Method and system for recommending consumption financial products based on graph clustering of granules

Abstract

The application belongs to the field of machine learning and graph data mining, and relates to a consumption financial commodity recommendation method based on graph clustering of granular ball calculation, which comprises the following steps: obtaining a financial transaction network graph data set and performing information extraction thereon to obtain graph data G; performing granular ball calculation on the graph data G through an adaptive splitting mechanism to finally generate a plurality of granular balls; adding edges between the granular balls according to the adjacency relationship between the granular balls to generate a granulation graph; constructing a target function of a granular ball similarity matrix; solving the target function of the granular ball similarity matrix to obtain the granular ball similarity matrix; performing symmetric non-negative on the granular ball similarity matrix, performing spectral clustering on the symmetric non-negative granular ball similarity matrix to obtain a commodity clustering result; and generating high-quality granular balls through the adaptive splitting mechanism to preliminarily cover and represent the transaction graph data, replacing fine-grained graph nodes with coarse-grained granular balls to perform a clustering task, and capturing global and local features through multi-granularity representation, so that fast and accurate node clustering is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of machine learning and graph data mining, and relates to a method and system for recommending consumer financial products based on graph clustering of granular ball calculation. BACKGROUND

[0002] In the field of consumption, transaction networks are an important type of data structure, widely used in customer behavior analysis and product recommendation. In transaction networks, customers are nodes and transaction relationships are edges. These networks usually have complex structures and a large number of node interactions. For such complex graph structures, traditional K-means and hierarchical clustering methods have difficulty effectively processing global and local structure information of nodes. Therefore, graph clustering technology has gradually become one of the mainstream methods in transaction network product recommendation.

[0003] Currently, there are mainly the following technologies:

[0004] Graph partition-based clustering: transaction networks are divided into multiple subgraphs by graph partitioning algorithms, and each subgraph represents a customer group. Common graph partitioning methods include spectral clustering and community detection. These methods usually optimize the cutting of the graph, such as minimizing the number of edges across groups and maximizing the closeness of customer transaction relationships within groups.

[0005] Graph embedding-based clustering: graph embedding methods embed customer nodes in transaction networks into low-dimensional vector spaces, and then cluster these embedded vectors. This method can compress high-dimensional network structures into low-dimensional spaces, capturing complex relationships between nodes, and is suitable for customer grouping and behavior prediction.

[0006] Graph neural network (GNN) based clustering: graph neural networks can learn the representation of each customer by aggregating the neighborhood information of nodes, and perform clustering based on these representations. GNN models can capture complex structures in transaction networks, especially when dealing with large-scale networks, and perform well.

[0007] Hybrid method-based clustering: some research combines multiple technical means, such as the combination of graph embedding and time series analysis, or the hybrid method of GNN and community detection. By integrating information from multiple dimensions, these methods can more comprehensively cluster customers in consumption networks.

[0008] In summary, the graph clustering method in the existing transaction network commodity recommendation method lacks multi-granularity analysis and high computational complexity, they usually operate on a single granularity, lack hierarchical modeling capability, and may ignore some global information, and due to the fine-grained input, when facing complex consumption data structure, high-order convolution or attention calculation will have a large memory overhead and computational complexity, processing is more difficult, and the calculation cost is significantly increased. SUMMARY

[0009] To solve the above technical problems, the present application adopts a graph clustering-based commodity recommendation method based on granular ball calculation, which includes:

[0010] S1, obtaining a financial transaction network graph data set, extracting information from the financial transaction network graph data set to obtain the extracted graph data G; the financial transaction network graph data includes commodities, commodity information and the relationship between commodities;

[0011] S2, performing granular ball calculation on the graph data G through an adaptive splitting mechanism to generate a plurality of granular balls; each granular ball includes a plurality of commodities with similar structures;

[0012] S3, adding edges between the granular balls according to the adjacency relationship between the granular balls to generate a granulation graph, calculating the adjacency matrix and feature matrix of the granulation graph to obtain the granulation graph G'=(V GB , E GB , X GB , A GB ), wherein V GB is a node set of the granulation graph, each node represents a granular ball, X GB is a feature matrix of the granulation graph, E GB is an edge set of the granulation graph, and A GB is an adjacency matrix of the granulation graph;

[0013] S4, constructing a target function of the granular ball similarity matrix; solving the target function of the granular ball similarity matrix to obtain the granular ball similarity matrix;

[0014] S5, performing symmetric non-negative on the granular ball similarity matrix, and performing spectral clustering on the symmetric non-negative granular ball similarity matrix to obtain a commodity clustering result;

[0015] S6, performing commodity recommendation according to the commodity clustering result.

