Consumer clustering method based on multi-modal graph comparative learning

Abstract

The invention discloses a consumer clustering method based on multi-modal graph comparative learning, and the method comprises the specific steps: (1) firstly processing a discrete multi-modal data set, obtaining neighbor nodes of each sample similarity top-k through employing a KNN method, and constructing an undirected graph structure; (2) obtaining two enhanced views from the obtained graph structure by using an edge disturbance data enhancement method, and then inputting the two enhanced views into a graph neural network to encode node feature representations of the two enhanced views; (3) performing comparative learning pre-training on the encoded feature representation by using an InfoNCE loss function, and finally training to obtain an optimized encoder; and (4) finally, inputting an originally constructed graph structure into a trained encoder to encode a final node feature representation, and then inputting the final node feature representation into a spectral clustering algorithm module for clustering to realize a clustering target. According to the method provided by the invention, discrete data is converted into a graph structure for processing, and a local structure between sample data can be captured; in addition, characteristics with higher discrimination can be learned by using comparative learning pre-training, so that a subsequent clustering algorithm can obtain more accurate clustering division.

Description

Technical Field

[0001] The present invention belongs to the field of node clustering in complex networks, and in particular relates to a consumer clustering method based on multimodal graph comparative learning. Background Art

[0002] With the continuous expansion of the economic market and increasingly fierce competition, companies need to develop effective marketing strategies to adapt to market changes. By analyzing multimodal data such as consumer purchase history, preferences, and needs, clustering algorithms are used to segment consumers into different groups. On this basis, by building customer portraits to reveal the demand characteristics, preferences, and behavior patterns of different consumer groups, companies can accurately grasp the target customer groups and implement more targeted marketing measures. This can not only improve the efficiency of marketing, but also effectively improve customer satisfaction and loyalty and promote sales growth. Furthermore, by building customer portraits, researchers can discover the characteristics of consumers in terms of health awareness, consumption motivation, and purchase decisions, thereby providing a scientific basis for product development and promotion. This study helps companies pay more attention to the actual needs of consumers in product design and improve the market adaptability of products.

[0003] In addition, conducting research on personalized marketing measures also has a far-reaching impact on enhancing brand value. In an era of information overload, consumers are faced with a large number of product choices, and personalized marketing can help companies stand out from the massive amount of information. By analyzing consumer behavior data, companies can tailor product recommendations and promotions for different customer groups, thereby enhancing consumers' willingness to buy and brand identity. This process not only optimizes resource allocation, but also enhances the company's market competitiveness.

[0004] In the field of cluster analysis, traditional clustering methods such as K-Means and hierarchical clustering rely on the original features of the data or simple similarity measures to divide the clusters. These methods usually assume that the data clusters have a specific shape, and their performance may decrease in high-dimensional space. In addition, traditional clustering methods often converge on local optimal solutions and may not capture the complex structure of the data.

[0005] Contrastive learning methods ultimately train an encoder to learn deep, discriminative feature representations of the data. By maximizing the similarity between positive pairs and minimizing the similarity between negative pairs, contrastive learning is able to identify and exploit complex structures and relationships in the data, thereby generating richer and more useful feature representations.

[0006] Based on this, in order to obtain a better clustering effect, the present invention combines contrastive learning with traditional clustering methods. After the contrastive learning pre-training stage, the learned feature representation will be used in the clustering algorithm to achieve the final cluster division. The advantage of this method is that it can significantly improve the accuracy and robustness of clustering. Finally, the clustering results are analyzed in depth in multiple dimensions. Explore the different preferences and behavior patterns of consumers to formulate personalized marketing strategies. This helps companies understand customer needs and preferences, improve customer retention, enhance brand competitiveness, and optimize products and services. Summary of the invention

[0007] Purpose of the invention: To address the problem of low brand repurchase rate among consumers in marketing, the present invention proposes a consumer clustering method based on multimodal graph comparative learning, which can learn a more obvious clustering division effect.

