Product recommendation method, system and equipment and storage medium
Through the combination of dual-channel feature extraction and the multi-head attention mechanism, the shortcomings of recommendation algorithms when users' investment behaviors change dynamically in the existing technology are solved, and more efficient and personalized financial product recommendations are achieved.
Patent Information
- Application Number
- CN202411739845.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-29
- Publication Date
- 2025-03-18
AI Technical Summary
When the existing technology deals with the dynamically changing investment behavior of users, the accuracy and personalization of the recommendation algorithm are insufficient, and the model training is difficult, and it depends on specific features, which reduces the generalization ability and has high demand for computing resources.
The dual-channel feature extraction method is adopted to process and integrate user characteristics and financial product characteristics through the multi-head self-attention layer and the multi-head cross-attention layer respectively to enhance the internal correlation and complementary information of the features, and reduce dependence on a large amount of labeled data.
It improves the accuracy and personalization of recommendations, enhances the generalization ability and training efficiency of the model, and reduces the computing resource requirements.
Smart Images
Figure CN119648334A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a product recommendation method, system, device and storage medium. Background Art
[0002] With the development of science and technology, personalized recommendation systems play an increasingly important role in the field of financial product recommendation. Traditional recommendation algorithms, such as collaborative filtering and content-based recommendations, can meet user needs to a certain extent, but they perform poorly when dealing with users' dynamically changing investment behaviors. To solve this problem, the dilated convolution model based on convolutional neural networks was introduced into the recommendation system, attempting to improve the accuracy and personalization of recommendations by constructing a time series of user investment behaviors.
[0003] Although convolutional neural networks can extract features from time series, due to the relatively small amount of feature information, the model may not be able to identify which features are more critical to the final recommendation decision. This not only increases the difficulty of model training, but may also cause the model to over-rely on certain specific features, reducing the generalization ability of the model. In order to overcome the problem of insufficient feature information, the model usually requires a large amount of labeled data for training. This not only increases the cost of data collection and preprocessing, but also prolongs the time for model training. In addition, for complex network structures, the demand for computing resources will also increase significantly. Summary of the invention
[0004] Based on the above problems, the present application provides a product recommendation method, system, device and storage medium.
[0005] The embodiments of the present application disclose the following technical solutions:
[0006] The first aspect of the present application discloses a product recommendation method, comprising:
[0007] Acquire a target data set, wherein the target data set includes data information related to user characteristics and characteristics of the financial product to be recommended;
[0008] Perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding;
[0009] Based on the multi-head self-attention layer, user characteristics and financial product characteristics are processed separately to enhance the internal correlation of user characteristics and financial product characteristics.
[0010] A multi-head cross attention layer is used to fuse user features and financial product features to obtain the first interactive feature sequence after the user feature data affects the financial product feature data, and the second interactive feature sequence after the financial product feature data affects the user feature data;
[0011] Based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer, a recommendation result of the model is determined.
[0012] In a possible implementation, the dual-path feature extraction of the target data to obtain the feature-encoded user features and financial product features includes:
[0013] Based on the embedding layer, two-way feature extraction is performed on the target data set to obtain user features and financial product features after feature encoding. The embedding layer is used to extract features from the target data set using a gated recurrent unit GRU, and the feature encoding result is controlled by adjusting the number of features and the number of recurrent layers of the hidden state of the GRU.
[0014] In a possible implementation, the model is obtained by training based on a training set, and the training set is determined in the following manner:
[0015] Get preliminary data;
[0016] Based on the prepared data, extracting user characteristic data and financial product characteristic data;
[0017] Determine the interest level of the user characteristic data and the financial product characteristic data respectively in combination with the user operation information;
[0018] The user feature data, the financial product feature data and the corresponding interest level are associated and stored to obtain a training set.
[0019] In a possible implementation, the training process of the model includes:
[0020] Get the training set;
[0021] Divide the training set into N parts equally, where N is a positive integer;
[0022] Use NM copies of the training set to train the model, and use the remaining M copies of the training set to evaluate the accuracy of the model, and repeat the training and evaluation process N times until all N copies of the training set have completed the training and evaluation of the model;
[0023] Calculate the average accuracy of N evaluations to get the final accuracy of the model.