[0016] The granular ball calculation on the graph data G includes:

[0017] S21, regarding all commodity nodes in the graph data G as a granular ball and performing splitting to obtain initial granular balls Wherein, f is the index of the initial granule, n is the number of commodity nodes in the graph data G;

[0018] S22, according to the principle of proximity to each initial granule gb f Bisplitting, obtaining two sub-spheres gb f , gb 1，f , gb 2，f of each initial granule gb f ;

[0019] S23, judging whether the sum of the densities of two sub-spheres gb 1，f , gb 2，f Is lower than the density of the corresponding parent sphere gb f If yes, the final sub-granule of the parent sphere gb 1，f Otherwise, go back to step S22, respectively bisplitting two sub-spheres gb 2，f According to the principle of proximity.

[0020] Splitting all commodity nodes in the graph data G into a granule and performing Including: taking the product of the commodity node degree and the shortest path distance between the commodity nodes as the index of the center point, taking the first Commodity nodes with the largest index as the initial center point, and dividing the remaining commodity nodes into the initial center point according to the shortest path principle, obtaining Initial granule gb f .

[0021] According to the principle of proximity to bisplitting the initial granule includes: selecting the two commodity nodes with the farthest path distance in the initial granule as the temporary center point, calculating the shortest path distance between the remaining commodity nodes in the initial granule and the temporary center point, and adding the remaining commodity nodes in the initial granule to the temporary center point with the closest distance, obtaining two sub-spheres.

[0022] The calculation process of the density of the granule includes: selecting the two commodity nodes with the farthest path distance in the granule, taking half of the path distance of the two commodity nodes as the radius of the granule, and dividing the number of commodity nodes in the granule by the radius of the granule, obtaining the density of the granule.

[0023] The feature matrix of the granule GB i In the granulated graph Wherein, x k Indicates the feature corresponding to the commodity node k in the i-th granule GB i |GB i |Indicates the number of commodity nodes in the granule.

[0024] The attribute information Processing the granulated graph by using graph convolution, obtaining the attribute information GBPower calculation is performed to obtain second-order neighbor information according to attribute information in the granulation graph and the second-order neighbor information to construct a granule similarity matrix.

[0025] The adjacency matrix A GB Power calculation includes: calculating the square of the adjacency matrix A GB A 2 GB , adding the adjacency matrix A GB and A 2 GB to obtain the second-order neighbor information f(A GB ).

[0026] The objective function of the granule similarity matrix is:

[0027]

[0028] Where S GB is the granule similarity matrix, alpha is a weighting parameter, superscript T is the transpose, denotes the square of the Frobenius norm, and f(A GB ) is the second-order neighbor information.

[0029] On the other hand, the present application adopts a kind of consumption financial product recommendation system, comprising:

[0030] Input module, for obtaining financial transaction network graph data;

[0031] Granule generation module, for generating a plurality of granules according to the financial transaction network graph data;

[0032] Granulation graph generation module, for generating a granulation graph according to a plurality of granules;

[0033] Granule similarity matrix calculation module, for calculating the granule similarity matrix according to the granulation graph;

[0034] Clustering module, for obtaining product clustering results according to the granule similarity matrix;

[0035] Product recommendation module, for product recommendation according to product clustering results.

[0036] Beneficial effects:

[0037] 1. Inspired by the human brain's "large-scale first" cognitive mechanism and the granular computing splitting mechanism, this invention uses granules of different granularity to cover and represent the data space of the financial transaction network graph, with excellent data representation capabilities. 2. This invention uses the characteristic matrix of granules instead of the characteristic matrix of graph nodes for clustering, and reconstructs the graph similarity matrix in a coarse-grained manner, enhancing the adaptability to complex topological structures. The number of granules is far less than the number of graph nodes, which greatly reduces the matrix size, reduces computing costs, and improves the algorithm speed. 3. This invention introduces the density of granule centers to control their splitting to ensure the similarity of the structures within the granules, thereby capturing more advanced community structures and improving the accuracy of product recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flowchart of a method for recommending products on a trading network based on granular sphere calculation provided by an embodiment of the present invention;

