Information recommendation method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention provides an information recommendation method and device, equipment, a storage medium and a program product, and relates to the field of artificial intelligence. The method comprises the steps of obtaining attribute information of a to-be-recommended user and behavior data information in a preset time period; respectively inputting the behavior data information into a plurality of long and short-term memory models to obtain interestingness of the to-be-recommended user on various product categories; the attribute information is input into a multi-head self-attention model, and feature representation is obtained and used for representing the probability that the to-be-recommended user depends on the attribute information to select various product categories; the interestingness and the feature representation of the to-be-recommended user on various product categories are input into a full-connection neural network, a product category recommendation result corresponding to the to-be-recommended user is obtained, and a discriminator in a semi-supervised generative adversarial network model is formed by the multiple long and short-term memory models, the multi-head self-attention model and the full-connection neural network. The method is used for achieving the effect of improving the accuracy of product category recommendation results.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to an information recommendation method, device, equipment, storage medium and program product. Background Art

[0002] In recent years, the financial products in banks have become more diverse and updated very quickly. Therefore, having an efficient product recommendation system can not only better promote its own products and reduce unnecessary expenses, but also improve customer satisfaction, thereby helping banks attract more customers and enhance their competitiveness.

[0003] In the prior art, traditional machine learning methods improve the objectivity and accuracy of product recommendations. However, traditional machine learning methods have problems with insufficient understanding of complex data relationships and insufficient adaptability. To this end, deep learning models can use complex networks to mine and understand the complex nonlinear relationships between user behavior data and product feature data, overcoming the difficulties of manual analysis and traditional machine learning in processing large-scale complex data. However, the use of deep learning models to recommend product categories still has the technical problem of low accuracy. Summary of the invention

[0004] The embodiments of the present application provide an information recommendation method, apparatus, device, storage medium and program product to improve the accuracy of product category recommendation results.

[0005] In a first aspect, an embodiment of the present application provides an information recommendation method, comprising:

[0006] Obtaining attribute information of the user to be recommended and behavioral data information within a preset time period;

[0007] Input the behavior data information into multiple long-short term memory models respectively to obtain the interest of the recommended user in various product categories;

[0008] Input the attribute information into the multi-head self-attention model to obtain feature representation, which is used to represent the probability of the recommended user selecting each product category based on the attribute information;

[0009] The interest and feature representation of the user to be recommended in various product categories are input into the fully connected neural network to obtain the product category recommendation results corresponding to the user to be recommended. Among them, multiple long short-term memory models, multi-head self-attention models and fully connected neural networks constitute the discriminator in the semi-supervised generative adversarial network model.

[0010] In a possible implementation manner, before obtaining the attribute information of the user to be recommended and the behavior data information within a preset time period, the method further includes:

[0011] Build the first semi-supervised generative adversarial network model;

[0012] Acquire multiple first training samples and multiple second training samples for training the first semi-supervised generative adversarial network model; the first training samples include: input information and labels corresponding to historical users, the labels are used to indicate the target product category corresponding to the historical users; the second training samples include: input information of historical users; the input information includes attribute information and behavior data information;

[0013] The first semi-supervised generative adversarial network model is trained according to multiple first training samples, multiple second training samples and a random noise vector to obtain a trained semi-supervised generative adversarial network model, wherein the first semi-supervised generative adversarial network model includes a first generator and a first discriminator, the first generator is used to generate pseudo input information according to the random noise vector, the first discriminator is used to distinguish the authenticity of the input information, and determine the corresponding prediction result according to the input information.

[0014] In a possible implementation, constructing a first semi-supervised generative adversarial network model includes:

[0015] Acquire product attribute data, process the product attribute data according to a product attribute clustering model to obtain multiple product categories, and determine second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in any product category;

[0016] For any product category, the first long short-term memory model is trained according to the second behavior data information of each historical user and the interest level of any product category to obtain a long short-term memory model corresponding to any product category after training, wherein multiple product categories correspond to multiple long short-term memory models;

[0017] The first semi-supervised generative adversarial network model is constructed based on multiple long short-term memory models, the first multi-head self-attention model and the first fully connected neural network.

[0018] In a possible implementation, the first long short-term memory model is trained according to the second behavior data information of each historical user and the interest level of any product category to obtain a trained long short-term memory model corresponding to any product category, including:

[0019] Determine a training set and a test set based on the second behavior data information and the interest level of any product category;

[0020] According to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model;

[0021] The second long short-term memory model is tested according to the test set to obtain a long short-term memory model.

[0022] In a possible implementation, according to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model, including:

[0023] Initialize the particle swarm to determine the initial position, speed, and number of iterations of each particle. The initial position is used to represent the initial model parameters of the first long short-term memory model.

[0024] For each particle, the first long short-term memory model is trained according to the initial position and the training set to obtain the prediction error, and the fitness of each particle is calculated according to the prediction error;

[0025] According to the fitness of each particle, the individual extreme value and the global extreme value are determined, and the position and speed of each particle are updated according to the individual extreme value and the global extreme value. The iterative process is repeated until the number of iterations is met and the target position of the particle corresponding to the global extreme value is obtained;

[0026] According to the model parameters corresponding to the target position, a second long short-term memory model is obtained.