[0024] In a possible implementation, the method further includes:
[0025] Get initial data;
[0026] Performing data cleaning on the initial data to obtain cleaned data;
[0027] Randomly sample the cleaned data to obtain preliminary data.
[0028] In a possible implementation, after performing dual-path feature extraction on the target data set to obtain feature-encoded user features and financial product features, the method further includes:
[0029] Determine the importance of each feature in the user features and financial product features after feature encoding;
[0030] Eliminate the user features and financial product features after feature coding whose importance does not meet the preset requirements.
[0031] In a possible implementation, obtaining a recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence, and the statistical pooling layer includes:
[0032] averaging the first interactive feature sequence and the second interactive feature sequence and performing concatenation to obtain a comprehensive feature vector;
[0033] The comprehensive feature vector is sent to the statistical pooling layer, and the recommendation probability is generated through the fully connected layer;
[0034] Based on the recommendation probability corresponding to the financial products, the recommendation results of the model are obtained.
[0035] The second aspect of the present application discloses a product recommendation system, comprising:
[0036] An acquisition unit, configured to acquire a target data set, wherein the target data set includes data information related to user characteristics and characteristics of the financial product to be recommended;
[0037] An extraction unit, used to perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding;
[0038] An enhancement unit is used to process user features and financial product features respectively based on a multi-head self-attention layer, and enhance the internal correlation of user features and financial product features;
[0039] A fusion unit, used to fuse user features and financial product features using a multi-head cross attention layer to obtain a first interactive feature sequence after the user feature data affects the financial product feature data, and a second interactive feature sequence after the financial product feature data affects the user feature data;
[0040] A determination unit is used to determine a recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer.
[0041] Compared with the prior art, this application has the following beneficial effects:
[0042] Obtaining the target data set ensures that the model can obtain the characteristics of users and financial products from multiple perspectives, thereby improving the accuracy of recommendations. By performing dual-path feature extraction on the target data set, the user characteristics and financial product characteristics after feature encoding are obtained respectively. Dual-path feature extraction ensures that the model obtains feature information from different paths, thereby capturing the characteristics of users and financial products more comprehensively. At the same time, the feature encoding process can enhance the expressiveness of the features, which helps the model identify which features are more critical for recommendation decisions. Using a multi-head self-attention layer to process user characteristics and financial product characteristics separately, enhance the internal correlation of the characteristics of user characteristics and financial product characteristics. The self-attention mechanism can automatically learn the relationship between different features, thereby helping the model to better understand and utilize these features. The multi-head design further enhances the expressiveness and flexibility of the model, enabling the model to capture the importance of features from multiple perspectives. The multi-head cross-attention layer is used to fuse user characteristics and financial product characteristics, and the first interactive feature sequence after the user feature data affects the financial product feature data, and the second interactive feature sequence after the financial product feature data affects the user feature data, is obtained. The cross-attention mechanism can capture the interaction between user characteristics and financial product characteristics, thereby generating richer interactive features. This not only improves the generalization ability of the model, but also enables the model to better adapt to different users and market environments. Based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer, the recommendation results of the model are determined. The statistical pooling layer can aggregate the interactive features and extract the most important information, thereby reducing the dependence on a large amount of labeled data. At the same time, the pooling operation can reduce the complexity of the model, reduce the demand for computing resources, and improve the training efficiency and generalization ability of the model. The recommendation results are determined by acquiring multi-source data, dual-path feature extraction, multi-head self-attention layer processing, multi-head cross-attention layer fusion and statistical pooling layer, which effectively solves the problem of single feature information and high computing resource requirements in the background technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0044] Figure 1 A flowchart of a product recommendation method provided in an embodiment of the present application;
[0045] Figure 2 A schematic diagram of the construction process provided for an embodiment of the present application;
[0046] Figure 3 A network structure diagram provided for an embodiment of the present application;
[0047] Figure 4 A schematic diagram of a cross-validation process provided in an embodiment of the present application;
[0048] Figure 5 A structural diagram of a product recommendation system provided in an embodiment of the present application. DETAILED DESCRIPTION
[0049] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0050] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the professional terms involved in the embodiments of the present application will be explained below.
