Recommendation system based on enhanced delivery operator graph clustering method

Abstract

The invention relates to the technical field of deep learning, and provides a recommendation system based on an enhanced delivery operator graph clustering method, and the system comprises a data processing module which is used for preprocessing original data, and obtaining a target numerical value vector inputted into an automatic encoder and graph data inputted into a graph convolutional network model; the model construction module is used for incorporating the representation learned by the automatic encoder into a graph convolutional network model through an enhanced delivery operator based on the target numerical value vector and the graph data to obtain a model combining the automatic encoder and the graph convolutional network; and the data recommendation module is used for carrying out theme recommendation by utilizing a model combining the automatic encoder and the graph convolutional network. According to the method, the weight is dynamically adjusted through the attention coefficient, and the enhanced delivery operator is designed to bring the representation learned by the automatic encoder into the corresponding graph convolutional network layer, so that a model in which the automatic encoder and the graph convolutional network are combined can be obtained, the discrimination capability of the model for feature importance can be enhanced, and the training quality of the model can be ensured.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a recommendation system based on an enhanced delivery operator graph clustering method. Background Art

[0002] In the application scenarios of personalized recommendation, social network and knowledge graph recommendation systems, the recommendation system can generate a personalized recommendation table for each user based on the user's historical behavior and preferences; it can recommend potential friends or content of interest through the relationship graph between users, and recommend related entities or information through the graph structure of entities and relationships; most of the existing recommendation systems face the problem that the interaction data between users and items is very sparse, that is, data sparsity, which makes it difficult for the recommendation system to accurately model user preferences; in the algorithms for graph data, especially graph clustering algorithms, although the effect is significant, the computational complexity is high; the recommendation system requires a large amount of user data for training, which will inevitably involve user privacy, and this issue is the most critical and important, and is the decisive factor in whether the recommendation system can be widely promoted and applied.

[0003] With the development of deep learning technology, deep clustering methods have become a trend, and the effective combination of deep embedding clustering analysis and autoencoders has become the mainstream of research. Deep clustering methods have achieved state-of-the-art development, but they simply focus on the attribute characteristics of the data itself and ignore the structure embedded in the data.

[0004] In order to reveal the structural information of data in deep clustering, the ability of graph convolutional networks to encode graph structures and node attributes has received in-depth attention. Kipf et al. proposed graph autoencoders (GAE) and variational graph autoencoders based on the AE framework to learn graph structural features. Pan et al. proposed an adversarial regularized graph autoencoder (ARGA) by introducing an adversarial regularizer. In addition, the structural information of the data indicates the potential similarity between samples. Graph data structures often have high-order structures in real-world networks. In order to effectively combine the structural information of the data with the learned representation of the autoencoder, BO et al. proposed a structured deep clustering network (SDCN), which aims to improve the representation of AE specifications into high-order graph regularization constraints through graph convolutional networks. The two influence each other. However, these existing works take it for granted that topological graph features and node attribute features have the same importance, make some predetermined assumptions about data distribution, and even directly ignore the noise in the graph structure.

[0005] With the introduction of the attention mechanism, graph clustering algorithms combined with the attention mechanism have also become more abundant. For example, Peng et al. proposed an attention-driven graph clustering network, which combines heterogeneous features with scale feature fusion mechanisms to reveal the multi-scale characteristics of the structure. Although deep clustering based on AE and graph convolutional networks has achieved great success, the representation learning of clustering tasks is still challenging.

[0006] Thanks to the powerful representation ability of deep neural networks, many deep clustering methods have been discovered in recent years and have achieved remarkable performance advantages. The idea of ​​AE has a profound enlightenment effect on clustering methods based on graph convolutional networks. Obviously, the ability of graph convolutional networks to encode graph structures and graph nodes is exactly what is needed to obtain data structure information. Starting from [1] using node features to obtain embedding based on graph convolutional networks, the application technology of graph convolutional networks in graph clustering has become more and more mature, but there is still a lot of room for improvement in clustering performance. SDCN integrates structural information into deep clustering. This is the first time that structural information has been clearly introduced into deep clustering. Subsequently, a series of improved methods based on this model emerged, such as Huang et al. proposed a Structured Deep Graph Clustering Network Based on ConsistencyConstraint (CC-DGC). These models all improve clustering performance through different learning representation methods.

[0007] Although graph convolutional networks have achieved remarkable achievements in different tasks of deep learning, some of their limitations have gradually been revealed. Graph convolutional networks use an equal weight strategy when processing different neighbor nodes of the same order, which may cause the neglect of some important information near the node. Based on this, attention-based graph networks have been proposed, and the graph attention network (GAT) has been proposed. It uses the attention mechanism to allow each node to be assigned different weights according to the characteristics of its neighboring nodes, thereby breaking the limitation of uneven weight distribution in graph convolutional networks. Subsequently, Peng et al. proposed a clustering method with a dual self-supervision mechanism (deep attention-guided graph clustering with dual self-supervision, DAGC). This model can learn information feature embedding based on the attention mechanism and design a dual self-supervision mechanism. These methods still have problems such as complex calculations and excessive structural information redundancy when aggregating multi-scale features.