[0027] In a possible implementation, training the first semi-supervised generative adversarial network model according to the plurality of first training samples, the plurality of second training samples, and the random noise vector to obtain a trained semi-supervised generative adversarial network model includes:

[0028] Inputting the random noise vector into the first generator in the first semi-supervised generative adversarial network model to generate pseudo input information;

[0029] Inputting the input information in the first training sample into the first discriminator in the first semi-supervised generative adversarial network model to obtain a prediction result, and determining the prediction loss according to the prediction result and the label in the first training sample;

[0030] Inputting the input information in the second training sample into the first discriminator to obtain a first probability value, and obtaining a first adversarial loss according to the first probability value, wherein the first probability value is used to indicate the probability that the input information is true;

[0031] Inputting the pseudo input information into the first discriminator to obtain a second probability value, and obtaining a second adversarial loss according to the second probability value, wherein the second probability value is used to indicate the probability that the pseudo input information is true;

[0032] The first discriminator is updated according to the prediction loss, the first adversarial loss and the second adversarial loss to obtain the second discriminator;

[0033] According to the second adversarial loss, the first generator is updated to obtain the second generator;

[0034] The second generator and the second discriminator are iteratively trained to obtain a semi-supervised generative adversarial network model.

[0035] In a second aspect, an embodiment of the present application provides an information recommendation device, including:

[0036] The first processing module is used to obtain attribute information of the user to be recommended and behavior data information within a preset time period;

[0037] The second processing module is used to input the behavior data information into multiple long-term and short-term memory models respectively to obtain the interest of the recommended user in various product categories;

[0038] The third processing module is used to input the attribute information into the multi-head self-attention model to obtain a feature representation, which is used to represent the probability of the recommended user selecting each product category based on the attribute information;

[0039] The fourth processing module is used to input the interest and feature representation of the user to be recommended in various product categories into the fully connected neural network to obtain the product category recommendation results corresponding to the user to be recommended, wherein multiple long short-term memory models, multi-head self-attention models and fully connected neural networks constitute the discriminator in the semi-supervised generative adversarial network model.

[0040] In a possible implementation manner, the first processing module is further configured to:

[0041] Build the first semi-supervised generative adversarial network model;

[0042] Acquire multiple first training samples and multiple second training samples for training the first semi-supervised generative adversarial network model; the first training samples include: input information and labels corresponding to historical users, the labels are used to indicate the target product category corresponding to the historical users; the second training samples include: input information of historical users; the input information includes attribute information and behavior data information;

[0043] The first semi-supervised generative adversarial network model is trained according to multiple first training samples, multiple second training samples and a random noise vector to obtain a trained semi-supervised generative adversarial network model, wherein the first semi-supervised generative adversarial network model includes a first generator and a first discriminator, the first generator is used to generate pseudo input information according to the random noise vector, the first discriminator is used to distinguish the authenticity of the input information, and determine the corresponding prediction result according to the input information.

[0044] In a possible implementation manner, the first processing module is further configured to:

[0045] Acquire product attribute data, process the product attribute data according to a product attribute clustering model to obtain multiple product categories, and determine second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in any product category;

[0046] For any product category, the first long short-term memory model is trained according to the second behavior data information of each historical user and the interest level of any product category to obtain a long short-term memory model corresponding to any product category after training, wherein multiple product categories correspond to multiple long short-term memory models;

[0047] The first semi-supervised generative adversarial network model is constructed based on multiple long short-term memory models, the first multi-head self-attention model and the first fully connected neural network.

[0048] In a possible implementation manner, the first processing module is further configured to:

[0049] Determine a training set and a test set based on the second behavior data information and the interest level of any product category;

[0050] According to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model;

[0051] The second long short-term memory model is tested according to the test set to obtain a long short-term memory model.

[0052] In a possible implementation manner, the first processing module is further configured to:

[0053] Initialize the particle swarm to determine the initial position, speed, and number of iterations of each particle. The initial position is used to represent the initial model parameters of the first long short-term memory model.

[0054] For each particle, the first long short-term memory model is trained according to the initial position and the training set to obtain the prediction error, and the fitness of each particle is calculated according to the prediction error;

[0055] According to the fitness of each particle, the individual extreme value and the global extreme value are determined, and the position and speed of each particle are updated according to the individual extreme value and the global extreme value. The iterative process is repeated until the number of iterations is met and the target position of the particle corresponding to the global extreme value is obtained;

[0056] According to the model parameters corresponding to the target position, a second long short-term memory model is obtained.

[0057] In a possible implementation manner, the first processing module is further configured to:

[0058] Inputting the random noise vector into the first generator in the first semi-supervised generative adversarial network model to generate pseudo input information;

[0059] Inputting the input information in the first training sample into the first discriminator in the first semi-supervised generative adversarial network model to obtain a prediction result, and determining the prediction loss according to the prediction result and the label in the first training sample;

[0060] Inputting the input information in the second training sample into the first discriminator to obtain a first probability value, and obtaining a first adversarial loss according to the first probability value, wherein the first probability value is used to indicate the probability that the input information is true;

[0061] Inputting the pseudo input information into the first discriminator to obtain a second probability value, and obtaining a second adversarial loss according to the second probability value, wherein the second probability value is used to indicate the probability that the pseudo input information is true;

[0062] The first discriminator is updated according to the prediction loss, the first adversarial loss and the second adversarial loss to obtain the second discriminator;

[0063] According to the second adversarial loss, the first generator is updated to obtain the second generator;

[0064] The second generator and the second discriminator are iteratively trained to obtain a semi-supervised generative adversarial network model.