[0051] The proposal of the attention mechanism is mainly inspired by human visual attention. That is, although the visual range of the human eye is relatively wide, the focus is often only on a small area, also known as the region of interest. When the human brain processes information, it will naturally extract useful information from the visual input and focus on some important parts of the input information. In essence, the attention mechanism relies on this idea to assign different weights to all information according to the importance of the input information. In this way, when processing information, the model can focus on important information with high weights and ignore unimportant information to help the model make decisions.
[0052] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.
[0053] As mentioned above, most existing financial product recommendation methods rely on manual experience to recommend popular financial products. This recommendation method is too subjective and cannot achieve personalized recommendations for users. With the advent of the big data era, although some data-based recommendation algorithms continue to emerge, such as collaborative filtering algorithms, they have shortcomings such as sparsity problems, cold start problems, scalability problems, and inability to integrate semantic information.
[0054] In addition, although the dilated convolutional neural network-based model attempts to solve the problems of traditional recommendation algorithms by constructing a time series of user investment behavior, it also has defects. Specifically, the model mainly relies on the characteristics of user investment behavior to construct a time series, ignoring other important factors that may affect user decision-making. For example, information such as the user's financial status, investment preferences, and market trends is not fully utilized, resulting in the recommendation model capturing relatively single information, which affects the accuracy of the recommendation.
[0055] Although convolutional neural networks can extract features from time series, due to the relatively small amount of feature information, the model may not be able to identify which features are more critical to the final recommendation decision. This not only increases the difficulty of model training, but may also cause the model to over-rely on certain specific features, reducing generalization capabilities.
[0056] In order to overcome the problem of insufficient feature information, the model usually requires a large amount of labeled data for training, which not only increases the cost of data collection and preprocessing, but also prolongs the time of model training. In addition, for complex network structures, the demand for computing resources will also increase significantly.
[0057] When the market environment changes or user behavior patterns shift, models based on fixed features may not be able to adjust in time, resulting in a decline in recommendation effectiveness. Especially in financial markets, users’ investment preferences and market conditions may change rapidly, which places higher demands on the real-time and adaptability of recommendation systems.
[0058] In summary, although the collaborative filtering algorithm and the dilated convolution model based on convolutional neural network in the existing technology have solved some basic problems of the recommendation system to a certain extent, they still have many limitations when facing large-scale and highly complex data.
[0059] In order to solve the above problems, an embodiment of the present application provides a product recommendation method, which extracts user features and financial product features respectively through dual-path feature extraction, uses a multi-head self-attention layer to capture the autocorrelation within each feature path, and then fuses user features and financial product features through a multi-head cross-attention layer, deeply explores the potential complementary information between features, and uses a weight matrix to update each feature path, so that the fused feature vector comprehensively includes the advantages of user features and financial product features, thereby improving the accuracy of the financial product recommendation model.
[0060] The user features and financial product features are extracted separately through dual-path feature extraction, ensuring the comprehensiveness and pertinence of feature representation. The multi-head self-attention layer is used to capture the autocorrelation within each feature path, effectively highlighting the features that are more critical to the recommendation decision and improving the accuracy of feature representation. The multi-head cross-attention layer is used to fuse user features and financial product features, realizing deep complementary information mining between features. This mechanism not only considers the importance of a single feature, but also emphasizes the correlation and interaction between different features, enhancing the expressive power of the model. Through the multi-head self-attention and multi-head cross-attention mechanisms, the complex relationship between users and financial products can be captured from multiple angles and levels, improving the accuracy and recommendation effect of the recommendation model. In the process of feature fusion, the weight matrix is used to update each feature path, so that the fused feature vector fully includes the advantages of user features and financial product features, further improving the accuracy and personalization of recommendations.