[0039] Figure 2 A schematic diagram of a detailed granulation map generation process according to an embodiment of the present invention;

[0040] Figure 3 This is a comparison diagram before and after granulation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, the present invention adopts a method for recommending consumer finance products based on graph clustering using granular sphere computing, including:

[0043] S1. Obtain a financial transaction network graph dataset, extract information from the financial transaction network graph dataset, and obtain the extracted graph data G = (V, E, X, A), where V = {v1, v2, ..., v n} is a set of commodity nodes, X={x1,x2,...,x n} is the feature matrix corresponding to the commodity node, n is the number of commodity nodes, E is the edge set, A={a ij}∈R n×n is the adjacency matrix, if the product node v i , v j When there is an edge between them, that is, (v i , v j )∈E, then a ij=1, otherwise a ij =0.

[0044] In one embodiment, product nodes include: loan products, credit cards, financial products, etc.; an edge represents the common purchase or usage relationship between two products. For example, if a customer often chooses two products at the same time, or often purchases a certain financial product after applying for a certain credit card, an edge is established between the two products. The characteristics of each product are composed of its descriptive information, such as rate of return, risk level, fee structure, target customer group, etc.

[0045] S2. Perform sphere calculation on the graph data G through an adaptive splitting mechanism to capture high-order neighbor information, aggregate nodes with similar structures into spheres, and ultimately generate multiple spheres; each sphere includes multiple product nodes; each sphere can be regarded as a subgroup that aggregates similar products.

[0046] like Figure 2 As shown, the particle-sphere calculation for the graph data G includes:

[0047] The first stage of initial sphere division: all commodity nodes in the graph data G are regarded as a sphere and Split, get Initial spheres, that is, the initial sphere set;

[0048] Specifically, the product of the commodity node degree and the shortest path between commodity nodes is used as the index of the center point in the hope of obtaining a uniform distribution; the node with the largest index is placed at the top. The commodity node is used as the initial center point, and the remaining commodity nodes are divided into the initial center point according to the shortest path principle. Initial sphere gb f .

[0049] Two-stage quality-controlled sphere splitting: each initial sphere is subjected to cyclic binary splitting according to the proximity principle and sphere splitting density gain to obtain the final sphere set;

[0050] Since the present invention divides the sphere into two parts based on the shortest path from the node to the center, it is possible that the shortest path from one node to two centers is the same. In this case, the node will be divided into the center point that participates in the judgment first, which may cause the gap between the two sub-spheres of the sphere to be too large. In order to avoid this situation as much as possible, the present invention selects the two farthest commodity nodes as temporary center points for sphere division, so that the nodes will be added more evenly based on the shortest path.

[0051] Specifically, the two commodity nodes farthest apart in the initial granule are selected as temporary center points, the distances between the remaining commodity nodes in the initial granule and the temporary center points are calculated, the remaining commodity nodes in the initial granule are added to the temporary center point closest to them, two sub-balls are obtained, and it is determined whether the sum of the densities of the two sub-balls is lower than the density of the parent ball, i.e. the density of the initial granule. If yes, the final granule is obtained; otherwise, the two sub-balls are respectively subjected to binary splitting, and the binary splitting process is cycled until the sum of the densities of the sub-balls is lower than the density of the parent ball, the splitting cycle is stopped, and the final granule set is obtained.

[0052] The calculation process of the density of the granule includes:

[0053] Calculation of the radius of the granule where d(i, j) represents the shortest path distance between commodity node i and commodity node j in the granule, which represents the maximum value in all commodity node pairs (i, j).

[0054] Calculation of the density of the granule where |GB| represents the number of commodity nodes in a single granule, the greater the density, the closer the connection between the commodity nodes in the granule, and the higher the quality of the granule.

[0055] S3, edges are added between the granules according to the adjacency relationship between the granules to generate a granulation graph; an adjacency matrix and a feature matrix of the granulation graph are calculated to obtain the granulation graph G'=(V GB , E GB , X GB , A GB );

[0056] where V GB ={GB1, GB2,..., GB N} is a set of granule nodes generated from the graph data set, and each node represents a granule;

[0057] is the feature matrix corresponding to the granule, N is the number of granule nodes in the granulation graph, and d is the dimension of the feature of each granule, where x k represents the feature corresponding to commodity node k in the i-th granule GB i , and | GB i | represents the number of commodity nodes in the granule GB i .