[0065] In a third aspect, an embodiment of the present application provides an information recommendation device, including: a memory, a processor;

[0066] The memory stores computer-executable instructions;

[0067] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0068] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementations of the first aspect.

[0069] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0070] The information recommendation method, apparatus, device, storage medium and program product provided in the embodiments of the present application construct a semi-supervised generative adversarial network model by utilizing a long short-term memory model, a multi-head self-attention model and a fully connected neural network. The long short-term memory model processes sequence data and can capture the time dependency and long-term dependency in the user behavior data information of the recommended user within the pre-examination time period, and predict the user's interest in different product categories. The multi-head self-attention model can capture the complex relationships in the attribute information and generate feature representations for the product categories by focusing on different feature subsets in the attribute information. The fully connected network further improves the accuracy of the product type recommendation results by capturing the relationship between the user's interest in different product categories and the feature representations generated by the multi-head self-attention model. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0072] Figure 1 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 1 ;

[0073] Figure 2 A schematic diagram of the structure of the long short-term memory model provided in the embodiment of the present application;

[0074] Figure 3 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 2 ;

[0075] Figure 4 A schematic diagram of the product classification results provided in the embodiment of the present application;

[0076] Figure 5 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 3 ;

[0077] Figure 6 A schematic diagram of the structure of a semi-supervised generative adversarial network model provided in an embodiment of the present application;

[0078] Figure 7 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 4 ;

[0079] Figure 8 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 5 ;

[0080] Fig. 9 A schematic diagram of the structure of an information recommendation device provided by this application;

[0081] Fig.10 A schematic diagram of the structure of an information recommendation device provided in this application.

[0082] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0083] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0084] In recent years, with the rapid development of the financial market, banks have launched a wide variety of financial products, and the update speed is getting faster and faster. In this environment, banks are faced with the challenge of how to effectively recommend suitable products to customers. An efficient product recommendation system can not only help banks better promote their products and reduce unnecessary marketing expenses, but also enhance the competitiveness of banks in the market while improving customer satisfaction.

[0085] Traditional product category recommendation systems usually rely on machine learning methods that predict products that customers may be interested in by analyzing historical data. Traditional machine learning methods have improved the objectivity and accuracy of recommendations to a certain extent. However, these methods have limitations when dealing with complex data relationships, especially when faced with multi-dimensional and multi-level user behavior data and product feature data, the adaptability of traditional methods is not strong enough.

[0086] To solve these problems, deep learning models were introduced into product category recommendation systems. Through its complex network structure, deep learning can better mine and understand the complex nonlinear relationship between user behavior data and product feature data. Compared with traditional methods, deep learning performs well in processing large-scale complex data sets, can automatically extract features, and reduce reliance on manual analysis.

[0087] However, despite the excellent performance of deep learning models in many aspects, existing technologies still face some challenges when using deep learning for product category recommendation. In particular, the accuracy of recommendations still needs to be improved. Therefore, how to further optimize deep learning models to achieve higher accuracy in product category recommendations has become an urgent problem to be solved.

[0088] The present application provides an information recommendation method, which constructs a semi-supervised generative adversarial network model by utilizing a long short-term memory model, a multi-head self-attention model, and a fully connected neural network. The long short-term memory model processes sequence data, and can capture the time dependency and long-term dependency in the user behavior data information of the recommended user within the pre-examination time period, and predict the user's interest in different product categories. The multi-head self-attention mechanism can capture the complex relationships in the attribute information, and generate feature representations of the product categories by focusing on different feature subsets in the attribute information. The fully connected network has solved the technical problem of poor accuracy of product type recommendation results in the prior art by capturing the user's interest in different product categories and the relationship between the feature representations generated by the multi-head self-attention model.

[0089] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0090] Figure 1 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 1 ,like Figure 1 As shown, the method includes:

[0091] S101: Acquire attribute information of a user to be recommended and behavior data information within a preset time period.

[0092] In this embodiment, the attribute information of the user to be recommended includes but is not limited to numerical characteristics such as the capital, age, enterprise size, registration time, etc. of the user to be recommended, as well as discrete characteristics such as the category and education level of the user to be recommended. Within a preset time period, various system logs, monitoring tools or special data analysis platforms are used to obtain the behavioral data information of the user to be recommended. For example, it includes the browsing history, search behavior, interactive behavior or consumption record of the user to be recommended on a specific application or platform, as well as the frequency and mode of using specific functions. Among them, the browsing history includes, for example, the pages visited, the dwell time, and the content clicked; the search behavior includes, for example, the searched keywords; and the interactive behavior includes, for example, likes, comments, and shares. By recording and organizing these behavioral data in detail, a comprehensive information set about the behavioral performance of the user to be recommended within the preset time period is formed. This information will provide an important basis for subsequent analysis and recommendation algorithms to achieve more accurate personalized recommendations.

