Financial transaction behavior recommendation prediction method and system based on message mask and cross-view comparative learning

Abstract

The invention provides a financial transaction behavior recommendation prediction method and system based on message masks and cross-view comparative learning, and the method specifically comprises the following steps: S1, building a comparative view of a user side through a graph neural network (GNN) based on the interaction information between a user and an article in a main view; s2, performing cross-view multi-level comparative learning on the main view and the comparative view, thereby deeply mining potential semantic correlation between the user and the article, improving the expression ability of a graph neural network, and obtaining an optimized graph neural network; and S3, according to interaction information between the user and the article, predicting behavior preferences of the user through the optimized graph neural network, and sorting future behaviors of the user according to the preferences.

Description

Technical Field

[0001] The present invention relates to the technical field of financial transaction behavior recommendation, and in particular to a financial transaction behavior recommendation prediction method and system based on message masking and cross-view contrastive learning. Background Art

[0002] With the continuous advancement and innovation of science and technology, financial transaction behaviors are undergoing profound changes. Mobile payment has rapidly become popular due to its convenience and efficiency, and the rise of digital currency has led the trend of payment revolution, and transaction behaviors and transaction channels are becoming increasingly diverse. At the same time, driven by the uncertainty of the global economic situation, technological progress, and changing regulatory requirements, financial security is receiving more and more attention from more people. How to predict future transaction behaviors and abnormal dynamics by analyzing customers' historical transaction data, market trends, and other relevant economic indicators is a very important issue. Recommendation systems are widely used on e-commerce platforms, short video platforms, etc. Recommendation systems based on collaborative filtering can predict users' future behaviors based on their interaction history. Therefore, collaborative filtering methods using recommendation systems can effectively solve the problems of financial transaction behavior prediction and anomaly prevention. Graph neural networks, as the main method of collaborative filtering, are in rapid development. LightGCN [1] Graph convolutional neural networks are used to aggregate embeddings and capture high-level interaction information between users and items, but the quality of data labels is required to be high. Subsequently, to address the problems of sparse labels and noise in real-world scenarios, graph comparative learning has gradually become popular and applied, and graph-based collaborative filtering recommendation systems have achieved a series of important breakthroughs. [2] Propose to randomly drop nodes (or edges) to generate contrast views, SimGCL [3] Granularity-controlled noise is introduced into the original structure, NCL [4] Combining structural information and global information to achieve consistent embedding, LightGCL [5] Approximate the SVD decomposition of the adjacency matrix to preserve key information.

[0003] However, despite the significant success of current graph neural networks in recommender systems, they still face the following challenges: 1) They lack the ability to mine information from the original embedding matrix when generating comparative views. 2) Frequent matrix operations destroy the original graph structure and result in high computational overhead. 3) They lack the ability to mine the underlying high-order connectivity between users and items (behaviors).

[0004] References

[0005] 1.HeX,DengK,WangX,etal.Lightgcn:Simplifying andpoweringgraphconvolutionnetwork for recommendation[C] / / Proceedings of the 43rdInternational ACM SIGIR conference on researchanddevelopmentinInformationRetrieval.2020:639-648.

[0006] 2.You Y,Chen T,Shen Y,et al.Graph contrastive learning automated[C] / / International ConferenceonMachineLearning.PMLR,2021:12121-12132.

[0007] 3.YuJ,YinH,XiaX,etal.Aregraphaugmentationsnecessary?simplegraphcontrastivelearning for recommendation[C] / / Proceedings of the 45thinternational ACM SIGIR conference on researchanddevelopmentininformationretrieval.2022:1294-1303.

[0008] 4.LinZ,TianC,HouY,etal.Improvinggraphcollaborativefilteringwithneighborhood-enriched contrastivelearning[C] / / ProceedingsoftheACMwebconference2022.2022:2320-2329.

[0009] 5.Cai X,Huang C,Xia L,et al.LightGCL:Simple yet effective graphcontrastive learning for recommendation[J].arXivpreprintarXiv:2302.08191,2023. Summary of the Invention

[0010] The present invention is made to solve the above problems, and its purpose is to provide a financial transaction behavior recommendation and prediction method and system based on message masking and cross-view contrastive learning.