[0058] E GB is an edge set, and as long as there is an edge between a pair of commodity nodes in two granules, there is an edge between the two granules, which is represented by an adjacency matrix A GB ={b ij}∈R N×N , which represents that if (GBi , GB j )∈E GB , then b ij = 1, otherwise b ij = 0.

[0059] Adjacency matrix of the granulated graph

[0060] The original graph data and the granulated graph data are shown in FIGS. 1 and 2, respectively. Figure 3

[0061] S4, constructing a target function of a granule-sphere similarity matrix according to the granulated graph; solving the target function of the granule-sphere similarity matrix to obtain the granule-sphere similarity matrix.

[0062] Specifically, first, the neighbor information in the granulated graph is aggregated and the information in the granulated graph is smoothed by using graph convolution, so that the attributes between adjacent granule-sphere nodes are more similar, and attribute information X' is obtained. GB The specific steps of the graph convolution are as follows:

[0063] First, the adjacency matrix is symmetrically normalized:

[0064]

[0065] where I is an identity matrix, L GB is the number of layers of the filter, and D GB is the degree matrix of the granulated graph.

[0066] Then, the graph filter is used to transform the attribute of the granule-sphere node to filter out the unwanted high-frequency signals and retain the geometric features of the graph, where m is a positive integer, and the formula is described as follows:

[0067]

[0068] For different complexity of the data set, the required m is different, preferably, m∈{1, 2,..., 6}.

[0069] The power calculation is performed on the adjacency matrix A GB , and the second-order neighbor information is accumulated and taken as a part of the similarity matrix: f(A GB ) = A GB + A 2 GB The second-order neighbor information represents the indirect association between commodities and reveals the potential purchase path.

[0070] ​The granule similarity matrix records the similarity between each pair of granule nodes, and the higher the similarity, the more likely it is that two granule nodes belong to the same group.

[0071] The granule similarity matrix objective function based on structure and features is as follows:

[0072]

[0073] Wherein, S GB is the granule similarity matrix, α is a trade-off parameter, controlling the amount of neighborhood information to be preserved, the superscript T is the transpose, represents the square of the Frobenius norm, f(A GB ) is the second-order neighborhood information.

[0074] By minimizing and and the difference between the granule similarity matrix S GB and the second-order neighborhood f(A GB ) calculated from the adjacency matrix A GB of the granule graph, the preservation of node feature representation is optimized, the loss of information is minimized in clustering or other forms of feature transformation, the high-order structural features of the graph are maintained, and the pairwise relationship between granules can not be completely simulated by the adjacency matrix alone, because it is usually sparse, so the second-order neighborhood is adopted.

[0075] The obtained S GB is as follows:

[0076]

[0077] S5, symmetrical non-negative is performed on the granule similarity matrix, and spectral clustering is performed on the symmetrical non-negative granule similarity matrix to obtain a commodity clustering result.

[0078] The symmetrical non-negative of the granule similarity matrix includes:

[0079]

[0080] Wherein, C GB is the symmetrical non-negative granule similarity matrix, || is the absolute value, and the superscript T is the transpose.

[0081] S6, commodity recommendation is performed according to the commodity clustering result.

[0082] Other commodities in the cluster where the user has browsed or purchased commodities are recommended to the user. If the user is interested in a commodity, other commodities in the same cluster are recommended because they are similar in attributes. For example, the user has purchased “short-term financial product A”, and “short-term financial product B” and “short-term financial product C” in the same cluster can be recommended.

[0083] In one embodiment, a consumption financial product recommendation system is adopted, comprising:

[0084] An input module is configured to acquire financial transaction network graph data;

[0085] A granule generation module is configured to generate a plurality of granules according to the financial transaction network graph data;

[0086] A granulated graph generation module is configured to generate a granulated graph according to the plurality of granules;

[0087] A granule similarity matrix calculation module is configured to calculate a granule similarity matrix according to the granulated graph;

[0088] A clustering module is configured to obtain a product clustering result according to the granule similarity matrix;

[0089] A product recommendation module is configured to recommend products according to the product clustering result.