[0008] The existing graph clustering method combining AE with graph convolutional network lacks an efficient and flexible fusion strategy, and ignores the reasonable importance distribution of node feature attributes and structural attributes; at the same time, the transfer operator of the existing method has certain limitations when combining the two types of node information.

[0009] Therefore, it is necessary to provide a recommendation system based on an enhanced delivery operator graph clustering method. Summary of the invention

[0010] The present invention provides a recommendation system based on an enhanced delivery operator graph clustering method, which dynamically adjusts weights through attention coefficients, and designs an enhanced delivery operator to more effectively incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer, unifying two different deep neural network structures, and obtaining a model combining the autoencoder and the graph convolutional network, which can enhance the model's ability to discriminate the importance of features. At the same time, the attention coefficient is introduced to adjust the weight ratio between the encoder and the graph convolutional network, which can ensure that the model automatically learns the optimal feature combination during the training process, rather than relying on a pre-assumed fixed ratio.

[0011] The present invention provides a recommendation system based on an enhanced delivery operator graph clustering method, comprising:

[0012] The data processing module is used to preprocess the raw data to obtain the target numerical vector for inputting the autoencoder and the graph data for inputting the graph convolutional network model;

[0013] The model building module incorporates the representation learned by the autoencoder into the graph convolutional network model through the enhanced delivery operator based on the target numerical vector and graph data, and obtains a model combining the autoencoder and the graph convolutional network;

[0014] The data recommendation module is used to recommend topics using a model that combines an autoencoder with a graph convolutional network.

[0015] Furthermore, the data processing module includes:

[0016] Remove stop words, punctuation marks and special characters from the text in the original data to obtain processed text;

[0017] Decompose the processed text into words or phrases, and use the bag-of-words model to convert the processed text into a target numerical vector;

[0018] Based on the original data, a KNN graph is constructed using the cosine similarity metric as the distance metric to convert the general data set into graph data and obtain the graph data in the input graph convolutional network model.

[0019] Furthermore, the model building module includes:

[0020] Input the target numerical vector into the autoencoder;

[0021] Input graph data into the graph convolutional network model;

[0022] The representation learned by the autoencoder is incorporated into the graph convolutional network model through the enhanced delivery operator to obtain a model combining the autoencoder and the graph convolutional network.

[0023] Furthermore, the target numerical vector is input into the autoencoder, including:

[0024] Extract the intrinsic features of the target numerical vector based on the traditional autoencoder;

[0025] The decoder is used to reconstruct the low-dimensional text representation of the input data through multiple fully connected layers. First, the intrinsic features of the data are extracted based on the traditional autoencoder. If there are L layers in the autoencoder, then Indicates Layer, the encoder is at The representation learned by the layer for:

[0026]

[0027] Among them, φ represents the Leaky ReLU activation function of the fully connected layer, and They are The weight matrix and bias of the layer encoder;

[0028] according to The original data X can be expressed as H (0) , the original data X can be compressed into a lower-dimensional data representation as input; the decoder reconstructs the input data through multiple fully connected layers, and the reconstructed data is represented as follows:

[0029]

[0030] in, Represents the reconstruction data, and Respectively represent The weight matrix and bias of the layer decoder, Representative The reconstructed data of the original data X of the layer input

[0031] The weights and biases are adjusted by the autoencoder through the back-propagation algorithm to minimize the loss function.

[0032] Furthermore, the representation learned by the autoencoder is incorporated into the graph convolutional network through enhanced delivery operators, including:

[0033] For each layer of information features in the deep neural network model, the weights are dynamically adjusted through the attention coefficient, and an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer;

[0034] The self-supervision module is introduced to unify the autoencoder and graph convolutional network layer in one framework to achieve the update of the deep neural network model.

[0035] Furthermore, the weights are dynamically adjusted through the attention coefficient, including:

[0036] Input graph adjacency matrix A and feature matrix X, for the weight matrix W of the neural network, the graph convolutional network The representation learned by the layer for:

[0037]

[0038] in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix, is the number of layers of the graph convolutional network, It is Feature matrix of layer nodes;

[0039] Introducing dynamic adjustment of attention coefficient and The weight ratio of is used to form an adaptive fusion strategy, which effectively combines the output of the autoencoder with the output of the graph convolutional network and The heterogeneity of the layers represents bonding.