[0011] The present invention provides a financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning, which has the following characteristics and specifically includes the following steps: S1, based on the interaction information between the user and the item in the main view, constructing a user-side comparative view through a graph neural network (GNN); S2, performing cross-view multi-level comparative learning on the main view and the comparative view, so as to deeply explore the potential semantic correlation between the user and the item, improve the expressive ability of the graph neural network, and obtain an optimized graph neural network; and S3, based on the interaction information between the user and the item, predicting the user's behavioral preferences through the optimized graph neural network, and sorting the user's future behaviors according to the preferences.

[0012] The financial transaction behavior recommendation and prediction method based on message masking and cross-view contrastive learning provided by the present invention may also have the following characteristics: wherein, in the graph-based recommendation method, the input user set is defined as U = {u1,u2,...,u n}, the input item (behavior) set is I = {i1,i2,...,i n}, the interaction matrix between the two is A. In graph contrastive learning, the core operating mechanism of the graph neural network is message passing, where a node sends messages to its neighbors and receives aggregated messages from its neighbors. The graph convolutional network (GCN) is a classic model in graph neural networks. It updates its own representation by aggregating information from neighboring nodes. Its aggregation update formula can be summarized as: in, represents the hidden embedding of the l+1th layer node u, σ represents the activation function, N u represents the neighbors of node u, e i,u represents the edge between node u and node i, AGG(·) represents the aggregation process of the graph neural network, and f(·) represents the message passing mechanism of the graph neural network.

[0013] The financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning provided by the present invention may also have the following features: wherein, the financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning also includes the following validity verification method: S4, through the four graph learning recommendation algorithms LightGCN, SGL, SimGCL, NCL, and LightGCL, the model effect is verified on three data sets of Yelp, Amazon and Ifashion, and the data is divided into training set and verification set.

[0014] In the financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning provided by the present invention, it can also have the following characteristics: wherein, step S1 includes the following sub-steps: S1-1, based on the interaction information between the user and the item in the main view, the user-item interaction matrix is input as the original adjacency matrix, and the original adjacency matrix is approximated by random SVD decomposition to obtain a reconstructed adjacency matrix; S1-2, the information of the reconstructed adjacency matrix is randomly masked by a graph neural network to obtain a mask matrix, and a comparative view of the user side is constructed by the reconstructed adjacency matrix and the mask matrix.

[0015] The financial transaction behavior recommendation and prediction method based on message masking and cross-view contrastive learning provided by the present invention may also have the following features: wherein, step S1-1 further includes the following sub-steps: S1-1-1, direct factorization of the matrix will produce a very high computational cost. Before SVD decomposition, the original matrix is first approximated with a q-order low-rank matrix. After obtaining the approximate low-rank matrix, matrix decomposition is performed. The computation process is expressed as follows: Approx(A,q), where represents an orthogonal matrix, Represents a diagonal matrix containing only q singular values of the original matrix, where q is the largest singular value; S1-1-2, after approximate SVD decomposition, the reconstructed adjacency matrix can be obtained for:

[0016] The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning provided by the present invention may also have the following features: wherein step S1-2 further includes the following sub-steps:

[0017] S1-2-1, using a matrix B that follows a Bernoulli distribution, randomly masks the matrix information to obtain a new matrix, and uses scaling to ensure that the expected values of the matrix before and after masking are equal. The calculation is as follows: Where M represents the original message aggregation matrix, represents the newly generated message aggregation matrix, p represents the masking rate, ⊙ represents element-wise matrix multiplication, and B represents the matrix with only 0 and 1 values obtained from Bernoulli sampling. S1-2-2 combines singular value decomposition with random information masking to construct a comparative view on the user side, which is calculated as follows: Where u represents a user, i represents an item, AGG(.) represents the message aggregation process of the graph convolutional network, and l represents the lth layer of the graph convolutional network. The comparative view calculation on the item side follows a similar principle:

[0018] The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning provided by the present invention may also have the following features: wherein step S2 includes the following sub-steps:

[0019] S2-1, in the first comparative interaction process, consider the semantic similarity between isomorphic neighbors in the comparative view and the main view. Since the user-item interaction graph is a bipartite graph, users aggregate and update themselves from the features of the interacted items. Therefore, it can be concluded that the embedding types of nodes at corresponding positions in the even layers are the same. The consideration of the odd layers is similar to that of the even layers.