[0061] It should be noted that the product recommendation method, system, device and medium provided in this application can be applied to the field of computer technology. The above is only an example and does not limit the application field of the product recommendation method, system, device and medium provided in this application. In addition, the embodiment of the present application may not limit the execution subject of the product recommendation. For example, the product recommendation method of the embodiment of the present application can be applied to data processing devices such as terminal devices or servers. Among them, the terminal device can be an electronic device such as a computer, a personal digital assistant (PDA). The server can be a stand-alone server, a cloud server, or a cluster server composed of multiple servers.
[0062] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0063] The following is an example to illustrate a product recommendation method provided by the present application. Figure 1 ,Should Figure 1 A flowchart of a product recommendation method provided in an embodiment of the present application, the method comprising:
[0064] S101: Obtain a target data set.
[0065] The target data set includes data information related to user characteristics and characteristics of the financial products to be recommended. User information includes but is not limited to age, gender, occupation, income level, investment preferences, etc.; financial product information covers product type, expected rate of return, investment period, risk rating, etc. This information is collected through legal channels to ensure data security and privacy. In addition, the user's transaction history is also collected, such as the financial products purchased, investment amount, holding time, etc., to more comprehensively understand the user's investment habits and preferences.
[0066] S102: Perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding.
[0067] After obtaining the target data set, the next step is to preprocess and extract features from the data. Two independent feature extraction paths are designed for user features and financial product features. Each path uses a different encoder model (for example, a neural network model) to process the corresponding features, and the final output is a high-dimensional feature vector that can reflect the core information of the original data.
[0068] For user characteristics, a neural network with an embedding layer can be used to capture the association between user attributes; and for financial product characteristics, a convolutional neural network (CNN) can be used to extract the static and dynamic characteristics of financial products.
[0069] S103. Based on the multi-head self-attention layer, the user characteristics and the financial product characteristics are processed separately to enhance the internal correlation of the user characteristics and the financial product characteristics.
[0070] After feature extraction, the multi-head self-attention mechanism is applied to further optimize user features and financial product features. The multi-head self-attention mechanism can help the model focus on different parts of the input features from different angles, thereby enhancing the ability of feature representation. For user features, this process can better understand the user's behavior patterns and preferences; for financial product features, it helps to discover the potential connections between different financial products.
[0071] S104. Use a multi-head cross attention layer to fuse user features and financial product features to obtain a first interactive feature sequence after the user feature data affects the financial product feature data, and a second interactive feature sequence after the financial product feature data affects the user feature data.
[0072] In order to effectively integrate user features and financial product features, a multi-head cross attention mechanism can be used. In this step, user features and financial product features interact with each other to generate two sets of interactive feature sequences. The first interactive feature sequence reflects how user features affect their cognition of financial product features, while the second interactive feature sequence reflects how financial product features in turn affect users' decision-making tendencies. This two-way information exchange helps the model more accurately predict products that users may be interested in.
[0073] S105. Determine a recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence, and the statistical pooling layer.
[0074] Finally, based on the above two sets of interactive feature sequences, a statistical pooling layer is used to further refine the features, reducing the dimensions while retaining the most important information. Subsequently, these features are converted into specific recommendation scores through one or more fully connected layers. These scores are used to measure the user's interest in each financial product. The recommendation model can sort these scores and recommend the most suitable financial products to users.
[0075] The following is an explanation of the model construction process. Figure 2 , Figure 2 A schematic diagram of the construction process provided for an embodiment of the present application.
[0076] The first stage is data preprocessing. In the process of building a recommendation model, the data set is the basis for building the model, and the quality of the data affects the training effect of the model. Therefore, in order to ensure the subsequent training of the recommendation model, data preprocessing is essential. The preprocessing steps are mainly divided into three steps: data cleaning, data sampling, and data set construction.
[0077] In a possible implementation, initial data is obtained; data cleaning is performed on the initial data to obtain cleaned data; random sampling is performed on the cleaned data to obtain preliminary data. Preliminary data is obtained; user characteristic data and financial product characteristic data are extracted based on the preliminary data; interest levels of the user characteristic data and financial product characteristic data are determined in combination with user operation information; the user characteristic data, financial product characteristic data and corresponding interest levels are associated and stored to obtain a training set.