[0090] The above embodiments further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above embodiments are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for recommending a consumption financial product based on graph clustering of granules, characterized by, The method comprises the following steps: S1, acquiring a financial transaction network graph data set, performing information extraction on the financial transaction network graph data set, and obtaining extracted graph data G; The financial transaction network graph data comprises commodities, commodity information, and relationships between commodities; S2, performing granular ball calculation on the graph data G through a self-adaptive splitting mechanism to generate a plurality of granular balls; each granular ball comprises a plurality of commodities; S3, adding edges between granules according to the adjacency relationship between the granules, generating a granulation graph G'=(V GB , E GB , X GB , A GB ), wherein V GB is a node set of the granulation graph, each node representing a granule, X GB is a feature matrix of the granulation graph, E GB is an edge set of the granulation graph, and A GB is an adjacency matrix of the granulation graph; S4, constructing a target function of a granular ball similarity matrix according to the granulated graph; solving the target function of the granular ball similarity matrix to obtain the granular ball similarity matrix; The objective function for constructing the granule similarity matrix according to the granulation graph comprises: processing the granulation graph by using graph convolution to obtain attribute information The adjacency matrix A GB The power calculation is performed to obtain second-order neighbor information, and the attribute information in the granulation graph and the second-order neighbor information to construct the objective function of the granule similarity matrix On the adjacency matrix A GB Power calculation includes: calculating the adjacency matrix A GB Square A 2 GB , the adjacency matrix A GB And A 2 GB Addition, get the second order neighbor information f(A GB ) The target function of the granular ball similarity matrix is: where S GB is the similarity matrix of granules, a is a trade-off parameter, the superscript T is the transpose, denotes the square of the Frobenius norm, f(A GB ) is the second-order neighbor information; S5, performing symmetric non-negative on the granular ball similarity matrix, and performing spectral clustering on the symmetric non-negative granular ball similarity matrix to obtain a commodity clustering result; S6, performing commodity recommendation according to the commodity clustering result.

2. The method according to claim 1, wherein the method is characterized by The granular ball calculation on the graph data G comprises: S21, all commodity nodes in the graph data G are regarded as one particle ball and are divided to obtain n initial particle balls gb , f , wherein f is an index of the initial particle ball, and n is a quantity of commodity nodes in the graph data G. S22, dividing each initial granulocyte gb according to the nearest principle f to obtain two sub-spheres gb f of each initial granulocyte gb 1,f , gb 2,f ; S23, judge whether the sum of the densities of the two child spheres gb 1,f , gb 2,f is lower than the density of the corresponding parent sphere gb f , if yes, get the final child sphere gb f ; otherwise, return to step S22, and split the two child spheres gb 1,f , gb 2,f according to the nearest principle.

3. The method according to claim 2, wherein the method is characterized by all commodity nodes in the graph data G are regarded as one granule and are divided into The splitting includes: taking the product of the commodity node degree and the shortest path distance between commodity nodes as the index of the center point, taking the first commodity nodes with the maximum index as the initial center points, and dividing the remaining commodity nodes into the initial center points according to the shortest path principle to obtain initial granules gb f .

4. The method according to claim 2, wherein the method is characterized by The bisection of the initial granular ball according to the nearest principle comprises: selecting two commodity nodes farthest from each other in the initial granular ball as temporary center points, calculating the shortest path distance between the remaining commodity nodes in the initial granular ball and the temporary center points, adding the remaining commodity nodes in the initial granular ball to the temporary center point closest to them to obtain two sub-balls.

5. The method according to claim 2, wherein the method is characterized by The calculation process of the density of the granular ball comprises: selecting two commodity nodes farthest from each other in the granular ball, taking half of the path distance of the two commodity nodes as the radius of the granular ball, and dividing the number of commodity nodes in the granular ball by the radius of the granular ball to obtain the density of the granular ball.

6. The method according to claim 1, wherein the method is characterized by Granular ball GB in granulation diagram i The characteristic matrix Among them, x k represents the i-th ball GB i The feature corresponding to the inner product node k, |GB i | represents the granular ball GB i The number of product nodes within the item.

7. A system for recommending a consumption financial product based on the graph clustering based on the particle calculation according to any one of claims 1 to 6, characterized by The method comprises the following steps: An input module is configured to acquire financial transaction network graph data; A granular ball generation module is configured to generate a plurality of granular balls according to the financial transaction network graph data; A granulated graph generation module is configured to generate a granulated graph according to the plurality of granular balls; A granular ball similarity matrix calculation module is configured to calculate a granular ball similarity matrix according to the granulated graph; A clustering module is configured to obtain a commodity clustering result according to the granular ball similarity matrix; A commodity recommendation module is configured to perform commodity recommendation according to the commodity clustering result.

Images