[0093] It should be noted that the attribute information of the user to be recommended and the behavior data information within a preset time period are mostly feature information.

[0094] In one possible implementation, for example, the above-mentioned attribute information and the behavior data information within a preset time period can be preprocessed to improve the quality of the attribute information and the behavior data information within the preset time period. The preprocessing process includes, for example: using one-hot encoding to process discrete data; normalizing numerical data; or cleaning and processing the attribute information and behavior data information, detecting outliers, deleting duplicate data, and standardizing the data. Among them, taking data cleaning and processing as an example, if the missing attribute value of a certain attribute information or behavior data information reaches 1 / 4 of the entire data, the entire data is discarded; if the missing attribute value does not reach 1 / 4, it can be confirmed by manual judgment whether the data is discarded or filled in using K-means. For example, the attribute mean of multiple attribute information or behavior data information with relatively complete attribute values ​​collected in the same time period is selected as the value of the missing attribute field of the data to be filled.

[0095] S102: Input the behavior data information into a plurality of long-term and short-term memory models respectively to obtain the interest of the user to be recommended in various product categories.

[0096] In this embodiment, multiple long-term and short-term memory models process the behavior data information respectively, so as to obtain the interest of the recommended user in each product category. The interest can be, for example, a quantitative indicator, such as a value from 0 to 1, and the higher the value, the greater the interest. For example, it can also be the total number of transactions of the corresponding products in each product category.

[0097] Among them, the long short-term memory (LSTM) model is a subtype of recurrent neural network. It combines feedback links to enable it to process time series data, such as behavioral data information. Figure 2 , Figure 2 A schematic diagram of the structure of the long short-term memory model provided in the embodiment of the present application. The long short-term memory model has multiple LSTM units, such as C1, C2, and C3. Taking the LSTM unit corresponding to C2 as an example, the LSTM unit includes a forget gate, an input gate, and an output gate. The forget gate is calculated by the formula:

[0098] f t =σ(W f ·[H t-1 ,X t ]+b f )

[0099] To determine which information is discarded from the cell state, where f t is the output of the forget gate, σ is the sigmoid activation function, W f is the weight matrix, b f is the bias term, Ht-1 is the hidden state of the previous time period, X t is the input for the current time period.

[0100] The input gate and candidate unit state are determined by the formula:

[0101] i t =σ(W i ·[H t-1 ,X t ]+b i )

[0102] C~ t =tanh(W c ·[H t-1 ,X t ]+b c )

[0103] decide which behavior data information will be added to the unit state, where i t is the output of the input gate, C~ t is the candidate cell state, W i and W c is the weight matrix, b i and b c is the bias term.

[0104] The results of the forget gate and the input gate are combined to update the cell state, which is achieved through the following formula:

[0105] C t =f t *C t-1 +i t *C~ t

[0106] Among them, C t is the updated cell state, and * indicates element-wise multiplication.

[0107] The output gate is based on

[0108] o t =σ(W o ·[H t-1 ,X t ]+b o )

[0109] H t =o t *tanh(C t )

[0110] Controls which behavior data information will be output from the unit state to the hidden state, where o t is the output of the output gate, H t is the hidden state in the current time period.

[0111] S103. Input the attribute information into the multi-head self-attention model to obtain feature representation.

[0112] In this embodiment, the attribute information is digitally encoded so that it can be processed by the model, and the encoded attribute information is input into the multi-head self-attention model, which processes the input attribute information. Inside the model, it calculates the degree of association and importance between each attribute information. Through the self-attention mechanism, the model can focus on the relationship between different attributes and pay more attention to important attributes. The multi-head self-attention model processes the input attribute information to obtain a feature representation. Among them, the feature representation is used to represent the probability that the user to be recommended selects various product categories based on the attribute information. The reason why this feature representation can be used to represent the probability that the user to be recommended selects various product categories based on the attribute information is that it captures the potential association between user attributes and product categories. It can be understood that it is a digital summary and refinement of the preferences and behavior patterns of the recommended users.

[0113] The Multi-Head Attention model is an extended form of the attention mechanism that is widely used in the Transformer model. It obtains the attention distribution of the input attribute information by running multiple independent attention mechanisms in parallel, thereby more comprehensively capturing the potential multiple semantic associations in the sequence.

[0114] S104: Input the interest degree and feature representation of the user to be recommended to each product category into a fully connected neural network to obtain a product category recommendation result corresponding to the user to be recommended.

[0115] In this embodiment, the interest and feature representation of the user to be recommended in various product categories are used as input data of the fully connected neural network. These input data will be propagated and calculated between the nodes of the fully connected neural network. The neurons in the fully connected neural network will process the input data. In this process, the fully connected neural network integrates and processes the input data through its internal trained parameters and connection structure. It will try to find the patterns and associations based on the combination of interest and feature representation. After calculation and processing, a result is finally output, that is, the product category recommendation result corresponding to the user to be recommended. It represents the product category that the fully connected neural network judges to be the most suitable for the user to be recommended based on the input data. Among them, the fully connected neural network adopts a simple single hidden layer neural network.