[0078] First, we obtain initial data from multiple sources, which may include but are not limited to user activity records of online financial platforms, third-party market research reports, publicly available financial product databases, etc. The initial data covers a wide range of user information and financial product information, such as user personal information, browsing history, purchase history, as well as detailed descriptions and historical performance of financial products.
[0079] After obtaining the initial data, the next step is to clean the data. The purpose of data cleaning is to remove noise and irrelevant information and improve data quality. First, data cleaning is performed to remove noise data and outliers, fill in missing values, and ensure data quality by performing operations such as deduplication, filtering, normalization, and processing outliers. Reliable data is filtered out from the original data to provide a data basis for subsequent data processing. Data cleaning operations include:
[0080] Deduplication: Remove duplicate items from the data to avoid adverse effects of duplicate data on the model;
[0081] Filtering: removing useless data, such as incomplete or erroneous data;
[0082] Normalization: Make the numerical range of data within a certain range to avoid adverse effects caused by numerical deviation;
[0083] Dealing with outliers: Data that deviates from normal values needs to be corrected or eliminated.
[0084] In one possible implementation, after data cleaning, a random sampling method can be used to extract training sets and test sets from the cleaned data to reduce the risk of overfitting. Usually, the ratio of training sets and test sets can be adjusted according to actual conditions, for example, 70% of the data is used for training and 30% of the data is used for testing.
[0085] Then, the data set is constructed. Based on the prepared data, data mining and machine learning techniques are used to extract user feature data and financial product feature data. User feature data may include the user's age, gender, income level, investment preferences, etc.; financial product feature data may cover product type, expected rate of return, risk level, etc. In the financial product recommendation system, the interaction data between users and financial products is saved one by one in the database. These interaction data need to be processed to divide the input original user features and financial product feature data. Therefore, it is necessary to extract user feature data and financial product feature data from the database.
[0086] In addition, the calculation of interest rating is based on the user's historical purchase behavior, browsing information and operation data. That is, the user-financial product interest rating is combined with the income obtained by the user from purchasing financial products, the user's browsing information and operation data, etc., and the rating is used as the output of the model. That is, the input of the model training is a certain user feature and a certain financial product feature, and the output is the user's interest in the financial product. The interactive data in the database is constructed into such an input and output data set for model training and prediction. Finally, the user feature data, financial product feature data and the corresponding interest level are associated to construct a training set for training the recommendation model. Each training sample contains the features of at least one user, the features of at least one financial product and the user's interest level in this product. The data set format is shown in Table 1.
[0087] Table 1. User-financial product interest dataset
[0088] Enter user characteristics Enter financial product features Interest rating uf11,uf12,…,uf1n pf11,pf12,…,pf1m Level 1 uf21,uf22,…,uf2n pf21,pf22,…,pf2m Level2 … … … ufi1,ufi2,…,ufin pfi1,pfi2,…,pfim level
[0089] Among them, ufi1, ufi2, …, ufin represent the first user feature, the second user feature to the nth user feature of the i-th data; pfi1, pfi2, …, pfim represent the first financial product feature, the second financial product feature to the m-th financial product feature of the i-th data; leveli represents the interest rating of the i-th data, where the interest rating is a number between 0.0 and 1.0, such as 0.2, 0.3, 0.4, and so on.
[0090] The second stage is feature extraction. The key to personalized recommendation is how to accurately describe the user's interests, and feature extraction is an important means to achieve this goal. The goal of feature extraction is to extract features that can effectively describe the user's interests and behaviors from the original data.
[0091] For the constructed data set, different features need to be evaluated and key features need to be screened out. Feature selection is an indispensable part of the personalized recommendation system, which includes two aspects: feature extraction and feature dimensionality reduction. The first aspect is feature extraction, which extracts useful features from the original data. This step can be achieved through a variety of methods, such as statistical analysis, machine learning algorithms, etc. The second aspect is feature dimensionality reduction, which reduces the computational complexity and improves the generalization ability of the model by reducing the number of features. Common feature dimensionality reduction methods include principal component analysis (PCA), linear discriminant analysis (LDA), etc.