[0040] Furthermore, the attention coefficient is dynamically adjusted. and The weight ratio of , to form an adaptive fusion strategy, also includes: learning the attention coefficient, specifically:

[0041] A shared linear transformation is performed on the feature representation and parameterized by a weight matrix W; the weight matrix W is calculated by a single-layer feedforward neural network to connect and And the weight vector α, then The attention score of a layer is:

[0042]

[0043] Among them, leakyReLU is an activation function. Score for attention;

[0044] The softmax function is used to normalize the attention scores to ensure that the sum of all attention coefficients is 1; the attention weight coefficient is expressed as follows:

[0045]

[0046] Among them, K represents any layer in the L layer, exp is the mathematical formula of the softmax function, and e (k) Respectively The attention scores of the layer and the k-th layer.

[0047] Furthermore, an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer, including: after obtaining the importance ratio of each feature, the delivery operator is designed to integrate the data attributes of the autoencoder into the feature expression in the graph convolutional network. The delivery operator is expressed as:

[0048]

[0049] in, It is the graph convolutional network in The data representation learned by the layer, In the autoencoder The reconstructed data representation learned by the layer, For the The attention weight coefficient of the layer;

[0050] The fused features are output through the graph convolutional network layer, i.e. The input of the layer is but The output of the layer is:

[0051]

[0052] in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix;

[0053] The enhanced delivery operator works L times in the entire model and is represented by the softmax function in the last layer. The output is:

[0054]

[0055] in, For the graph convolutional network in The data representation learned by the layer, In the autoencoder The reconstructed data representation learned by the layer, W is a learnable parameter. When the network training is good, the predicted cluster label can be directly obtained through Z.

[0056] Furthermore, the model combining the autoencoder and the graph convolutional network is used to recommend topics, and the topic recommendation effect is evaluated and optimized, specifically:

[0057] Obtain multiple similar models combining autoencoders and graph convolutional networks for horizontal comparison;

[0058] A model combining multiple similar autoencoders and graph convolutional networks is used to recommend topics, and multiple first topic recommendation results are obtained;

[0059] The topic recommendation effect and the plurality of first topic recommendation results are used as data sources, respectively input into a preset application scenario simulation model, simulate the application scenario, and obtain a simulation result of the application scenario;

[0060] According to the evaluation rules of the simulation results, the topic recommendation effect and the multiple first topic recommendation results are rated and scored for application effect respectively; and an average score value is obtained according to the rating and scoring results;

[0061] If the rating score of the topic recommendation effect is less than the average score, the feasibility of model optimization is judged and analyzed for the model combining the autoencoder and the graph convolutional network. If model optimization feasibility exists, the model combining the autoencoder and the graph convolutional network is optimized. If model optimization feasibility does not exist, the topic recommendation performance of the model combining the autoencoder and the graph convolutional network is calibrated for the optimal recommendation performance so as to better apply it to different scenarios.

[0062] Furthermore, it also includes adaptive matching of different recommendation application scenarios based on the topic recommendation performance of the model combining the autoencoder and the graph convolutional network, specifically:

[0063] Obtain a model combining an autoencoder and a graph convolutional network, and obtain its calibrated optimal recommendation performance;

[0064] Obtain recommended application scenarios and recommended requirements for the recommended application scenarios;

[0065] According to the recommendation requirements, the personalized features of the recommended application scenarios are obtained;

[0066] Based on the set knowledge graph network structure, the corresponding optimal recommendation performance is obtained by utilizing the subject correspondence between the personalized features in the knowledge graph network structure and the optimal recommendation performance, and a model combining the autoencoder and the graph convolutional network corresponding to the optimal recommendation performance is obtained.

[0067] Compared with the prior art, the present invention has the following advantages and beneficial effects: dynamically adjusting weights through attention coefficients, and designing an enhanced delivery operator to more effectively incorporate the representations learned by the autoencoder into the corresponding graph convolutional network layer, unifying two different deep neural network structures, and obtaining a model combining the autoencoder and the graph convolutional network, which can enhance the model's ability to discriminate the importance of features. At the same time, introducing the attention coefficient to adjust the weight ratio between the encoder and the graph convolutional network can ensure that the model automatically learns the optimal feature combination during the training process, rather than relying on a pre-assumed fixed ratio.

[0068] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0069] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 It is a schematic diagram of the recommendation system structure based on the enhanced delivery operator graph clustering method;

[0072] Figure 2 It is a structural diagram of the data processing module;

[0073] Figure 3 Build a schematic diagram of the module structure for the model. DETAILED DESCRIPTION

[0074] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0075] The present invention provides a recommendation system based on an enhanced delivery operator graph clustering method, such as Figure 1 As shown, including:

[0076] The data processing module is used to preprocess the raw data to obtain the target numerical vector for inputting the autoencoder and the graph data for inputting the graph convolutional network model;

[0077] The model building module incorporates the representation learned by the autoencoder into the graph convolutional network model through the enhanced delivery operator based on the target numerical vector and graph data, and obtains a model combining the autoencoder and the graph convolutional network;

[0078] The data recommendation module is used to recommend topics using a model that combines an autoencoder with a graph convolutional network.