[0116] It should be noted that, since each long short-term memory model outputs the interest level of the corresponding product category, and the multi-head self-attention model outputs the corresponding feature representation, the feature representation is used for the probability of the recommended user selecting each product category based on the attribute information. Therefore, when the fully connected neural network processes the interest level and feature representation of each product category, it will generate the probability of recommending each product category, and then select the product category with the maximum probability from the probabilities of recommending each product category as the product category recommendation result corresponding to the recommended user.

[0117] Among them, multiple long short-term memory models, multi-head self-attention models and fully connected neural networks constitute the discriminator in the semi-supervised generative adversarial network model. When attribute information and behavioral data information within a preset time period are input into the semi-supervised generative adversarial network model, these different model components work together. The long short-term memory model obtains the interest of various product categories based on the behavioral data information. The self-attention model can focus on global relationships and patterns, and obtain feature representations based on attribute information. The fully connected neural network integrates the outputs of the above two models as inputs and makes the final discriminant decision to obtain the product category recommendation results corresponding to the recommended user. Through such a combination, the semi-supervised generative adversarial network model can analyze the input data more comprehensively and deeply, thereby improving the accuracy of product category recommendation results.

[0118] It should be noted that, in addition to being able to recommend product categories of banking products, the semi-supervised generative adversarial network model provided in this embodiment can also be used to recommend other information, such as product recommendations on e-commerce platforms, information or video recommendations on content platforms, etc.

[0119] An information recommendation method provided by an embodiment of the present application constructs a semi-supervised generative adversarial network model by utilizing a long short-term memory model, a multi-head self-attention model, and a fully connected neural network. The long short-term memory model processes sequence data and can capture the time dependency and long-term dependency in the user behavior data information of the recommended user within the pre-examination time period, and predict the user's interest in different product categories. The multi-head self-attention model can capture the complex relationships in the attribute information and generate feature representations for the product categories by focusing on different feature subsets in the attribute information. The fully connected network further achieves the technical effect of improving the accuracy of the product type recommendation results by capturing the relationship between the user's interest in different product categories and the feature representations generated by the multi-head self-attention model.

[0120] Figure 3 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 2 ,like Figure 3 As shown, in this embodiment Figure 1Based on the embodiment, a possible implementation of the information recommendation method is described in detail, and the method includes:

[0121] S301. Acquire product attribute data, process the product attribute data according to a product attribute clustering model to obtain multiple product categories, and determine the second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in any product category.

[0122] In this embodiment, the product attribute data include unit price, release time, product rating, holding period, etc., and the product attribute data are all numerical feature attributes. In order to avoid product classification errors caused by insufficient human experience, a Gaussian Mixture Model (GMM) is used to train a product attribute clustering model based on all product attribute data. The product attribute data is processed according to the product attribute clustering model to obtain multiple product categories, such as M product categories, and data such as cluster centers and products of each category are also obtained. Among them, product categories include, for example, loan products, deposit products, credit products, etc. By using the Gaussian mixture clustering model algorithm, we can better explore the similarity relationship between the attributes of products, divide them into multiple product categories, and divide products with similar attribute characteristics into one cluster. The classification results are as follows: Figure 4 As shown, different colors indicate differences between product categories.

[0123] Next, the second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in any product category are determined, wherein the interest of each historical user in any product category is equivalent to the label corresponding to each historical user.

[0124] Specifically, for example, M product categories are obtained, and the mth product category is selected. The second behavior data information of all historical users under the mth product category within a preset historical time period and the interest of each historical user in the mth product category are obtained from the database. The second behavior data information is preprocessed to ensure the quality and accuracy of the data. Among them, the interest of each historical user in the mth product category can be, for example, the total number of transactions of each product in the mth product category by the historical user within the preset historical time period.

[0125] S302: For any product category, determine a training set and a test set according to the second behavior data information and the interest level of any product category.

[0126] In this embodiment, for example, the mth product category is continued as an example: for the mth product category, the second behavior data information and the interest level of the mth product category are divided according to a preset ratio, for example, 80% of the total data is used for training, and 20% is used for prediction to be divided into a training set and a prediction set. It should be noted that the second behavior data information of a historical user corresponds to the interest level of the historical user in the mth product category.

[0127] S303, initializing the particle swarm to determine the initial position and speed of each particle and the number of iterations.

[0128] In this embodiment, the first long short-term memory model is optimized using a particle swarm algorithm, which is an optimization algorithm based on swarm intelligence and seeks the optimal solution by simulating the behavior of a flock of birds or a school of fish.

[0129] Before optimizing the first long short-term memory model, the particle swarm is initialized, and the initial position and velocity are randomly assigned to each particle. The initial position is usually randomly selected in the search space, while the velocity can be set to zero or initialized according to a preset strategy (such as Gaussian distribution). During initialization, the current position of each particle is set to its individual optimal position, that is, the individual extremum, and the optimal position is selected from all particles as the global optimal position, that is, the global extremum. At the same time, the maximum number of iterations or convergence conditions of the algorithm are set. This determines the length of time the algorithm runs and the search accuracy.