[0092] Therefore, feature selection can improve data validity, screen out features that are highly correlated with the target variable, and improve the prediction accuracy of the model. On the other hand, it can reduce computational complexity, reduce the number of features, and reduce the time and resource consumption of model training and reasoning. And by removing redundant and noisy features, it can improve the generalization ability and stability of the model.
[0093] Feature extraction is mainly achieved through the embedding layer. The embedding layer is a method that converts high-dimensional raw feature data into low-dimensional dense vectors, which can better capture the relationship between features. The specific implementation of the embedding layer is as follows:
[0094] GRU encodes input features through its internal gating mechanism (update gate and reset gate) to generate hidden state vectors. These hidden state vectors can capture long-term dependencies in sequence data. In the actual model training, it is necessary to dynamically adjust the number of features and loop layers of the GRU's hidden state according to the training effect of the model. By changing these parameters, the result of feature encoding can be controlled to better adapt to different data sets and application scenarios.
[0095] Use a GRU to encode user features and generate a user feature vector. User features may include the user's age, gender, and historical purchase records. Use a GRU to encode user features and generate a user feature vector. User features may include the user's age, gender, and historical purchase records. Fuse the user feature vector with the financial product feature vector to form the final input feature vector for subsequent recommendation model training.
[0096] The third stage is model construction. The network structure diagram of the model is as follows Figure 3 As shown, Figure 3 The network structure diagram provided for the embodiment of the present application mainly includes the following key parts:
[0097] Input layer: The input of the model is divided into two parts: original user characteristics and financial product characteristics. User characteristics may include the user's age, gender, occupation, income level, historical purchase records, etc. Financial product characteristics may include product type, rate of return, risk level, investment period, etc.
[0098] Embedding layer: The input user features and financial product features are first extracted through the embedding layer. The embedding layer converts the high-dimensional raw feature data into a low-dimensional dense vector to better capture the relationship between features. The embedding layer is mainly implemented by the gated recurrent unit (GRU), which encodes the input features through its internal gating mechanism to generate a hidden state vector.
[0099] Fully connected layer: The feature data processed by the embedding layer is further mapped through the fully connected layer. The fully connected layer maps the feature data to the label space of the feature data sample and generates a feature vector.
[0100] Multi-head self-attention layer: The encoded feature vector is fed into the multi-head self-attention layer. The main purpose of the multi-head self-attention mechanism is to find the autocorrelation within the single-channel feature data and enhance the representation of useful information in the feature data. Through multiple attention heads, the model can capture the relationship between features from different angles and improve the model's expressiveness.
[0101] Since user characteristics and financial product characteristics belong to two different dimensions, the interaction of the characteristics of these two dimensions can obtain the user's interest in the financial product. The present invention realizes the interaction of different feature data through a multi-head cross attention layer.
[0102] For user feature data, Query(q) is the user feature data itself. Key(k) and Value(v) are the financial product feature data. The important features in the financial product feature data k are queried through the user feature data q, and the corresponding weight matrix is obtained. Then, it is multiplied by the financial product feature data v to obtain the interactive feature sequence after the user feature data affects the financial product feature data.
[0103] For the characteristic data of financial products: Query (q) is the characteristic data of the financial products itself. Key (k) and Value (v) are the characteristic data of the users. The important features in the characteristic data k of the user are queried through the characteristic data q of the financial products, and the corresponding weight matrix is obtained. Then, the weight matrix is multiplied by the user characteristic data v to obtain the interactive characteristic sequence after the influence of the characteristic data of the financial products on the user characteristic data. The two interactive characteristic sequences are averaged (Mean) and then concatenated and sent to the statistical pooling layer.
[0104] The statistical pooling layer is used to further extract the statistical information of the features, such as mean, variance, etc., to enhance the robustness of the model. Finally, the recommendation result of the model is obtained through formula (1).
[0105]
[0106] Among them, μ and σ in formula 1 represent the mean and standard deviation obtained by the statistical pooling layer, which means Cascade operator, f c represents the fully connected layer, the SoftMax layer is used to generate the recommendation probability for the feature data set, and y^ represents the result of the model recommendation.