[0079] The working principle of the above technical solution is: in order to realize the recommendation system based on the enhanced delivery operator graph clustering method, the present invention proposes a data processing module for preprocessing the original data to obtain the target numerical vector of the input autoencoder and the graph data in the input graph convolutional network model; a model construction module is proposed, based on the target numerical vector and the graph data, the representation learned by the autoencoder is incorporated into the graph convolutional network model through the enhanced delivery operator, and an enhanced delivery operator with the attention coefficient as the weight is used as a bridge to fuse the data features output from the autoencoder and the graph convolutional network model to obtain a model combining the autoencoder and the graph convolutional network; a data recommendation module is proposed for using the model combining the autoencoder and the graph convolutional network for topic recommendation.

[0080] The beneficial effects of the above technical solution are: by adopting the solution provided in this embodiment, the weights are dynamically adjusted through the attention coefficient, and an enhanced delivery operator is designed to more effectively incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer, unifying two different deep neural network structures, and obtaining a model combining the autoencoder and the graph convolutional network, which can enhance the model's ability to discriminate the importance of features. At the same time, the attention coefficient is introduced to adjust the weight ratio between the encoder and the graph convolutional network, which can ensure that the model automatically learns the optimal feature combination during the training process, rather than relying on a pre-assumed fixed ratio.

[0081] In one embodiment, Figure 2 As shown, the data processing module includes:

[0082] Remove stop words, punctuation marks and special characters from the text in the original data to obtain processed text;

[0083] Decompose the processed text into words or phrases, and use the bag-of-words model to convert the processed text into a target numerical vector;

[0084] Based on the original data, a KNN graph is constructed using the cosine similarity metric as the distance metric to convert the general data set into graph data and obtain the graph data in the input graph convolutional network model.

[0085] The working principle of the above technical solution is: in order to realize the processing of the original data, the present invention first removes the stop words, punctuation marks and special characters in the text of the original data to obtain the processed text; then decomposes the processed text into words or phrases, and uses the bag-of-words model to convert the processed text into a target numerical vector; then, based on the original data, a KNN graph is constructed using the cosine similarity metric as a distance metric. After the KNN graph is constructed, the nodes in the graph represent users or items, and the edges represent the similarity or closeness between the nodes, thereby helping to discover user groups or categories and providing a basis for personalized recommendations, so as to convert general data sets into graph data volumes and obtain graph data in the input graph convolutional network model.

[0086] The beneficial effect of the above technical solution is: by adopting the solution provided by this embodiment, by processing the original data, it can be ensured that the data after processing provides conditions for subsequent use.

[0087] In one embodiment, Figure 3 As shown, the model building modules include:

[0088] Input the target numerical vector into the autoencoder;

[0089] Input graph data into the graph convolutional network model;

[0090] The representation learned by the autoencoder is incorporated into the graph convolutional network model through the enhanced delivery operator to obtain a model combining the autoencoder and the graph convolutional network.

[0091] The working principle of the above technical solution is: in order to realize the construction of the model, the present invention first inputs the target numerical vector into the autoencoder; then inputs the graph data into the graph convolutional network model; finally, the representation learned by the autoencoder is incorporated into the graph convolutional network model through the enhanced delivery operator to obtain a model combining the autoencoder and the graph convolutional network.

[0092] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, the construction of a model combining an autoencoder and a graph convolutional network can provide a basis for the implementation of subsequent topic recommendations.

[0093] In one embodiment, the target numerical vector is input into the autoencoder, including:

[0094] Extract the intrinsic features of the target numerical vector based on the traditional autoencoder;

[0095] The decoder is used to reconstruct the low-dimensional text representation of the input data through multiple fully connected layers. First, the intrinsic features of the data are extracted based on the traditional autoencoder. If there are L layers in the autoencoder, then Indicates Layer, the encoder is at The representation learned by the layer for:

[0096]

[0097] Among them, v represents the Leaky ReLU activation function of the fully connected layer, and They are The weight matrix and bias of the layer encoder;

[0098] according to The original data X can be expressed as H (0) , the original data X can be compressed into a lower-dimensional data representation as input; the decoder reconstructs the input data through multiple fully connected layers, and the reconstructed data is represented as follows:

[0099]

[0100] in, Represents the reconstruction data, and Respectively represent The weight matrix and bias of the layer decoder, Representative The reconstructed data of the original data X of the layer input

[0101] The weights and biases are adjusted by the autoencoder through the back-propagation algorithm to minimize the loss function.