[0008] Technical solution: The present invention proposes a consumer clustering method based on multimodal graph contrast learning, comprising the following steps:

[0009] (1) Processing discrete multi-datasets, using the KNN method to obtain the top-k neighbor nodes of each sample similarity, and constructing an undirected graph structure;

[0010] (2) The edge perturbation data augmentation method is used on the obtained undirected graph structure to obtain two enhanced views, and then the two enhanced views are input into the graph neural network to encode their node feature representations;

[0011] (3) The encoded node feature representation uses the InfoNCE loss function for contrastive learning pre-training, and finally trains an optimized encoder;

[0012] (4) The originally constructed undirected graph structure is input into the trained encoder to encode the final node feature representation, which is then input into the spectral clustering algorithm module for clustering to achieve the clustering goal.

[0013] Furthermore, the implementation process of step (1) is as follows:

[0014] Assume that the original data X∈R N×d , where each row x i represents the i-th sample, N is the number of samples, and d is the dimension. For each sample, first find its top-K similar neighbors and set edges to connect them with the neighbors, thus forming a graph structure G.

[0015] Calculate the sample similarity matrix S∈R N×N There are many ways to construct a KNN graph. Here are three common methods for constructing a KNN graph:

[0016] 1) Euclidean distance. The similarity calculation formula between samples i and j is:

[0017] S ij =-||x i -x j || 2

[0018] This method uses negative Euclidean distance as a similarity measure. The smaller the distance, the higher the similarity. It is suitable for data with consistent dimensional scales.

[0019] 2) Cosine similarity. The similarity calculation formula between samples i and j is:

[0020]

[0021] This method determines the similarity of two vectors by measuring their similarity in direction without considering their size, and is suitable for vector data.

[0022] 3) Dot product. The similarity calculation formula between samples i and j is:

[0023]

[0024] The larger the dot product result calculated by this method, the more similar the two samples are. This method is applicable to any type of data.

[0025] After calculating the similarity matrix S∈R N×N Finally, this paper selects the top-K similar points of each sample as its neighbors, constructs an undirected k-nearest neighbor graph, and then obtains the adjacency matrix A from the non-graph data. The final graph can be expressed as: G = (A, X).

[0026] Furthermore, the implementation process of step (2) is as follows:

[0027] After obtaining the sample data G of the graph structure, in order to obtain a richer representation of sample nodes, the present invention uses edge perturbation data enhancement to generate two enhanced views G required for contrastive learning. 1 ′ and G 2 ′.

[0028] After obtaining the two enhanced views, the present invention inputs the two enhanced views into GCN to extract the node representation of the sample. For the graph G = (A, X), the process of GCN extracting the node representation is defined as:

[0029]

[0030] Where D is the degree matrix, I is the identity matrix, Z (k) is the node representation of the kth layer, W (k)is the weight matrix of the kth layer, and σ is the nonlinear activation function.

[0031] The node representation Z obtained for the two enhanced views 1 and Z 2 ,In order to improve the efficiency of contrastive learning training, the two representations are normalized before the contrastive loss, as follows:

[0032]

[0033] Finally, the normalized node representations are input into the contrastive learning framework for training.

[0034] Furthermore, the implementation process of step (3) is as follows:

[0035] The present invention does not directly use the node representation after GCN encoding for clustering analysis, but first trains by using the contrastive learning method to bring similar sample nodes closer and dissimilar sample nodes farther apart, in preparation for the subsequent clustering task.

[0036] Specifically, the present invention regards the corresponding nodes in the two enhanced views as a positive sample pair (such as z 1 ' i and z 2 ' i is a positive sample pair), and the rest of the nodes are considered negative sample pairs (such as z 1 ' i and z′ 2j is a negative sample pair). The loss function optimized by contrastive learning can be expressed as:

[0037]

[0038] Where τ is the temperature parameter and sim(·) is the similarity calculation function.

[0039] After the above operations, the model is finally optimized through the contrast loss function, thereby training and optimizing an encoder for subsequent clustering analysis.

[0040] Furthermore, the implementation process of step (4) is as follows:

[0041] After training the sample data, the present invention uses spectral clustering to cluster the samples. Compared with traditional clustering methods, such as K-Means and hierarchical clustering, spectral clustering is applicable to most data distributions, requires much less computation, and is easy to implement.