[0020] The same nodes should have more similar semantic information and node features. Therefore, the cross-view semantic information is combined to mine the similarity between the contrast view and the main view. Therefore, contrastive learning is performed between the embeddings of the lth layer in the contrast view and the (l-2)th layer in the main view, where the same nodes in different views and layers are regarded as positive sample pairs, and nodes in different positions are regarded as negative sample pairs. The calculation is as follows:

[0021]

[0022] in, represents the user embedding of the corresponding layer of the contrast view, represents the user embedding of the corresponding layer of the main view, s() represents the similarity calculation function, τ1 represents the temperature coefficient, and U represents other users in the user set except u. The loss calculation method on the item side is similar to it:

[0023]

[0024] The loss formula calculation process for the first comparative interaction is:

[0025]

[0026] S2-2, in the second comparative interaction process, explore the potential preference associations between users and items in different views.

[0027] In a recommendation system, when a user interacts with an item, the user often has similar characteristics to other neighbors who interact with the item. According to the assumption of the recommendation system, similar users tend to interact with the same items. Therefore, the embedding matrix of the even-numbered contrast view on the user side is compared with the embedding matrix of the odd-numbered main view on the item side for comparative learning. This method can discover high-order hidden relationships between users and items. Through comparative learning, similar node pairs are brought closer to each other, and dissimilar node pairs are pulled apart. The calculation process of the user-side comparative learning loss is as follows: The meanings of the parameters in this formula are the same as before. Similarly, the calculation process of item-side contrastive learning loss is as follows: Finally, the second contrastive learning loss between users and items is calculated as follows: S2-3, in the interaction graph between users and items, nodes are embedded with ID information and therefore lack rich attribute information. Therefore, in the convolution process of the graph convolutional network, discarding feature transformation, activation function, and self-loop process can reduce algorithm redundancy and improve algorithm performance. Therefore, the embedding of each layer of the main view and the comparison view is simply summed as the output of the final graph convolutional network layer. The summation process is calculated as follows: Where E represents the feature embedding matrix of the main view, and G represents the feature embedding matrix of the comparison view. In S2-4, after mining the potential high-level information between users and between users and items, we use the InfoNCE loss to perform comparative learning between the main view and the comparison view at the same level. The calculation is as follows: where τ is the temperature coefficient and a represents a single element in the set.

[0028] The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning provided by the present invention may also have the following features: wherein step S3 includes the following sub-steps:

[0029] The goal of the recommendation system is to predict user behavior preferences based on user interaction behaviors and rank users' future behaviors based on their preferences. The Bayesian personalized ranking loss (BPRloss) is used to predict user interaction behaviors. The goal is to maximize the predicted preference probability of items that users actually interact with (positive samples) relative to items that they do not interact with (negative samples). The calculation process is as follows: σ represents the activation function, represents the prediction of positive sample interaction, It represents the prediction of the interaction of negative samples. The goal is to widen the gap between positive and negative samples as much as possible. The overall optimization goal of model training includes considering the BRP loss of the recommendation system prediction task and the InfoNCE loss of three-way contrastive learning. Finally, the total loss function of the model optimization is obtained by adding up the losses of each part. The calculation is as follows: Among them, λ1, λ2, and λ3 represent the weight coefficients of contrastive learning loss, and Θ represents the regularization term of the model.

[0030] The present invention also provides a financial transaction behavior recommendation and prediction system based on message masking and cross-view comparative learning, which has the following characteristics, including: a comparative view construction module, which constructs a user-side comparative view through a graph neural network (GNN) based on the interaction information between the user and the item in the main view; a graph neural network optimization module, which performs cross-view multi-level comparative learning on the main view and the comparative view, thereby deeply mining the potential semantic correlation between the user and the item, improving the expressive ability of the graph neural network, and obtaining an optimized graph neural network; and a user behavior prediction module, which predicts the user's behavior preferences through the optimized graph neural network based on the interaction information between the user and the item, and sorts the user's future behavior according to the preferences.