[0102] The working principle of the above technical solution is: in order to realize the input of the target numerical vector into the autoencoder, the present invention proposes to extract the intrinsic characteristics of the target numerical vector based on the traditional autoencoder; in order to maintain the attribute characteristics of the original data, the task of the decoder is to decompress the low-dimensional data and reconstruct it back to the original data form output, and use the decoder to reconstruct the text low-dimensional representation of the input data through multiple fully connected layers; the encoder adjusts the weights and biases through the back-propagation algorithm to minimize the loss function.

[0103] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, the encoding of the target numerical vector can be completed by inputting the target numerical vector into the automatic encoder, which is beneficial to subsequent data processing operations.

[0104] In one embodiment, the representation learned by the autoencoder is incorporated into the graph convolutional network through an enhanced delivery operator, including:

[0105] For each layer of information features in the deep neural network model, the weights are dynamically adjusted through the attention coefficient, and an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer;

[0106] The self-supervision module is introduced to unify the autoencoder and graph convolutional network layer in one framework to achieve the update of the deep neural network model.

[0107] The working principle of the above technical solution is to incorporate the representation learned by the autoencoder into the graph convolutional network through an enhanced delivery operator, including:

[0108] For each layer of information features in the deep neural network model, the weights are dynamically adjusted through the attention coefficient, and an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer;

[0109] The self-supervision module is introduced to unify the autoencoder and graph convolutional network layer in one framework to achieve the update of the deep neural network model.

[0110] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, the representation learned by the autoencoder is incorporated into the graph convolutional network through the enhanced delivery operator, which can ensure that a fusion model of the autoencoder and the graph convolutional network layer is obtained.

[0111] In one embodiment, the weights are dynamically adjusted by the attention coefficient, including:

[0112] Input graph adjacency matrix A and feature matrix X, for the weight matrix W of the neural network, the graph convolutional network The representation learned by the layer for:

[0113]

[0114] in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix, is the number of layers of the graph convolutional network, Z (l) It is Feature matrix of layer nodes;

[0115] Introducing dynamic adjustment of attention coefficient and The weight ratio of is used to form an adaptive fusion strategy, which effectively combines the output of the autoencoder with the output of the graph convolutional network and The heterogeneity of the layers represents bonding.

[0116] The working principle of the above technical solution is as follows: In order to dynamically adjust the weights through the attention coefficient, the present invention first inputs the adjacency matrix A and the feature matrix X of the graph, and for the weight matrix W of the neural network, the graph convolution network The representation learned by the layer for:

[0117]

[0118] in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix, is the number of layers of the graph convolutional network, It is Feature matrix of layer nodes; then introduce dynamic adjustment of attention coefficient and The weight ratio of is used to form an adaptive fusion strategy, which effectively combines the output of the autoencoder with the output of the graph convolutional network and The heterogeneity of the layers represents bonding.

[0119] The beneficial effects of the above technical solution are: adopting the solution provided by this embodiment, by introducing the attention coefficient to dynamically adjust and The weight ratio can ensure the formation of an adaptive fusion strategy.

[0120] In one embodiment, the attention coefficient is dynamically adjusted and The weight ratio of , to form an adaptive fusion strategy, also includes: learning the attention coefficient, specifically:

[0121] A shared linear transformation is performed on the feature representation and parameterized by a weight matrix W; the weight matrix W is calculated by a single-layer feedforward neural network to connect and And the weight vector α, then The attention score of a layer is:

[0122]

[0123] Among them, leakyReLU is an activation function. Score for attention;

[0124] The softmax function is used to normalize the attention scores to ensure that the sum of all attention coefficients is 1; the attention weight coefficient is expressed as follows:

[0125]

[0126] Among them, K represents any layer in the L layer, exp is the mathematical formula of the softmax function, and e (k) Respectively The attention scores of the layer and the k-th layer.

[0127] The working principle of the above technical solution is: introduce dynamic adjustment of attention coefficient and The weight ratio of , to form an adaptive fusion strategy, by performing a shared linear transformation on the feature representation and parameterizing it through the weight matrix W; the weight matrix W is calculated by a single-layer feedforward neural network and used to connect and And the weight vector α, calculate the The attention scores of the layers are normalized using the softmax function to ensure that the sum of all attention coefficients is 1.

[0128] The beneficial effects of the above technical solution are: adopting the solution provided by this embodiment, by performing a shared linear transformation on the feature representation and performing parameterization through a weight matrix; and obtaining the first The attention score and attention weight coefficient of the layer can ensure the dynamic adjustment of the attention coefficient and The weight ratio.