[0107] The fourth stage is model training. The constructed recommendation model needs to be fully trained to achieve the expected performance. During the training process, we use the constructed data set to train the model. In order to effectively avoid the overfitting or underfitting state of the model, the present invention adopts a variety of strategies:
[0108] Standardize or normalize user features and financial product features to ensure that the features are on the same scale, which helps accelerate model convergence. Handle missing values, fill or delete missing data, and ensure data integrity. Add L1 or L2 regularization terms to the model to prevent the model from being too complex and causing overfitting. Adjust the regularization parameters to balance the complexity and generalization ability of the model. During the training process, if the performance on the validation set no longer improves, terminate the training early to avoid overfitting. Use methods such as grid search or random search to find the best hyperparameter combination to improve model performance.
[0109] In one possible implementation, a training set is obtained for the model; the training set is divided into N parts on average, where N is a positive integer; the model is trained using NM parts of the training set, and the remaining M parts of the training set are used to evaluate the accuracy of the model, and the training and evaluation process is repeated N times until the N parts of the training set have completed training and evaluation of the model; the average of the accuracies obtained from the N evaluations is calculated to obtain the final accuracy of the model.
[0110] For example, the value of N can be 5, and the value of M can be 1. That is, when evaluating the model, the 5-fold cross-validation method is used. The principle is to divide the original feature data set into 5 parts, use 4 of them as training sets to train the model, and use the remaining 1 part as a test set to evaluate the model, and repeat the above process 5 times. The advantage of this method is that it does not require additional data splitting to avoid data waste, and it can prompt the model to learn feature data from multiple aspects to avoid the model falling into local extreme values. The 5-fold cross-validation process is as follows: Figure 4 As shown, Figure 4 The schematic diagram of the cross-validation process provided in the embodiment of the present application is as follows: The average value of the recommended accuracy of the model Accuracy of the cross-validation model is calculated 5 times as the final accuracy Acc of the model.
[0111] Therefore, this application has the following beneficial effects:
[0112] Deep learning models can effectively process large-scale, high-dimensional data sets, thanks to their powerful feature learning capabilities. This capability enables deep learning models to automatically extract useful features from large amounts of user behavior data without the need for manual feature design.
[0113] Through deep network structures, deep learning models can capture the complexity and diversity of user interests, such as the relationship between users' long-term preferences and short-term interests, which is crucial for providing more personalized recommendation services. Compared with traditional linear models, deep learning models can establish more complex nonlinear relationships, which helps improve the accuracy of predictions, especially when dealing with nonlinearly changing data.
[0114] Moreover, the model proposed in the present invention mainly relies on the attention mechanism, in which the self-attention mechanism is used to calculate the autocorrelations within user features and financial product features respectively, which is equivalent to giving higher weights to features that are more conducive to decision-making within the two features; secondly, since users interact with financial products to obtain their interest in financial products, user features and financial product features should be fused during the training of the model. The present invention fuses the two features through multi-head interactive attention and learns from each other, so that the fused feature vector has the advantages of both user features and financial product features, thereby achieving efficient and accurate personalized financial product recommendations.
[0115] The above are some specific implementations of the product recommendation method provided in the embodiment of the present application. Based on this, the present application also provides a corresponding product recommendation system. The system provided in the embodiment of the present application will be introduced from the perspective of functional modularization. Figure 5 A structural diagram of a product recommendation system provided in an embodiment of the present application.
[0116] The system comprises:
[0117] An acquisition unit 110 is used to acquire a target data set, wherein the target data set includes data information related to user characteristics and characteristics of the financial product to be recommended;
[0118] An extraction unit 111 is used to perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding;
[0119] An enhancement unit 112 is used to process user features and financial product features respectively based on a multi-head self-attention layer to enhance the internal correlation of user features and financial product features;
[0120] A fusion unit 113 is used to fuse user features and financial product features using a multi-head cross attention layer to obtain a first interactive feature sequence after the user feature data affects the financial product feature data, and a second interactive feature sequence after the financial product feature data affects the user feature data;
[0121] The determination unit 114 is used to determine the recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer.