[0031] The financial transaction behavior recommendation and prediction system based on message masking and cross-view comparative learning provided by the present invention may also have the following features: it also includes a validity verification module, which verifies the model effect on three data sets of Yelp, Amazon and Ifashion through four graph learning recommendation algorithms, LightGCN, SGL, SimGCL, NCL and LightGCL, and the data are divided into training set and verification set.

[0032] Functions and effects of the invention

[0033] The financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning according to the present invention can provide theoretical support for financial transaction behavior, thereby achieving high-quality and lightweight recommendation prediction effects. The method proposed in the present invention improves 3%-8% compared to the basic algorithm, and achieves good prediction performance on three data sets, while maintaining low computational cost and optimal prediction accuracy. The algorithm can increase the ability to explore original information while reducing sample variance and maintaining information diversity, and alleviate the overfitting problem; it can simplify calculations and improve the lightweight of the algorithm model through specific graph reconstruction; by strengthening the cross-view multi-layer interaction between the main view and the comparison view, the model's ability to mine potential high-order connectivity relationships between nodes can be greatly improved; the generation of comparison views based on graph decomposition and reconstruction with random message discarding and multi-level cross-view comparative learning can fully improve the model's representation ability and prediction accuracy, and provide theoretical support for the prediction of financial transaction behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 2 is a schematic diagram of a financial transaction behavior recommendation and prediction method based on message masking and cross-view contrast learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the financial transaction behavior recommendation and prediction method based on message masking and cross-view comparative learning of the present invention.

[0036] This paper proposes a simple yet effective recommendation system model based on message masking and cross-view contrastive learning to predict financial transaction behavior. It constructs contrastive views by randomly aggregating information during graph reconstruction and then performs cross-view contrast interaction based on these contrastive views.

[0037] The two core parts of this method: graph reconstruction technology of message masking and contrastive learning of cross-view interaction can effectively solve the main problems currently faced by collaborative filtering recommendation systems and achieve efficient and accurate recommendations and predictions.

[0038] First, in the process of constructing the contrast view, we approximate the original adjacency matrix through random SVD decomposition, reducing the computational complexity of graph reconstruction while basically preserving the original graph structure.

[0039] Subsequently, the message mask of Bernoulli distribution is used to optimize the aggregation process of graph neural network, thereby improving the ability to mine the original matrix information while retaining more information, alleviating the problem of overfitting, and improving the generalization of graph neural network.

[0040] Then, we use the cross-view contrastive learning paradigm to explore the complex relationships of high-order connectivity between users and high-level users, and users and high-level products, to achieve high-precision contrastive learning.

[0041] Definition: In the graph-based recommendation method, the input user set is U = {u1,u2,...,u n}, the input item (behavior) set is I = {i1,i2,...,i n}, and their interaction matrix is A. In graph contrastive learning, the core of GNN's operating mechanism is message passing, where a node sends messages to its neighbors and receives aggregated messages from its neighbors. GCN is a classic model in GNN, which updates its own representation by aggregating information from neighboring nodes. Its aggregation update formula can be summarized as

[0042]

[0043] in, represents the hidden embedding of the l+1th layer node u, σ represents the activation function, N u represents the neighbors of node u, e i,u represents the edge between node u and node i, AGG(·) represents the aggregation process of the graph neural network, and f(·) represents the message passing mechanism of the graph neural network.

[0044] From the perspective of deep neural networks, we propose a recommendation system model based on message masking and cross-view contrastive learning to support financial transaction behavior prediction.

[0045] Figure 1 2 is a schematic diagram of a financial transaction behavior recommendation and prediction method based on message masking and cross-view contrast learning in an embodiment of the present invention.

[0046] like Figure 1 As shown in the figure, the core is divided into three steps: SVD graph reconstruction with random message dropping, multiple cross-view comparison interactions, and multi-loss and multi-objective joint optimization. The specific steps are detailed as follows:

[0047] S1, based on the interaction information between users and items in the main view, constructs a user-side comparative view through a graph neural network (GNN).