[0129] In one embodiment, an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer, including: after obtaining the importance ratio of each feature, designing a delivery operator to integrate the data attributes of the autoencoder into the feature expression in the graph convolutional network, and the delivery operator is expressed as:

[0130]

[0131] in, It is the graph convolutional network in The data representation learned by the layer, In the autoencoder The reconstructed data representation learned by the layer, For the The attention weight coefficient of the layer;

[0132] The fused features are output through the graph convolutional network layer, i.e. The input of the layer is but The output of the layer is:

[0133]

[0134] in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix;

[0135] The enhanced delivery operator works L times in the entire model and is represented by the softmax function in the last layer. The output is:

[0136]

[0137] in, For the graph convolutional network in The data representation learned by the layer, In the autoencoder The reconstructed data representation learned by the layer, W is a learnable parameter. When the network training is good, the predicted cluster label can be directly obtained through Z.

[0138] The working principle of the above technical solution is: The design principle of the enhanced delivery operator of the present invention is: From a mathematical point of view, the formula As the weight of the attention mechanism, the balance between models is dynamically adjusted. This is similar to the Lagrange multiplier method in mathematical optimization problems, which is used to find the optimal solution under constraints. The delivery process is achieved by weighted averaging because this weighted mechanism can enhance the model's ability to distinguish the importance of features. The improved enhanced delivery operator introduces dynamic weights. The invention provides the model with the ability to adaptively adjust the fusion ratio of feature representation. This design allows the model to automatically learn the optimal feature combination during training instead of relying on a pre-set fixed ratio. In addition, this weighting strategy helps the model to conduct more detailed exploration in the feature space, thereby improving the clustering performance. After obtaining the importance ratio of each feature, the invention designs a delivery operator to integrate the data attributes of the autoencoder into the feature expression in the graph convolutional network. The fused features are output through the graph convolutional network layer. In unsupervised clustering, the goal of the graph convolutional network is to use the structural information of the graph and the similarity of node features for clustering. Through the multi-layer convolution operation of the graph convolutional network, the node features gradually aggregate their own features and the information of neighboring nodes, so that the feature vectors of similar nodes are mapped from the high-dimensional feature space to a new low-dimensional space, thereby better capturing the similarity between nodes. The enhanced delivery operator works L times in the entire model, and the softmax function is used to represent the last layer.

[0139] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, by designing an enhanced delivery operator, the predicted cluster labels can be directly obtained through learnable parameters, allowing the model to automatically learn the optimal feature combination during the training process instead of relying on a pre-set fixed ratio.

[0140] In one embodiment, the topic recommendation is performed using a model combining an autoencoder and a graph convolutional network, and the topic recommendation effect is evaluated and optimized, specifically:

[0141] Obtain multiple similar models combining autoencoders and graph convolutional networks for horizontal comparison;

[0142] A model combining multiple similar autoencoders and graph convolutional networks is used to recommend topics, and multiple first topic recommendation results are obtained;

[0143] The topic recommendation effect and the plurality of first topic recommendation results are used as data sources, respectively input into a preset application scenario simulation model, simulate the application scenario, and obtain a simulation result of the application scenario;

[0144] According to the evaluation rules of the simulation results, the topic recommendation effect and the multiple first topic recommendation results are rated and scored for application effect respectively; and an average score value is obtained according to the rating and scoring results;

[0145] If the rating score of the topic recommendation effect is less than the average score, the feasibility of model optimization is judged and analyzed for the model combining the autoencoder and the graph convolutional network. If model optimization feasibility exists, the model combining the autoencoder and the graph convolutional network is optimized. If model optimization feasibility does not exist, the topic recommendation performance of the model combining the autoencoder and the graph convolutional network is calibrated for the optimal recommendation performance so as to better apply it to different scenarios.

[0146] The working principle of the above technical solution is: using a model combining an autoencoder and a graph convolutional network to recommend topics, and also includes evaluating and optimizing the topic recommendation effect, specifically:

[0147] Obtain multiple similar models combining autoencoders and graph convolutional networks for horizontal comparison;

[0148] A model combining multiple similar autoencoders and graph convolutional networks is used to recommend topics, and multiple first topic recommendation results are obtained;

[0149] The topic recommendation effect and the plurality of first topic recommendation results are used as data sources, respectively input into a preset application scenario simulation model, simulate the application scenario, and obtain a simulation result of the application scenario;

[0150] According to the evaluation rules of the simulation results, the topic recommendation effect and the multiple first topic recommendation results are rated and scored for application effect respectively; and an average score value is obtained according to the rating and scoring results;

[0151] If the rating score of the topic recommendation effect is less than the average score, the feasibility of model optimization is judged and analyzed for the model combining the autoencoder and the graph convolutional network. If model optimization feasibility exists, the model combining the autoencoder and the graph convolutional network is optimized. If model optimization feasibility does not exist, the topic recommendation performance of the model combining the autoencoder and the graph convolutional network is calibrated for the optimal recommendation performance so as to better apply it to different scenarios.

