2-6As shown, this invention provides an efficient recommendation method based on user conversations and social networks, utilizing dynamic user conversation content and related social network information for joint and efficient recommendations. Platforms such as Twitter, YouTube, and Zhihu allow users to create and share information while simultaneously consuming information created by other users. Therefore, recommendation systems deployed on these platforms are crucial for displaying information of interest to users and improving long-term user engagement. However, in practical system applications, users' personal preferences and tendencies are dynamic; users may show significant interest in certain things and information within a specific period. Furthermore, in real social life, there are certain mutual influences between users' historical behaviors, and these influences also change dynamically over a certain period. Therefore, how to model conversation-level user behavior sequences while simultaneously conducting joint analysis of users' dynamic social relationships is a challenge. The influence on the current user is characterized by the impact of subset conversations of other users on the current user. The entire system is constructed using recurrent neural networks (RNNs) and graph attention networks to model user behavior and the influence of other users on the current user. This variant of a conversation-based recommendation system combined with user ID can be called Personalized Conversation-Based Recommendation (PSR). Because the aforementioned phenomena exist in many online services such as e-commerce and video sharing websites, session-based recommendation systems have significant advantages and practical value, and therefore have received much attention in recent years.

[0083] like Figure 4 The proposed efficient recommendation solution is illustrated below. First, a heterogeneous knowledge graph is built from social networks and all user historical behaviors. Then, a heterogeneous graph network is used to learn fused knowledge of users and items, mainly including social relationships, user-session interactions, and transitions. Given a user and their current session, relevant user and session representations are retrieved and fed into the PSR model.

[0084] Because user and session representations are socially aware, the SSR model can leverage knowledge from social networks to provide accurate recommendations. This adapts a PSR model to an SSR model. Since SSR is an upgraded version of PSR, it takes into greater consideration the influence of other session information on the user.

[0085] This computational framework is effective because the model can pre-compute user and session representations for accurate recommendations. Therefore, during the recommendation process, the SSR model only needs to process the current session, making it as efficient as the original PSR model. This scheme is as follows... Figure 4 As shown, it mainly consists of the following three parts: 1) Knowledge graph embedding module; 2) Personalized conversation recommendation module; 3) Prediction module; The following will introduce the three modules mentioned above in detail.

[0086] First, the knowledge graph embedding module mainly utilizes heterogeneous knowledge graph networks, such as... Figure 3 As shown, this module completes the deep feature representation of users and sessions. It mainly consists of two task modules, as follows:

[0087] The first task module is to establish a heterogeneous knowledge graph network K, whose main components are user historical behavior D and social relationship network S. Specifically, D and S represent the original user historical behavior and the information contained in the social relationship network, respectively. The information is synthesized and coarse-grained through a series of processing steps in the heterogeneous network.

[0088] The second task module involves learning the hidden information in users and sessions within K. K = (V, E, A, R, φ, ψ) represents the corresponding heterogeneous graph knowledge network. V represents the user and historical behavior entries contained in the social network and user history behavior; A represents the set of different node types; R represents the set of different edge types; φ represents the mapping process between each node and its corresponding type; and ψ represents the mapping relationship between each edge and its type. E represents different edge information. The edge types include: user-user, user-session, session-user, and session-session. The edge information between a user and a session represents the interaction relationship between the user and that session, indicating that the user has participated in or is currently participating in that session. The edge information between sessions represents whether there is a transitional relationship or correlation between the two sessions. This high scalability allows for the flexible addition of different user and corresponding session information to the network, while also enabling the filtering and preprocessing of node and session information to eliminate those that do not meet the requirements.

[0089] After completing the above settings, the user node and session information need to be input into the corresponding heterogeneous graph neural network for sufficient learning. This neural network consists of L layers, with different layers interconnected through weights, H... l [v] represents node v corresponding to layer l of the neural network (this node can be the user or the corresponding session content). In this L-layer neural network, each layer is connected to the previous layer through neurons. The connection here is the convolutional layer connection, which by default is the mathematical relationship y = wx + b, and then it is activated by an activation function.