[0048] Contrastive learning effectively addresses popularity bias by obtaining representations of users and items. However, many methods remove edges from the adjacency matrix before reconstructing the graph, a necessary step to prevent overfitting during model training. However, this random edge dropout mechanism results in a significant loss of information about the original graph structure, which compromises the expressive power of GNNs. Therefore, we improve the message dropout method during contrastive view generation, aiming to produce more robust and accurate embeddings.

[0049] Step S1 specifically includes the following sub-steps:

[0050] S1-1, based on the interaction information between the user and the item in the main view, input the user-item interaction matrix as the original adjacency matrix, and approximate the original adjacency matrix through random SVD decomposition to obtain a reconstructed adjacency matrix.

[0051] Step S1-1 specifically includes the following sub-steps:

[0052] S1-1-1, direct factorization of the matrix will produce a very high computational cost. Before SVD decomposition, the original matrix is first approximated with a q-order low-rank matrix. After obtaining the approximate low-rank matrix, the matrix decomposition is performed. The calculation process is expressed as:

[0053]

[0054] in represents an orthogonal matrix, represents a diagonal matrix containing only the q singular values of the original matrix, where q is the largest singular value.

[0055] S1-1-2, after approximate SVD decomposition, the reconstructed adjacency matrix can be obtained for:

[0056]

[0057] S1-2, randomly masking the information of the reconstructed adjacency matrix through a graph neural network to obtain a mask matrix, and constructing a user-side comparison view through the reconstructed adjacency matrix and the mask matrix.

[0058] Random masking is applied during the message passing phase of the graph neural network. The masking step of aggregating matrix information can also be viewed as a form of Bernoulli sampling process.

[0059] Step S1-2 specifically includes the following sub-steps:

[0060] S1-2-1, using a matrix B that follows a Bernoulli distribution, randomly masks the matrix information to obtain a new matrix, and uses scaling to ensure that the expected values of the matrix before and after masking are equal. The calculation is as follows:

[0061]

[0062] Where M represents the original message aggregation matrix, represents the newly generated message aggregation matrix, p represents the masking rate, ⊙ represents the element-wise matrix multiplication, and B represents the matrix with only 0 and 1 values obtained from Bernoulli sampling.

[0063] S1-2-2 combines singular value decomposition with random information masking to construct the user-side comparison view. The calculation is as follows:

[0064]

[0065] Where u represents a user, i represents an item, AGG(.) represents the message aggregation process of the graph convolutional network, and l represents the lth layer of the graph convolutional network. The comparative view calculation on the item side follows a similar principle:

[0066]

[0067] S2, performing cross-view multi-level comparative learning on the main view and the comparison view, thereby deeply mining the potential semantic correlation between the user and the object, improving the expression ability of the graph neural network, and obtaining an optimized graph neural network.

[0068] While generating comparative views based on SVD can preserve key collaborative filtering information, it lacks deep mining of the underlying semantic correlations between items and users, which is crucial for improving the expressive power of graph neural networks. Therefore, after constructing the comparative views, cross-view, multi-level comparative learning is required for the main and comparative views.

[0069] Step S2 specifically includes the following sub-steps:

[0070] In S2-1, during the first comparative interaction, we consider the semantic similarity between isomorphic neighbors in the comparative view and the main view. Since the user-item interaction graph is a bipartite graph, users aggregate and update themselves based on the features of the interacted items. Therefore, we can conclude that the embedding types of nodes at corresponding positions in the even layers are the same. The considerations for odd layers are similar to those for even layers.

[0071] The same nodes should have more similar semantic information and node features. Therefore, the cross-view semantic information is combined to mine the similarity between the contrast view and the main view. Therefore, contrastive learning is performed between the embeddings of the lth layer in the contrast view and the (l-2)th layer in the main view, where the same nodes in different views and layers are regarded as positive sample pairs, and nodes in different positions are regarded as negative sample pairs. The calculation is as follows:

[0072]

[0073] in, represents the user embedding of the corresponding layer of the contrast view, represents the user embedding of the corresponding layer of the main view, s() represents the similarity calculation function, τ1 represents the temperature coefficient, and U represents other users in the user set except u. The loss calculation method on the item side is similar to it:

[0074]

[0075] The loss formula calculation process for the first comparative interaction is:

[0076]

[0077] S2-2, in the second comparative interaction process, the potential preference associations between users and items are explored in different views.