[0152] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, by evaluating and optimizing the topic recommendation effect, it is convenient to better play the function of the model combining the autoencoder and the graph convolutional network.

[0153] In one embodiment, it also includes adaptively matching different recommendation application scenarios according to the topic recommendation performance of the model combining the autoencoder and the graph convolutional network, specifically:

[0154] Obtain a model combining an autoencoder and a graph convolutional network, and obtain its calibrated optimal recommendation performance;

[0155] Obtain recommended application scenarios and recommended requirements for the recommended application scenarios;

[0156] According to the recommendation requirements, the personalized features of the recommended application scenarios are obtained;

[0157] Based on the set knowledge graph network structure, the corresponding optimal recommendation performance is obtained by utilizing the subject correspondence between the personalized features in the knowledge graph network structure and the optimal recommendation performance, and a model combining the autoencoder and the graph convolutional network corresponding to the optimal recommendation performance is obtained.

[0158] The working principle of the above technical solution is: it also includes adaptive matching of different recommendation application scenarios according to the topic recommendation performance of the model combining the autoencoder and the graph convolutional network, specifically:

[0159] Obtain a model combining an autoencoder and a graph convolutional network, and obtain its calibrated optimal recommendation performance;

[0160] Obtain recommended application scenarios and recommended requirements for the recommended application scenarios;

[0161] According to the recommendation requirements, the personalized features of the recommended application scenarios are obtained;

[0162] Based on the set knowledge graph network structure, the corresponding optimal recommendation performance is obtained by utilizing the subject correspondence between the personalized features in the knowledge graph network structure and the optimal recommendation performance, and a model combining the autoencoder and the graph convolutional network corresponding to the optimal recommendation performance is obtained.

[0163] The beneficial effect of the above technical solution is: by adopting the solution provided in this embodiment, by adaptively matching different recommended application scenarios, it can be ensured that the model combining the autoencoder and the graph convolutional network is better adapted to the application scenario.

[0164] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A recommendation system based on enhanced delivery operator graph clustering method, characterized in that: include: The data processing module is used to preprocess the raw data to obtain the target numerical vector for inputting the autoencoder and the graph data for inputting the graph convolutional network model; The model building module incorporates the representation learned by the autoencoder into the graph convolutional network model through the enhanced delivery operator based on the target numerical vector and graph data, and obtains a model combining the autoencoder and the graph convolutional network; The data recommendation module is used to recommend topics using a model that combines an autoencoder with a graph convolutional network.

2. A recommendation system based on enhanced delivery operator graph clustering method according to claim 1, characterized in that: The data processing module includes: Remove stop words, punctuation marks and special characters from the text in the original data to obtain processed text; Decompose the processed text into words or phrases, and use the bag-of-words model to convert the processed text into a target numerical vector; Based on the original data, a KNN graph is constructed using the cosine similarity metric as the distance metric to convert the general data set into graph data and obtain the graph data in the input graph convolutional network model.

3. The recommendation system based on the enhanced delivery operator graph clustering method according to claim 1, characterized in that: Model building modules include: Input the target numerical vector into the autoencoder; Input graph data into the graph convolutional network model; The representation learned by the autoencoder is incorporated into the graph convolutional network model through the enhanced delivery operator to obtain a model combining the autoencoder and the graph convolutional network.

4. A recommendation system based on enhanced delivery operator graph clustering method according to claim 3, characterized in that: The target numeric vector is fed into the autoencoder as follows: Extract the intrinsic features of the target numerical vector based on the traditional autoencoder; The decoder is used to reconstruct the low-dimensional text representation of the input data through multiple fully connected layers. First, the intrinsic features of the data are extracted based on the traditional autoencoder. If there are L layers in the autoencoder, then l represents the lth layer, and the representation H learned by the encoder at the lth layer is (l) for: Among them, φ represents the Leaky ReLU activation function of the fully connected layer, and are the weight matrix and bias of the l-th layer encoder respectively; According to H (l) , the original data X can be expressed as H (0) , the original data X can be compressed into a lower-dimensional data representation as input; the decoder reconstructs the input data through multiple fully connected layers, and the reconstructed data is represented as follows: in, Represents the reconstruction data, and Represent the weight matrix and bias of the l-th layer decoder respectively, Represents the reconstructed data of the original input data X of the l-1th layer The weights and biases are adjusted by the autoencoder through the back-propagation algorithm to minimize the loss function.

5. The recommendation system based on the enhanced delivery operator graph clustering method according to claim 3, characterized in that: The representations learned by the autoencoder are incorporated into the graph convolutional network through enhanced delivery operators, including: For each layer of information features in the deep neural network model, the weights are dynamically adjusted through the attention coefficient, and an enhanced delivery operator is designed to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer; The self-supervision module is introduced to unify the autoencoder and graph convolutional network layer in one framework to achieve the update of the deep neural network model.