[0130] The initial position is used to represent the initial model parameters of the first long short-term memory model, and the model parameters include, for example, the learning rate, batch size, number of neurons, etc. of the first long short-term memory model.

[0131] In a possible implementation, the size of the particle swarm may be determined, for example, based on the problem scale or computing resources. Generally, the more particles there are, the stronger the algorithm's search capability is, but the computing cost also increases accordingly.

[0132] S304 . For each particle, the first long short-term memory model is trained according to the initial position and the training set to obtain a prediction error, and the fitness of each particle is calculated according to the prediction error.

[0133] In this embodiment, the initial position of each particle represents the initial model parameters of the first long short-term memory model. Therefore, the training set is input into the first long short-term memory model to obtain the prediction result corresponding to each historical user in the training set, and the prediction result and the real result are analyzed to obtain the prediction error corresponding to each historical user in the training set. The real result refers to the interest of each historical user in the training set in the mth product category.

[0134] The prediction error corresponding to each historical user is calculated to obtain the fitness of each particle. By calculating the fitness of each particle, the quality of the model parameters represented by each particle is measured, providing a basis for subsequent optimization and search processes, such as selecting particles with high fitness for further operations or updating the position of particles, etc., so as to gradually find a better model parameter configuration and improve the performance of the first long short-term memory model.

[0135] S305. According to the fitness of each particle, the individual extreme value and the global extreme value are determined, and the position and speed of each particle are updated according to the individual extreme value and the global extreme value. The iterative process is repeated until the number of iterations is met to obtain the target position of the particle corresponding to the global extreme value.

[0136] In this embodiment, after obtaining the fitness of each particle, for each particle itself, its current fitness is compared with the fitness it has experienced before (i.e., individual extreme value). If the current fitness is better, the individual extreme value is updated as the current fitness, and the corresponding position at this time is recorded. Then, among the individual extreme values ​​of all particles, the one with the best fitness is found as the global extreme value. Then, the position and speed of each particle are updated. When updating the speed, factors such as the current speed of the particle, the distance from the individual extreme value, and the distance from the global extreme value are usually considered. The updated position is calculated based on the new speed and current position. In this way, the update of position and speed is completed. Then this process is repeated continuously, and the optimal solution is continuously approached in the iterative process. When the set number of iterations is reached, the position of the particle corresponding to the global extreme value is the final target position, which represents the optimal model parameters. Through the particle swarm method, a more extensive and rapid search can be carried out in the parameter space, improving the efficiency of finding a better solution, thereby improving the performance of the first long short-term memory model.

[0137] S306. Obtain a second long short-term memory model according to the model parameters corresponding to the target position.

[0138] In this embodiment, after the model parameters corresponding to the target position are determined, these parameters are applied to the first long short-term memory model to form a second long short-term memory model with specific structure and performance.

[0139] S307. Test the second long short-term memory model according to the test set to obtain a long short-term memory model, and construct a first semi-supervised generative adversarial network model according to multiple long short-term memory models, the first multi-head self-attention model and the first fully connected neural network.

[0140] In this embodiment, the second long short-term memory model is tested using a test set to measure the performance of the model on new data, such as accuracy, recall rate, F1 value and other indicators, so as to determine whether its performance meets the requirements. If it meets the requirements, a long short-term memory model is obtained.

[0141] A first semi-supervised generative adversarial network model is constructed based on multiple long short-term memory models, a first multi-head self-attention model, and a first fully connected neural network. Since the long short-term memory model has been trained, it means that some model parameters in the first semi-supervised generative adversarial network model have been determined. By integrating multiple long short-term memory models and the first multi-head self-attention model using the first fully connected neural network, the first semi-supervised generative adversarial network model has stronger adaptability and generalization ability, and can be effectively learned and generated with a small amount of labeled data and a large amount of unlabeled data, thereby improving the performance and application scope of the model in complex tasks.

[0142] It should be noted that multiple product categories correspond to multiple long short-term memory models. For example, if there are M product categories, there are M long short-term memory models.

[0143] S308, obtaining a plurality of first training samples and a plurality of second training samples for training a first semi-supervised generative adversarial network model;

[0144] In this embodiment, the first training sample includes input information and labels corresponding to historical users. The labels are used to indicate the target product categories corresponding to historical users. The target product categories can be, for example, the real choices of historical users, which helps the model learn the association between different input information and specific product categories. The second training sample only includes input information of historical users, which plays an auxiliary role in semi-supervised learning, allowing the first semi-supervised generative adversarial network model to mine potential patterns and features from a large amount of unlabeled data. Input information includes attribute information and behavioral data information.

[0145] S309: Training the first semi-supervised generative adversarial network model according to the multiple first training samples, the multiple second training samples, and the random noise vector to obtain a trained semi-supervised generative adversarial network model.

[0146] In this embodiment, the first semi-supervised generative adversarial network model includes a first generator and a first discriminator. The first generator is used to generate pseudo input information based on a random noise vector, and the first discriminator is used to distinguish the authenticity of the input information and determine the corresponding prediction result based on the input information.