[0078] In a recommendation system, when a user interacts with an item, the user often has similar characteristics to other neighbors who interact with the item. According to the assumption of the recommendation system, similar users tend to interact with the same items. Therefore, the embedding matrix of the even-numbered contrast view on the user side is compared with the embedding matrix of the odd-numbered main view on the item side for comparative learning. This method can discover high-order hidden relationships between users and items. Through comparative learning, similar node pairs are brought closer to each other, and dissimilar node pairs are pulled apart. The calculation process of the user-side comparative learning loss is as follows:

[0079]

[0080] The meanings of the parameters in this formula are the same as before. Similarly, the calculation process of item-side contrastive learning loss is as follows:

[0081]

[0082] Finally, the second contrastive learning loss between users and items is calculated as follows:

[0083]

[0084] S2-3, in the interaction graph between users and items, the nodes are embedded with ID information, so they lack rich attribute information. Therefore, in the process of convolution of the graph convolutional network, discarding feature transformation, activation function and self-loop process can reduce the redundancy of the algorithm and improve the performance of the algorithm.

[0085] Therefore, we simply sum the embeddings of each layer of the main view and the contrast view to serve as the output of the final graph convolutional network layer. The summation process is calculated as follows:

[0086]

[0087] Among them, E represents the feature embedding matrix of the main view, and G represents the feature embedding matrix of the comparison view.

[0088] In S2-4, after mining the potential high-level information between users and between users and items, we use the InfoNCE loss to perform comparative learning between the main view and the comparison view at the same level. The calculation is as follows:

[0089]

[0090] where τ is the temperature coefficient and a represents a single element in the set.

[0091] S3, based on the interaction information between the user and the item, predict the user's behavioral preferences through the optimized graph neural network, and sort the user's future behaviors according to the preferences.

[0092] The goal of the recommendation system is to predict user behavior preferences based on user interaction behaviors and rank users' future behaviors based on their preferences. The Bayesian personalized ranking loss (BPRloss) is used to predict user interaction behaviors. The goal is to maximize the predicted preference probability of items that users actually interact with (positive samples) relative to items that they do not interact with (negative samples). The calculation process is as follows:

[0093]

[0094] σ represents the activation function, represents the prediction of positive sample interaction, Represents the prediction of the interaction of negative samples, and the goal is to widen the gap between positive and negative samples as much as possible.

[0095] The overall optimization goal of model training includes considering the BRP loss of the recommendation system prediction task and the InfoNCE loss of three-way contrastive learning. Finally, the losses of each part are added together to obtain the total loss function of model optimization, which is calculated as follows:

[0096]

[0097] Among them, λ1, λ2, and λ3 represent the weight coefficients of contrastive learning loss, and Θ represents the regularization term of the model.

[0098] S4, in order to verify our proposed financial transaction behavior recommendation prediction method based on message masking and cross-view contrastive learning, the model effect is verified on three datasets: Yelp, Amazon and Ifashion through four graph learning recommendation algorithms: LightGCN, SGL, SimGCL, NCL, and LightGCL. The data is divided into training set and validation set.

[0099] The present invention also provides a financial transaction behavior recommendation and prediction system based on message masking and cross-view contrastive learning, comprising:

[0100] The comparison view construction module constructs a user-side comparison view through a graph neural network (GNN) based on the interaction information between the user and the item in the main view.

[0101] The graph neural network optimization module performs cross-view multi-level comparative learning on the main view and the comparison view, thereby deeply mining the potential semantic correlation between users and objects, improving the expressive power of the graph neural network, and obtaining an optimized graph neural network.

[0102] The user behavior prediction module predicts the user's behavior preferences based on the interaction information between the user and the item through the optimized graph neural network, and sorts the user's future behavior according to the preferences.

[0103] The validity verification module verifies the model effect on three datasets: Yelp, Amazon, and Ifashion, using four graph learning recommendation algorithms: LightGCN, SGL, SimGCL, NCL, and LightGCL. The data is divided into training set and validation set.