[0042] Before performing spectral clustering, the trained encoder is used to obtain the final sample node representation. The present invention uses the K nearest neighbor method to construct the similarity matrix S∈R using the new sample node representation.N×N , and then calculate the Laplacian matrix L of the graph based on the similarity matrix: L = DS, where D is the degree matrix. Then calculate the eigenvalues ​​and eigenvectors of the Laplacian matrix L. Use the eigenvectors corresponding to the smallest k non-zero eigenvalues ​​to form a new feature space. Finally, use clustering algorithms such as K-Means in these feature spaces to cluster the sample data to obtain the final result.

[0043] A device according to the present invention includes a memory and a processor, wherein:

[0044] A memory for storing computer programs that can be run on the processor;

[0045] The processor is used to execute the probiotic product user spectrum clustering algorithm steps based on contrastive learning as described above when running the computer program.

[0046] The computer-readable storage medium of the present invention stores a computer program, and the computer program is loaded and executed by a processor to implement the probiotic product user spectrum clustering algorithm steps based on contrastive learning.

[0047] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0048] (1) The present invention converts discrete consumer data sets into graph structures for processing, thereby effectively capturing the intrinsic geometric relationships of consumer data and providing important information about the similarity of data points.

[0049] (2) The present invention performs data enhancement on the original consumer data so that the model can learn different forms of graph data, making it more flexible and robust in the face of changes in the real world.

[0050] (3) The present invention adopts contrastive learning pre-training to optimize the feature representation of consumer data and provide high-quality feature input for subsequent clustering tasks.

[0051] The above improvements improve the model's clustering effect on consumer data and help the model form clearer inter-class distinctions and intra-class compactness in the feature space. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is the overall flow chart of the probiotics product user spectrum clustering algorithm based on contrastive learning;

[0053] Figure 2 Construct a graph structure diagram for KNN;

[0054] Figure 3 Schematic diagram of data enhancement and encoding;

[0055] Figure 4 Schematic diagram of pre-training for contrastive learning;

[0056] Figure 5 Schematic diagram of spectral clustering algorithm;

[0057] Figure 6 A schematic diagram for visualizing the clustering effect. DETAILED DESCRIPTION

[0058] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0059] like Figure 1 As shown, the present invention provides a consumer clustering method based on multimodal graph comparative learning, and the specific steps are as follows:

[0060] Step 1: Process the discrete data set, use the KNN method to obtain the top-k neighbor nodes of each sample similarity, and construct an undirected graph structure, such as Figure 2 shown.

[0061] The core goal of this step is to help companies achieve more accurate customer clustering analysis by building a graph structure based on the KNN (K-Nearest Neighbor) algorithm. By exploring consumer behavior and preferences, we first use the KNN algorithm to convert the similarity between data points into graph edges, and regard each data point as a node in the graph, thereby building a complete KNN graph. This graph structure can effectively capture the intrinsic geometric relationship of the data, provide important information about the similarity of data points, and provide the necessary graph structure foundation for subsequent graph convolutional networks (GCNs) and contrastive learning. Through this method, the structural information of the data is optimized, which facilitates more accurate clustering and analysis of subsequent deep learning models, thereby providing companies with more personalized and accurate marketing strategies.

[0062] Assume that the original data X∈R N×d , where each row x i represents the i-th sample, N is the number of samples, and d is the dimension. For each sample, first find its top-K similar neighbors and set edges to connect them with the neighbors, thus forming a graph structure G.

[0063] Calculate the sample similarity matrix S∈R N×N There are many ways to construct a KNN graph. Here are three common methods for constructing a KNN graph:

[0064] 1) Euclidean distance. The similarity calculation formula between samples i and j is:

[0065] Sij =-||x i -x j || 2

[0066] This method uses negative Euclidean distance as a similarity measure. The smaller the distance, the higher the similarity. It is suitable for data with consistent dimensional scales.

[0067] 2) Cosine similarity. The similarity calculation formula between samples i and j is:

[0068]

[0069] This method determines the similarity of two vectors by measuring their similarity in direction without considering their size, and is suitable for vector data.

[0070] 3) Dot product. The similarity calculation formula between samples i and j is:

[0071]

[0072] The larger the dot product result calculated by this method, the more similar the two samples are. This method is applicable to any type of data.