[0090] In the corresponding neural network layer, Layer 1, two parts were calculated and analyzed: first, the influence of the relationship network, and second, user preferences. For user ν, if user u follows another user in the relationship network, then this user's following behavior can be expressed as:

[0091]

[0092] Where W and b represent the learning parameters related to the network. This represents the mutual attention strength between two users. The formula primarily expresses the connection strength between two different users, i.e., the mutual attention strength. How to influence users through conversation can be illustrated by the following expression:

[0093]

[0094] Here, W and b represent the learning parameters related to the network. Note that here, because we are considering user preferences, which primarily refer to different items (or equivalently, how items influence users), we only consider how messages from items to users affect users. Therefore, we do not need to consider the impact of messages from users to items. To comprehensively evaluate the impact of user preferences and conversational social networks, we utilize a hierarchical aggregation scheme to handle the corresponding information evaluation.

[0095]

[0096]

[0097] in This represents the various neighbor information of user u within the entire social network and historical behavior network information. The `Aggregate` operation represents hierarchical aggregation, achieving the acquisition of high-level information by hierarchically aggregating the input user social network information and user historical behavior information. By merging the above transmitted information, the second-level aggregation result can be obtained:

[0098]

[0099] The output of the above formula This represents high-level feature information calculated from a user's social network and related historical behavior information. To further address the issues arising from the imbalance of user node information (which includes the user's social network information and historical behavior information; imbalance mainly refers to a few extreme cases, such as a user who clicked and generated a lot of historical behavior information but only has one followed friend), an attention-focusing scheme is introduced to handle the trust issue in the system.

[0100]

[0101] Here, ψ(e) represents different edge types. The importance of information transmitted through an edge e = (v, u) can be calculated, and then measured by a score. The specific calculation process is as follows:

[0102]

[0103] This corresponds to the network's output, used to evaluate and measure the importance of the information. It can be normalized using the Softmax function.

[0104]

[0105] The influence of all the neighbor information on a user can be calculated as follows:

[0106]

[0107] To further aggregate and calculate the influence of all neighboring nodes on user u, we calculate all the information transmitted through user υ, as shown in the following formula. The main function of this formula is to further express the influence of different nodes on the current user through user υ, thus completing the linear transformation of the feature space. Therefore, the influence of any user on the current user can be realized through the above scheme:

[0108]

[0109] Finally, the new node information is expressed by comprehensively processing the aggregated hierarchical information and the old node information. At the same time, after completing the linear transformation of the corresponding feature information, nonlinear processing is performed by the ReLU activation function.

[0110]

[0111] Among them, H l [u] represents the lth th The expression of each layer neuron at each different node u simultaneously incorporates a user's personal preferences and social influences; therefore, the final output H of the network... L This is the result after fully learning about the user and related conversational influences. The above formula uses a general non-linear function to avoid overfitting during the learning process; ReLU is a common activation function.

[0112] Secondly, there is the personalized conversation embedding module, the second part of the entire recommendation system. This module takes as input the higher-order representations of users and conversations processed by the previous module. After processing the relevant input information, it obtains embedding feature information strongly correlated with user preferences and conversation content. The first step of this module is to extract relevant user and conversation information from the previous module through embedding information lookup. The second step is to calculate personalized conversation embedding information extraction based on the user and conversation feature information obtained in the first step, as shown below:

[0113] s Per =Θ(u KG ,S KG )

[0114] Here, Θ represents the extraction of personalized recommendation information through the calculation module. For recommendation schemes that do not consider the combined influence of user and conversation, S can be obtained by adding a calculation component regarding personal preferences during the calculation process. KG and U KG These represent the representation of user historical behavior information and user social network information in the graph embedding network, respectively. The main function of this module is to obtain the dynamic personal preferences s of the current user in the current session using a graph neural network. Per The process mainly consists of three parts. First, a weighted directed graph neural network G = (V, E) is used to evaluate user u's personal preferences in the current session. Second, it needs to fully learn the contextual feature vector for session S, which can be expressed by the following formula:

[0115] x i =S KG [t]‖u KG

[0116] To fully describe the transmission of information, we use... and To express the input and output relationship for a node, we can further express the contextual feature vector relationship by obtaining information from the two inputs and outputs mentioned above, as follows:

[0117]

[0118]

[0119]

[0120] Among them W in With W out These correspond to the network's learning parameters, a and b, respectively. iThis represents the hierarchical aggregation information corresponding to node i. Merging the obtained contextual feature information, including the initialization feature vector, yields:

[0121]

[0122] r i =σ(W r (a i ||x i )+b r )

[0123]

[0124] Where W and b represent the learning parameters during the model training process, a i This represents information about neighboring nodes, while x i This represents the initialized feature vector. h i This represents the context feature vector of node i. Ultimately, we can obtain the importance score of node i:

[0125] ò i =p T σ(W(h i ||h last t ||u KG )+r)

[0126] Where h last t This represents the context feature vector corresponding to the last entry in S. The calculation based on user personalized preferences is a weighted sum of all context feature vectors. The calculation of the weights involves non-linear processing using the Softmax function.

[0127]

[0128] in This represents the output of the fully connected layer of the network, which is converted into the corresponding normalized probability value through the Softmax function. This completes the calculation of the three main modules. In the actual training and computation process, cross-entropy is used as the loss function for the entire network's training and convergence. The calculation of cross-entropy is shown below:

[0129]

[0130] The core innovations of this patent are mainly reflected in the following aspects:

[0131] 1) By feeding users’ personal historical behavior trajectory and social networks as unified input into a heterogeneous knowledge graph network, knowledge embedding information is obtained from users’ social relationship network and historical conversations through heterogeneous graph neural networks. At the same time, this network can process heterogeneous knowledge graph networks with semi-static input information.

[0132] 2) All the different modules involved in the system architecture can be calculated in advance in actual applications, so there is no complex amount of calculation in actual application deployment;

[0133] 3) This architecture has excellent compatibility, and can integrate conversation-based recommendation systems by incorporating node information from social networks, thereby further improving system performance;

[0134] Example 4

[0135] To further evaluate the performance of the efficient recommendation scheme based on user conversations and social networks, the publicly available datasets Gowalla and Foursquare were selected. Gowalla primarily contains social network and regular user conversation information, while Foursquare includes user social network information obtained from other video sharing platforms in addition to user conversation information. To further adapt to efficient network training, all datasets were split daily. The acquired datasets were then divided into test and training sets in a 3:7 ratio for model training and testing. Statistical information related to the datasets is as follows: Figure 5-6 As shown. Figure 5 The image shows the different content contained in two different databases, which mainly contain user clicks, sessions, users, items, and corresponding social relationships. This database is also a typical collection of user surrounding information in a recommendation system. Figure 6 The figure shows the recommendation accuracy of two different databases under different recommendation networks. As can be seen from the figure, different networks have different recommendation accuracies for two different datasets. The recommendation scheme proposed in this paper shows excellent performance compared with other schemes on both different datasets.

[0136] To obtain optimal model training parameters, grid parameter scanning and validation set analysis are used to acquire the training parameters that best support model performance. The hyperparameters involved mainly include embedding block size: {32, 64, 96, 128}, and learning rate: {10...} -4 10 -3 10 -2 10 -1The number of layers in the graph neural network is {1, 2, 3}. During model training, the batch size is set to 128, and the optimizer Adam is used to continuously adjust the relevant parameters in the entire network to optimize the steps. In practical applications, to further compare the performance differences of different recommendation schemes, representative recommendation schemes were selected and evaluated on a unified dataset. The accuracy of the recommendation model was evaluated by measuring whether the user's predictions matched the actual results in the first 10 predictions. Through comparative analysis of the results, it can be seen that traditional conversation-based recommendation systems often fail to fully mine the feature information in social networks and related conversation information. It can be seen that the more fully a new model integrates and processes the information involved, the higher the recommendation accuracy of the entire system. Compared with traditional recommendation schemes, the proposed scheme computes contextualized conversation information embeddings to understand the user's attention to each item and changes in personal preferences. Therefore, the network can learn better methods to more accurately capture user dynamic information and personalized interests. Therefore, the efficient recommendation scheme based on user conversations and social networks can be deployed within the platform, integrating and processing conversation information from social, gaming, and transaction sections within the platform. After inputting the corresponding network model, we obtain the prediction of the user's next stage of the session.