6. A recommendation system based on enhanced delivery operator graph clustering method according to claim 5, characterized in that: Dynamically adjust weights through attention coefficients, including: Input graph adjacency matrix A and feature matrix X, weight matrix W of neural network, representation Z learned by layer l of graph convolutional network (l) for: in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix, l is the number of layers of the graph convolutional network, Z (l) is the feature matrix of the l-th layer nodes; Introducing attention coefficient to dynamically adjust Z (l) and H (l) to form an adaptive fusion strategy that effectively combines the output of the autoencoder with the graph convolutional network and combines the heterogeneous representations of the lth layer.

7. A recommendation system based on enhanced delivery operator graph clustering method according to claim 6, characterized in that: Introducing attention coefficient to dynamically adjust Z (l) and H (l) The weight ratio of , to form an adaptive fusion strategy, also includes: learning the attention coefficient, specifically: A shared linear transformation is performed on the feature representation and parameterized by a weight matrix W; the weight matrix W is calculated by a single-layer feed-forward neural network and used to connect Z (l) and H (l) , and the weight vector α, then the attention score of the lth layer is: Among them, leakyReLU is an activation function, e (l) Score for attention; The softmax function is used to normalize the attention scores to ensure that the sum of all attention coefficients is 1; the attention weight coefficient is expressed as follows: Among them, K represents any layer in the L layer, exp is the mathematical formula of the softmax function, e (l) and e (k) are the attention scores of the lth layer and the kth layer respectively.

8. The recommendation system based on the enhanced delivery operator graph clustering method according to claim 6, characterized in that: Design an enhanced delivery operator to incorporate the representation learned by the autoencoder into the corresponding graph convolutional network layer, including: after obtaining the importance ratio of each feature, design a delivery operator to integrate the data attributes of the autoencoder into the feature expression in the graph convolutional network. The delivery operator is expressed as: Among them, Z (l-1) is the data representation learned by the graph convolutional network at layer l-1, H (l-1) is the reconstructed data representation learned in the l-1th layer of the autoencoder, α (l-1) is the attention weight coefficient of the l-1th layer; The fused features are output through the graph convolutional network layer, that is, the input of the lth layer is The output of layer l is: in, Add the adjacency matrix of self-loops to the adjacency matrix A, which allows the node representation itself to be included when updating the node representation. is a diagonal matrix; The enhanced delivery operator works L times in the entire model and is represented by the softmax function in the last layer. The output is: Among them, Z (l) is the data representation learned by the graph convolutional network at layer l, H (l) It is the reconstructed data representation learned in the lth layer of the autoencoder, W is a learnable parameter, and when the network training is good, the predicted cluster label can be directly obtained through Z.

9. The recommendation system based on the enhanced delivery operator graph clustering method according to claim 1, characterized in that: The model combining the autoencoder and the graph convolutional network is used for topic recommendation, and the topic recommendation effect is evaluated and optimized, specifically: Obtain multiple similar models combining autoencoders and graph convolutional networks for horizontal comparison; A model combining multiple similar autoencoders and graph convolutional networks is used to recommend topics, and multiple first topic recommendation results are obtained; The topic recommendation effect and the plurality of first topic recommendation results are used as data sources, respectively input into a preset application scenario simulation model, simulate the application scenario, and obtain a simulation result of the application scenario; According to the evaluation rules of the simulation results, the topic recommendation effect and the application effects of the multiple first topic recommendation results are rated and scored respectively; And obtain the average score value according to the rating and scoring results; If the rating score of the topic recommendation effect is less than the average score, the feasibility of model optimization is judged and analyzed for the model combining the autoencoder and the graph convolutional network. If model optimization feasibility exists, the model combining the autoencoder and the graph convolutional network is optimized. If model optimization feasibility does not exist, the topic recommendation performance of the model combining the autoencoder and the graph convolutional network is calibrated for the optimal recommendation performance so as to better apply it to different scenarios.

10. A recommendation system based on enhanced delivery operator graph clustering method according to claim 9, characterized in that: It also includes adaptive matching of different recommendation application scenarios based on the topic recommendation performance of the model combining the autoencoder and the graph convolutional network, specifically: Obtain a model combining an autoencoder and a graph convolutional network, and obtain its calibrated optimal recommendation performance; Obtain recommended application scenarios and recommended requirements for the recommended application scenarios; According to the recommendation requirements, the personalized features of the recommended application scenarios are obtained; Based on the set knowledge graph network structure, the corresponding optimal recommendation performance is obtained by utilizing the subject correspondence between the personalized features in the knowledge graph network structure and the optimal recommendation performance, and a model combining the autoencoder and the graph convolutional network corresponding to the optimal recommendation performance is obtained.

Images