[0073] After calculating the similarity matrix S∈R N×N Finally, this paper selects the top-K similar points of each sample as its neighbors, constructs an undirected k-nearest neighbor graph, and then obtains the adjacency matrix A from the non-graph data. The final graph can be expressed as: G = (A, X).

[0074] Step 2: Use the edge perturbation data augmentation method to obtain two enhanced views on the obtained undirected graph structure, and then input the two enhanced views into the graph neural network to encode their node feature representations, such as Figure 3 shown.

[0075] Data augmentation is a crucial part of contrastive learning. Appropriate data augmentation can significantly improve the performance of the model and its ability to process unknown data. This helps the model to be better applied in the real world and improves the generalization ability of the model.

[0076] Therefore, after obtaining the sample data G of the graph structure, in order to obtain a richer sample node representation, the present invention uses edge perturbation data enhancement to generate two enhanced views G required for contrastive learning 1 ′ and G 2 ′.

[0077] After obtaining the two enhanced views, the present invention inputs the two enhanced views into GCN to extract the node representation of the sample. For the graph G = (A, X), the process of GCN extracting the node representation is defined as:

[0078]

[0079] Where D is the degree matrix, I is the identity matrix, Z (k) is the node representation of the kth layer, W (k) is the weight matrix of the kth layer, and σ is the nonlinear activation function.

[0080] The node representation Z obtained for the two enhanced views 1 and Z 2 ,In order to improve the efficiency of contrastive learning training, the two representations are normalized before the contrastive loss, as follows:

[0081]

[0082] Finally, the normalized node representations are input into the contrastive learning framework for training.

[0083] Step 3: The encoded node feature representation uses the InfoNCE loss function for contrastive learning pre-training, and finally trains to obtain an optimized encoder, such as Figure 4 shown.

[0084] Contrastive learning can learn rich and diverse feature representations by comparing the similarities and differences between different samples. InfoNCE is a commonly used contrastive learning loss function that compares the similarity of positive samples with that of negative samples, and in the process of minimizing the loss function, it shortens the distance between positive samples and increases the distance between negative samples, thereby achieving the purpose of contrastive learning.

[0085] Based on this, the present invention does not directly use the node representation after GCN encoding for clustering analysis, but first trains by using the graph contrast learning method to bring similar sample nodes closer and dissimilar sample nodes farther apart, in preparation for the subsequent clustering task.

[0086] Specifically, the present invention regards the corresponding nodes in the two enhanced views as a positive sample pair (such as z 1 ' i and z 2 ' i is a positive sample pair), and the rest of the nodes are considered negative sample pairs (such as z 1 ' i and z′ 2j is a negative sample pair). The loss function optimized by contrastive learning can be expressed as:

[0087]

[0088] Where τ is the temperature parameter and sim(·) is the similarity calculation function.

[0089] After the above operations, the model is finally optimized through the contrast loss function, thereby training and optimizing an encoder for subsequent clustering analysis.

[0090] Step 4: Input the original undirected graph structure into the trained encoder to encode the final node feature representation, and then input it into the spectral clustering algorithm module for clustering to achieve the clustering goal, such as Figure 5 shown.

[0091] After training the sample data, the present invention uses spectral clustering to cluster the samples. Compared with traditional clustering methods, such as K-Means and hierarchical clustering, spectral clustering is applicable to most data distributions, requires much less computation, and is easy to implement.

[0092] Before performing spectral clustering, the trained encoder is used to obtain the final sample node representation. The present invention uses the K nearest neighbor method to construct the similarity matrix S∈R using the new sample node representation. N×N , and then calculate the Laplacian matrix L of the graph based on the similarity matrix: L = DS, where D is the degree matrix. Then calculate the eigenvalues ​​and eigenvectors of the Laplacian matrix L. Use the eigenvectors corresponding to the smallest k non-zero eigenvalues ​​to form a new feature space. Finally, use clustering algorithms such as K-Means in these feature spaces to cluster the sample data to obtain the final result.

[0093] A device according to the present invention includes a memory and a processor, wherein:

[0094] A memory for storing computer programs that can be run on the processor;

[0095] The processor is used to execute the probiotic product user spectrum clustering algorithm steps based on contrastive learning as described above when running the computer program.