[0147] It should be noted that the labels of the first training samples help guide the first semi-supervised generative adversarial network model to learn the correspondence between different inputs and target categories, while the second training samples can help the first semi-supervised generative adversarial network model learn common features from a wider data distribution. The introduction of random noise vectors increases the flexibility and creativity of the model, enabling it to generate new data. By continuously adjusting the parameters of the first generator and the first discriminator, the first semi-supervised generative adversarial network model gradually achieves better performance in generating data and discriminating the authenticity of data, and finally obtains a trained semi-supervised generative adversarial network model that can effectively process and generate relevant data.

[0148] An information recommendation method provided in an embodiment of the present application can effectively classify products into different categories by processing product attribute data through a product attribute clustering model, which helps to analyze and understand the market positioning and user needs of products in more detail. The LSTM model is optimized using a particle swarm algorithm to improve the prediction accuracy of the model. By combining multiple long short-term memory models, multi-head self-attention models, and fully connected neural networks to construct a semi-supervised generative adversarial network model, the advantages of different models can be fully utilized to improve the performance of the overall model, thereby achieving the effect of improving the product category recommendation results.

[0149] Figure 5 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 3 ,like Figure 5 As shown, in this embodiment Figure 3 Based on the embodiment, a possible implementation method of training a first semi-supervised generative adversarial network model according to multiple first training samples, multiple second training samples and a random noise vector to obtain a trained semi-supervised generative adversarial network model is described in detail, and the method includes:

[0150] S501. Input a random noise vector into a first generator in a first semi-supervised generative adversarial network model to generate pseudo input information.

[0151] In this embodiment, after the random noise vector enters the first generator of the first semi-supervised generative adversarial network model as input, the first generator will try to convert the noise vector into pseudo input information that looks like real data based on its learned patterns and rules to deceive the first discriminator. The real data refers to the input information corresponding to the historical users in the first training sample and the input information corresponding to the historical users in the second training sample.

[0152] S502: Input the input information in the first training sample into the first discriminator in the first semi-supervised generative adversarial network model to obtain a prediction result, and determine the prediction loss according to the prediction result and the label in the first training sample.

[0153] In this embodiment, when the input information (real data) in the first training sample is input into the first discriminator of the first semi-supervised generative adversarial network model. The first discriminator processes it and outputs a prediction result. The prediction result is displayed, for example, in the form of a scalar value, indicating the predicted product category corresponding to the predicted historical user. The prediction loss is obtained by calculating the difference between the predicted product category and the label in the first training sample (the target product category corresponding to the historical user). This prediction loss measures the classification performance of the first discriminator on the real data.

[0154] S503: Input the input information in the second training sample into the first discriminator to obtain a first probability value, and obtain a first adversarial loss according to the first probability value, wherein the first probability value is used to indicate the probability that the input information is true.

[0155] In this embodiment, after the input information of the second training sample (unlabeled input information) is input into the first discriminator. The first discriminator will output a probability value indicating that the input information is true, that is, a first probability value. The first adversarial loss is calculated based on this first probability value. For example, if the first probability value is close to 1, it means that the first discriminator tends to believe that the input information is true; if it is close to 0, it means that the first discriminator tends to believe that the input information is false. Since the input information is real data, the first discriminator should output a high probability (close to 1), indicating that it believes the input is real. Therefore, the first adversarial loss is usually measured based on the ability of the first discriminator to output a high probability for real data.

[0156] S504: Input the pseudo input information into the first discriminator to obtain a second probability value, and obtain a second adversarial loss according to the second probability value, wherein the second probability value is used to indicate the probability that the pseudo input information is true.

[0157] In this embodiment, the first discriminator processes the pseudo input information and outputs a probability value, namely the second probability value. The second probability value is used to indicate the probability that the pseudo input information is true. Ideally, for the pseudo input information, the first discriminator should output a low probability (close to 0), indicating that it believes that the pseudo input information is false. The second adversarial loss is calculated using the second probability value. In this case, the second adversarial loss is measured based on the ability of the first discriminator to output a low probability for the pseudo input information generated by the first generator. This means that the first discriminator can identify which data is generated by the first generator.

[0158] S505. Update the first discriminator according to the prediction loss, the first adversarial loss, and the second adversarial loss to obtain a second discriminator.

[0159] In this embodiment, during the training process of the semi-supervised generative adversarial network model, the update of the first discriminator is based on a combination of multiple loss functions, including prediction loss, first adversarial loss, and second adversarial loss. These losses are weighted to obtain a total loss, and the gradient of the total loss with respect to the parameters of the first discriminator is calculated through the back propagation algorithm. The calculated gradient is then used by the optimization algorithm to adjust the parameters of the first discriminator to minimize the total loss, and finally the second discriminator is obtained.

[0160] It should be noted that, during weighted calculation, the weights corresponding to the prediction loss, the first adversarial loss, and the second adversarial loss depend on the training strategy and objectives. For example, the weights can be set manually. If the first semi-supervised generative adversarial network model is expected to focus on the prediction accuracy of labeled input information, the weight corresponding to the prediction loss is set to a larger value; if the adversarial nature of the first generator and the first discriminator is expected to be enhanced, the weight corresponding to the second adversarial loss is set to a larger value.