[0104] Those skilled in the art will appreciate that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A financial transaction behavior recommendation and prediction method based on message masking and cross-view contrastive learning, characterized by: The specific steps include: S1, based on the interaction information between users and items in the main view, constructs a user-side comparative view through a graph neural network (GNN); S2, performing cross-view multi-level comparative learning on the main view and the comparison view, thereby deeply mining the potential semantic correlation between users and items, improving the expressive power of the graph neural network, and obtaining an optimized graph neural network; as well as S3, based on the interaction information between the user and the item, predict the user's behavioral preferences through the optimized graph neural network, and sort the user's future behaviors according to the preferences.

2. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 1, characterized in that: in, In the graph-based recommendation method, the input user set is U = {u1,u2,...,u n }, the input item (behavior) set is I = {i1,i2,...,i n }, the interaction matrix between the two is A. In graph contrastive learning, the core operating mechanism of the graph neural network is message passing, where a node sends messages to its neighbors and receives aggregated messages from its neighbors. The graph convolutional network (GCN) is a classic model in graph neural networks. It updates its own representation by aggregating information from neighboring nodes. Its aggregation update formula can be summarized as: in, represents the hidden embedding of the l+1th layer node u, σ represents the activation function, N u represents the neighbors of node u, e i,u represents the edge between node u and node i, AGG(·) represents the aggregation process of the graph neural network, and f(·) represents the message passing mechanism of the graph neural network.

3. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 1, characterized in that: in, The financial transaction behavior recommendation and prediction method based on message masking and cross-view contrastive learning also includes the following validity verification method: S4, through the four graph learning recommendation algorithms LightGCN, SGL, SimGCL, NCL, and LightGCL, the model effect is verified on three datasets: Yelp, Amazon, and Ifashion. The data is divided into training set and validation set.

4. The financial transaction behavior recommendation and prediction method based on message masking and cross-view contrastive learning according to claim 2, Its characteristics are: in, The step S1 includes the following sub-steps: S1-1, based on the interaction information between users and items in the main view, input the user-item interaction matrix as the original adjacency matrix, and approximate the original adjacency matrix through random SVD decomposition to obtain a reconstructed adjacency matrix; S1-2, randomly masking the information of the reconstructed adjacency matrix through a graph neural network to obtain a mask matrix, and constructing a user-side comparison view through the reconstructed adjacency matrix and the mask matrix.

5. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 4, characterized in that: in, The step S1-1 further includes the following sub-steps: S1-1-1, direct factorization of the matrix will produce a very high computational cost. Before SVD decomposition, the original matrix is first approximated with a q-order low-rank matrix. After obtaining the approximate low-rank matrix, the matrix decomposition is performed. The calculation process is expressed as: in represents an orthogonal matrix, represents a diagonal matrix containing only q singular values of the original matrix, where q is the largest singular value; S1-1-2, after approximate SVD decomposition, the reconstructed adjacency matrix can be obtained for:

6. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 4, characterized in that: in, The step S1-2 further includes the following sub-steps: S1-2-1, using a matrix B that follows a Bernoulli distribution, randomly masks the matrix information to obtain a new matrix, and uses scaling to ensure that the expected values of the matrix before and after masking are equal. The calculation is as follows: Where M represents the original message aggregation matrix, represents the newly generated message aggregation matrix, p represents the masking rate, ⊙ represents the element-wise matrix multiplication, and B represents the matrix with only 0 and 1 values obtained from Bernoulli sampling; S1-2-2 combines singular value decomposition with random information masking to construct the user-side comparison view. The calculation is as follows: Where u represents a user, i represents an item, AGG(.) represents the message aggregation process of the graph convolutional network, and l represents the lth layer of the graph convolutional network. The comparative view calculation on the item side follows a similar principle:

7. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 2, characterized in that: in, The step S2 includes the following sub-steps: S2-1, in the first comparative interaction process, consider the semantic similarity between isomorphic neighbors in the comparative view and the main view. Since the user-item interaction graph is a bipartite graph, users aggregate and update themselves from the features of the interacted items. Therefore, it can be concluded that the embedding types of nodes at corresponding positions in the even layers are the same. The consideration of the odd layers is similar to that of the even layers. The same nodes should have more similar semantic information and node features. Therefore, the cross-view semantic information is combined to mine the similarity between the contrast view and the main view. Therefore, contrastive learning is performed between the embeddings of the lth layer in the contrast view and the (l-2)th layer in the main view, where the same nodes in different views and layers are regarded as positive sample pairs, and nodes in different positions are regarded as negative sample pairs. The calculation is as follows: in, represents the user embedding of the corresponding layer of the contrast view, represents the user embedding of the corresponding layer of the main view, s() represents the similarity calculation function, τ1 represents the temperature coefficient, and U represents other users in the user set except u. The loss calculation method on the item side is similar to it: The loss formula calculation process for the first comparative interaction is: S2-2, in the second comparative interaction process, explore the potential preference associations between users and items in different views, In a recommendation system, when a user interacts with an item, the user often has similar characteristics to other neighbors who interact with the item. According to the assumption of the recommendation system, similar users tend to interact with the same items. Therefore, the embedding matrix of the even-numbered contrast view on the user side is compared with the embedding matrix of the odd-numbered main view on the item side for comparative learning. This method can discover high-order hidden relationships between users and items. Through comparative learning, similar node pairs are brought closer to each other, and dissimilar node pairs are pulled apart. The calculation process of the user-side comparative learning loss is as follows: The meanings of the parameters in this formula are the same as before. Similarly, the calculation process of item-side contrastive learning loss is as follows: Finally, the second contrastive learning loss between users and items is calculated as follows: S2-3, in the interaction graph between users and items, nodes are embedded with ID information, so they lack rich attribute information. Therefore, in the process of convolution of the graph convolutional network, discarding feature transformation, activation function and self-loop process can reduce the redundancy of the algorithm and improve the performance of the algorithm. Therefore, we simply sum the embeddings of each layer of the main view and the contrast view to serve as the output of the final graph convolutional network layer. The summation process is calculated as follows: Where E represents the feature embedding matrix of the main view, and G represents the feature embedding matrix of the comparison view; In S2-4, after mining the potential high-level information between users and between users and items, we use the InfoNCE loss to perform comparative learning between the main view and the comparison view at the same level. The calculation is as follows: where τ is the temperature coefficient and a represents a single element in the set.

8. The method for recommending and predicting financial transaction behaviors based on message masking and cross-view contrastive learning according to claim 2, characterized in that: in, The step S3 includes the following sub-steps: The goal of the recommendation system is to predict user behavior preferences based on user interaction behaviors and rank users' future behaviors based on their preferences. The Bayesian personalized ranking loss (BPRloss) is used to predict user interaction behaviors. The goal is to maximize the predicted preference probability of items that users actually interact with (positive samples) relative to items that they do not interact with (negative samples). The calculation process is as follows: σ represents the activation function, represents the prediction of positive sample interaction, Represents the prediction of the interaction of negative samples, the goal is to widen the gap between positive and negative samples as much as possible, The overall optimization goal of model training includes considering the BRP loss of the recommendation system prediction task and the InfoNCE loss of three-way contrastive learning. Finally, the losses of each part are added together to obtain the total loss function of model optimization, which is calculated as follows: Among them, λ1, λ2, and λ3 represent the weight coefficients of contrastive learning loss, and Θ represents the regularization term of the model.

9. A financial transaction behavior recommendation and prediction system based on message masking and cross-view contrastive learning, characterized by: include: The comparison view construction module uses a graph neural network (GNN) to construct a user-side comparison view based on the interaction information between users and items in the main view; A graph neural network optimization module performs cross-view multi-level comparative learning on the main view and the comparison view, thereby deeply mining the potential semantic correlation between users and items, improving the expression ability of the graph neural network, and obtaining an optimized graph neural network; and The user behavior prediction module predicts the user's behavior preferences based on the interaction information between the user and the item through the optimized graph neural network, and sorts the user's future behavior according to the preferences.

10. The financial transaction behavior recommendation and prediction system based on message masking and cross-view contrastive learning according to claim 9 is characterized in that: Also includes: The validity verification module verifies the model effect on three datasets: Yelp, Amazon, and Ifashion, using four graph learning recommendation algorithms: LightGCN, SGL, SimGCL, NCL, and LightGCL. The data is divided into training set and validation set.

Images