[0096] The computer-readable storage medium of the present invention stores a computer program, and the computer program is loaded and executed by a processor to implement the probiotic product user spectrum clustering algorithm steps based on contrastive learning.

[0097] Experimental verification:

[0098] Questionnaire survey is the main method used in this survey. We conducted online and offline surveys on different consumer groups. Taking probiotic products as an example, we learned about their basic information, probiotic product usage, problems that they expect probiotics to improve, consumption of X probiotic series products, and factors affecting purchase intention. A total of 1,205 questionnaires were distributed in the formal survey of this study, and 1,193 valid questionnaires were collected, with a total recovery rate of 99%. The questionnaire was set up in five parts: basic information, probiotic product usage, problems that they expect probiotics to improve, consumption of X probiotic series products, and factors affecting purchase intention, with a total of 83 options.

[0099] The present invention adopts the "Wenjuxing" online questionnaire survey method, which effectively avoids the problems of respondents missing answers and giving irregular answers. During the survey process, we marked the questionnaires with low cooperation, fast filling speed, casual attitude and abandoned midway to ensure that the quality of the recovered valid samples is high and there is no missing or incomplete data. After the questionnaire is distributed, the data is directly exported from the background of the Wenjuxing software, eliminating manual entry, reducing the risk of errors, and shortening the project execution cycle. Since the main purpose of the study is to increase the repurchase rate of product X, this article only analyzes the data for people who have used it, that is, 547 out of 1193 copies are studied, and finally the user's repurchase rate is used as the evaluation indicator of the model.

[0100] The proposed probiotics user spectrum clustering algorithm based on contrastive learning takes the GCN model commonly used in graph data networks as the base network, combines the contrastive learning method to train on the sample data set we collected, and finally achieves efficient clustering of sample data.

[0101] In order to compare the advantages of the method of the present invention over other network models, this experiment compared it with the commonly used clustering algorithms K-means, GMM (Gaussian mixture model), and spectral clustering (hereinafter referred to as SC) algorithm, and compared the clustering effects of these methods after comparative learning pre-training.

[0102] In this experiment, 300 epochs were uniformly set when training the comparative learning model. The k value was set to 6 or 10 when KNN constructed the graph structure. The learning rate during model training was set to 5e-5, and the weight_decay was set to 5e-4 by default.

[0103] The clustering effect of this method on the dataset is shown in Table 1.

[0104] Table 1 Clustering effect of different methods (repurchase rate%)

[0105]

[0106] From the data in the table, it can be seen that the method proposed in the present invention can improve the effects of various commonly used clustering algorithms. For the K-means algorithm, the highest repurchase rate of cluster grouping increased from 73.63% to 90.63%, while the GMM algorithm increased from 70.27% to 91.59%. Although the highest repurchase rate of ordinary SC algorithm clustering is relatively high, the clustering effect is not very ideal. After contrastive learning training, the highest repurchase rate has decreased, but the clustering effect has been greatly improved. As can be seen from the table, the clustering effect of the contrastive learning combined with spectral clustering method is obvious, and the analysis of each population after clustering will be more accurate using this method.

[0107] Finally, this paper also conducts a visual analysis of the results of each clustering, such as Figure 6 shown.

[0108] Figure 6 The clustering visualization effects of each method before and after contrastive learning are added are shown. After t-SNE is used to reduce the dimensionality of high-dimensional data, it can be clearly observed that contrastive learning significantly improves different methods. After adding contrastive learning pre-training, the clustering effect is significantly improved, and the separation between classes is significantly improved. Figure 6 It fully demonstrates that the method proposed by the present invention effectively enhances the intra-class compactness and inter-class separation, presents the best clustering effect, and facilitates the formulation of personalized marketing strategies for each cluster group.