[0137] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

[0138] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0139] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A recommendation method based on user conversations and social networks, characterized in that, Includes the following steps: Construction of Knowledge Graph Embedding Network: Utilizing users' social network information and historical behavior information, a knowledge graph embedding network is constructed to obtain in-depth information related to user sessions and social networks; Training and learning of heterogeneous knowledge graph networks: Building heterogeneous knowledge graph networks using knowledge graph embedding networks and learning the hidden information in users and sessions within the heterogeneous knowledge graph networks; The construction of the heterogeneous knowledge graph network is specifically as follows: based on the different types of edge information in the heterogeneous knowledge graph network, it is determined whether the user has an interaction relationship with the session. If there is an interaction relationship, the edge information is retained; otherwise, it is deleted. The learning of the heterogeneous knowledge graph network specifically includes learning user preferences, remapping user and session relationships, and representing new nodes. Edge types include: user-user, user-session, session-user, and session-session. The edge information between a user and a session represents the interaction relationship between the user and the session, indicating that the user has participated in the session operation in the past or is currently participating. The edge information between sessions represents whether there is a transition relationship or correlation between the two sessions. User history includes items or services purchased, selected, and rated by the user, as well as user click information; The remapping of user and session relationships specifically involves: finding users who follow each other, and based on the intensity of attention between users, using network-related learning parameters, value assessments of different edge types to form information evaluation, and importance formed by hierarchical aggregation of input user social network information and user historical behavior information, as well as evaluation and measurement of the importance of output information, to obtain a linear mapping of the feature vectors corresponding to all nodes and sessions, so that all feature information is mapped to a new feature space. Embedding of personalized conversation features: Extract user and conversation information from the learned heterogeneous knowledge graph network, use the computing module to calculate the characteristic conversation embedding information extraction for each user, obtain the user's personal preferences and the context feature vector in the conversation, use the personal preferences and context feature vector to perform weighted summation to obtain the final calculation function of user preferences, and obtain the final expression of user preferences.

2. The recommendation method based on user conversations and social networks according to claim 1, characterized in that, The construction of the knowledge graph embedding network is specifically as follows: The establishment of the input layer: The user's social network information and / or historical behavior information outside the user's social network information are formatted and input to form the user's input layer, which mainly includes social network layer information as edge information and user historical session information as node information. The establishment of knowledge graph neural networks: For each user, a corresponding convolutional neural network is established. Specifically, the user input layer and the convolutional neural network are combined to form a convolutional neural network with at least two layers. The output of the graph network layer is as follows: The graph neural network extracts and processes the user's historical conversation information and social relationship network information to obtain the user's personal preference feature information, and then calculates the user's personal preference evaluation based on the personal preference feature information.

3. The recommendation method based on user conversations and social networks according to claim 1, characterized in that, The learning of user preferences specifically involves: calculating user preferences by comprehensively weighting the information that the user pays attention to and the conversation information related to the user.

4. The recommendation method based on user conversations and social networks according to claim 1, characterized in that, The embedding of personalized session features also includes the processing of context feature vectors in the session. Specifically, by acquiring input and output information, the feature vector relationship expressing the context is obtained. By merging context feature information and combining the feature vector relationship, the context feature vector in the session is obtained.

Images