[0122] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the solutions provided by the embodiments of the present application.
[0123] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the method described in any embodiment of the present application.
[0124] The computer storage medium stores codes, and when the codes are executed, a device executing the codes implements the method described in any embodiment of the present application.
[0125] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.
[0126] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.
[0127] It should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.
[0128] It should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0129] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0130] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.
[0131] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A product recommendation method, characterized in that: include: Acquire a target data set, wherein the target data set includes data information related to user characteristics and characteristics of the financial product to be recommended; Perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding; Based on the multi-head self-attention layer, user characteristics and financial product characteristics are processed separately to enhance the internal correlation of user characteristics and financial product characteristics. A multi-head cross attention layer is used to fuse user features and financial product features to obtain the first interactive feature sequence after the user feature data affects the financial product feature data, and the second interactive feature sequence after the financial product feature data affects the user feature data; Based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer, a recommendation result of the model is determined.
2. The product recommendation method according to claim 1, characterized in that: The dual-path feature extraction of the target data to obtain the user features and financial product features after feature encoding includes: Based on the embedding layer, two-way feature extraction is performed on the target data set to obtain user features and financial product features after feature encoding. The embedding layer is used to extract features from the target data set using a gated recurrent unit GRU, and the feature encoding result is controlled by adjusting the number of features and the number of recurrent layers of the hidden state of the GRU.
3. The product recommendation method according to claim 1, characterized in that: The model is obtained based on training of a training set, and the training set is determined in the following manner: Get preliminary data; Based on the prepared data, extracting user characteristic data and financial product characteristic data; Determine the interest level of the user characteristic data and the financial product characteristic data respectively in combination with the user operation information; The user feature data, the financial product feature data and the corresponding interest level are associated and stored to obtain a training set.
4. The product recommendation method according to claim 3, characterized in that: The training process of the model includes: Get the training set; Divide the training set into N parts equally, where N is a positive integer; Use NM copies of the training set to train the model, and use the remaining M copies of the training set to evaluate the accuracy of the model, and repeat the training and evaluation process N times until all N copies of the training set have completed the training and evaluation of the model; Calculate the average accuracy of N evaluations to get the final accuracy of the model.
5. The product recommendation method according to claim 3, characterized in that: The method further comprises: Get initial data; Performing data cleaning on the initial data to obtain cleaned data; Randomly sample the cleaned data to obtain preliminary data.
6. The product recommendation method according to claim 1, characterized in that: After performing dual-path feature extraction on the target data set to obtain the user features and financial product features after feature encoding, the method further includes: Determine the importance of each feature in the user features and financial product features after feature encoding; Eliminate the user features and financial product features after feature coding whose importance does not meet the preset requirements.
7. The product recommendation method according to claim 1, characterized in that: The method of obtaining a recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer includes: averaging the first interactive feature sequence and the second interactive feature sequence and performing concatenation to obtain a comprehensive feature vector; The comprehensive feature vector is sent to the statistical pooling layer, and the recommendation probability is generated through the fully connected layer; Based on the recommendation probability corresponding to the financial products, the recommendation results of the model are obtained.
8. A product recommendation system, characterized in that: include: An acquisition unit, configured to acquire a target data set, wherein the target data set includes data information related to user characteristics and characteristics of the financial product to be recommended; An extraction unit, used to perform dual-path feature extraction on the target data set to obtain user features and financial product features after feature encoding; An enhancement unit is used to process user features and financial product features respectively based on a multi-head self-attention layer, and enhance the internal correlation of user features and financial product features; A fusion unit, used to fuse user features and financial product features using a multi-head cross attention layer to obtain a first interactive feature sequence after the user feature data affects the financial product feature data, and a second interactive feature sequence after the financial product feature data affects the user feature data; A determination unit is used to determine a recommendation result of the model based on the first interactive feature sequence, the second interactive feature sequence and the statistical pooling layer.
9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the product recommendation method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the product recommendation method according to any one of claims 1 to 7.