Claims

1. A consumer clustering method based on multimodal graph contrastive learning, characterized in that: The following steps are involved: (1) Process the discrete multimodal data set, use the KNN method to obtain the top-k neighbor nodes of each sample similarity, and construct an undirected graph structure; (2) The edge perturbation data augmentation method is used on the obtained undirected graph structure to obtain two enhanced views, and then the two enhanced views are input into the graph neural network to encode their node feature representations; (3) The encoded node feature representation uses the InfoNCE loss function for contrastive learning pre-training, and finally trains an optimized encoder; (4) The originally constructed undirected graph structure is input into the trained encoder to encode the final node feature representation, which is then input into the spectral clustering algorithm module for clustering to achieve the clustering goal.

2. The consumer clustering method based on multimodal graph contrast learning according to claim 1, characterized in that: The specific steps of step (1) are: Assume that the original data X∈R N×d , where each row x i represents the i-th sample, N is the number of samples, and d is the dimension; for each sample, first find its top-K similar neighbors and set edges to connect it with the neighbors, thus forming a graph structure G; Calculate the sample similarity matrix S∈R N×N The methods include the following three: 1) Euclidean distance; The similarity calculation formula between samples i and j is: S ij =-||x i -x j || 2 By using negative Euclidean distance as the similarity measure, the smaller the distance, the higher the similarity, which is suitable for data with consistent dimensional scales; 2) Cosine similarity; the similarity calculation formula between samples i and j is: The similarity of two vectors is determined by measuring their similarity in direction, regardless of their magnitude, and is suitable for vector data. 3) Dot product; the similarity calculation formula between samples i and j is: The larger the dot product result obtained by calculation, the more similar the two samples are, which is applicable to any type of data; After calculating the similarity matrix S∈R N×N Finally, the top-K similar points of each sample are selected as its neighbors, an undirected k-nearest neighbor graph is constructed, and then the adjacency matrix A is obtained from the non-graph data; the final graph is represented as: G = (A, X).

3. The method for consumer clustering based on multimodal graph contrastive learning according to claim 2, characterized in that: The specific steps of step (2) are: After obtaining the sample data G of the graph structure, in order to obtain richer sample node representation, edge perturbation data enhancement is used to generate two enhanced views G1′ and G2′ required for contrastive learning; After obtaining the two enhanced views, the two enhanced views are input into GCN to extract the node representation of the sample; for the graph G = (A, X), the process of GCN extracting node representation is defined as: Where D is the degree matrix, I is the identity matrix, Z (k) is the node representation of the kth layer, W (k) is the weight matrix of the kth layer, σ is the nonlinear activation function; For the node representations Z1 and Z2 obtained from the two enhanced views, in order to improve the efficiency of contrastive learning training, the two representations are normalized before performing contrastive loss. The method is as follows: Finally, the normalized node representations are input into the contrastive learning framework for training.

4. The method for consumer clustering based on multimodal graph contrastive learning according to claim 3, characterized in that: The specific steps of step (3) are: First, we use contrastive learning to train similar sample nodes closer together and dissimilar sample nodes farther apart, in preparation for the subsequent clustering task. Specifically, the corresponding nodes in the two enhanced views are regarded as a positive sample pair, and the remaining nodes are regarded as negative sample pairs. The loss function optimized by contrastive learning is expressed as: Where τ is the temperature parameter, sim(·) is the similarity calculation function; After the above operations, the model is finally optimized through the contrast loss function, thereby training and optimizing an encoder for subsequent clustering analysis.

5. The method for consumer clustering based on multimodal graph contrastive learning according to claim 4, characterized in that: The specific steps of step (4) are: After training the sample data, the spectral clustering method is used to cluster the samples; Before performing spectral clustering, the trained encoder is used to obtain the final sample node representation; The K nearest neighbor method is used to construct the similarity matrix S∈R using the new sample node representation N×N , and then calculate the Laplacian matrix L of the graph based on the similarity matrix: L = DS, where D is the degree matrix; then calculate the eigenvalues ​​and eigenvectors of the Laplacian matrix L; use the eigenvectors corresponding to the smallest k non-zero eigenvalues ​​to form a new feature space; finally, use clustering algorithms such as K-Means in these feature spaces to cluster the sample data to obtain the final result.

6. A device for consumer clustering algorithm based on graph contrast learning, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the steps of a consumer spectrum clustering algorithm based on contrastive learning as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the steps of the consumer user clustering algorithm based on graph contrast learning as described in any one of claims 1 to 5.

Images