[0161] In one possible implementation, during the model training process, the learning rate of the second discriminator uses an adaptive learning rate. The adaptive learning rate method can dynamically adjust the learning rate according to feedback during the training process. For example, in the early stage of training, the learning rate can be large to accelerate convergence; and in the later stage of training, the learning rate can be reduced to improve the accuracy and stability of the model.

[0162] S506. Update the first generator according to the second adversarial loss to obtain a second generator.

[0163] In this embodiment, the second adversarial loss is calculated to generate a gradient with respect to the parameters of the first generator, and then the calculated gradient is used by the optimization algorithm to adjust the parameters of the first generator to minimize the second adversarial loss, and finally obtain the second discriminator.

[0164] S507. Iteratively train the second generator and the second discriminator to obtain a semi-supervised generative adversarial network model.

[0165] In this embodiment, by continuously alternating the training of the second generator and the second discriminator, in this iterative process, the two sides continuously promote and improve each other and gradually reach a state of equilibrium. For example, the maximum number of iterations can be preset in advance. When the maximum number of iterations is met, a relatively mature semi-supervised generative adversarial network model is obtained, which can better complete the generation task and the discrimination task under semi-supervision, and can be used in the scenario of product category recommendation.

[0166] It should be noted that the second discriminator includes multiple long short-term memory models, multi-head self-attention models and fully connected neural networks. Since the multiple long short-term memory models have been trained before being integrated into the second discriminator, it means that the model parameters of the multiple long short-term memory models have been set. Then the process of training the second discriminator in the above steps S501-S507 is to determine the model parameters of the multi-head self-attention model and the fully connected neural network. When the training of the second discriminator is completed, the model parameters of the multi-head self-attention model and the fully connected neural network have also been determined.

[0167] The information recommendation method provided in the embodiment of the present application generates pseudo input information by inputting a random noise vector into the first generator, as well as inputting and discriminating the training samples, calculating the prediction loss, adversarial loss, etc., and updating and iteratively training the first discriminator and the first generator accordingly. Doing so enables the model to better learn the data distribution and characteristics, and improves the generation and discrimination capabilities of the semi-supervised generative adversarial network model. At the same time, by using the semi-supervised learning method, full use is made of labeled and unlabeled data, and the generalization ability and adaptability of the semi-supervised generative adversarial network model are improved, thereby improving the accuracy of the product category recommendation results.

[0168] In one possible implementation, Figure 6 A schematic diagram of the structure of the semi-supervised generative adversarial network model provided in the embodiment of the present application is shown in FIG. Figure 6 As shown in FIG. 1 , the semi-supervised generative adversarial network model includes a generator and a discriminator. The discriminator includes multiple long short-term memory models and a multi-head self-attention model. Multiple long short-term memory models and multi-head self-attention models are connected through a fully connected neural network. Multiple long short-term memory models constitute an LSTM interest prediction module, and the multi-head self-attention model includes an embedding layer and a Transformer structure. The generator is used to generate pseudo input information (x * ), as pseudo samples. The dataset includes multiple first training samples (x, y) and multiple second training samples (x). The first training sample includes the input information (x) and label (y) corresponding to the historical user, and the label is used to indicate the target product category corresponding to the historical user. The second training sample only includes the input information (x) of the historical user. The input information includes attribute information and behavior data information. The dataset and pseudo input information are input into the discriminator, which is used to distinguish the authenticity of the input information and determine the corresponding prediction results based on the input information. By continuously optimizing the parameters of the generator and the discriminator, the final semi-supervised generative adversarial network model is obtained.

[0169] In one possible implementation, Figure 7 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 4 ,like Figure 7 As shown, a possible implementation of the information recommendation method is described, and the method includes:

[0170] The first data of historical users is obtained, and the first data includes attribute information, behavior data information and product attribute data, and the first data is processed and standardized. According to the product attribute data, a Gaussian mixture clustering model (GMM) is used for training to form a product attribute clustering model, and a classification result of the product category is obtained.

[0171] Based on the classification results, taking the mth product category as an example, obtain the labels corresponding to all users in the mth product category and the behavior information data within a preset historical time period. The long short-term memory model is trained according to the labels and behavior information data, and the particle swarm algorithm is used to optimize the long short-term memory model to determine whether the long short-term memory model is optimal. If so, a trained long short-term memory model is obtained.

[0172] Based on the classification result, the attribute information data of the first historical user is obtained, wherein the first historical user is partially labeled and partially unlabeled, and the label is used to indicate the digital code of the product category that the first historical user should recommend,

[0173] The multi-head self-attention model is trained using the attribute information data, where the multi-head self-attention model is used as part of the discriminator in the semi-supervised generative adversarial network model. Therefore, the generator in the semi-supervised generative adversarial network model can be used to train the multi-head self-attention model. During the training process, it is determined whether the data set input into the multi-head self-attention model reaches a preset scale, where the data set refers to the input attribute information data and the pseudo attribute information data generated by the generator. If so, it is determined whether the model training cycle reaches a preset value (epoch value). If so, a trained multi-head self-attention model is obtained.

[0174] The multi-head self-attention model and multiple long short-term memory models are connected through a fully connected neural network to obtain the discriminator in the semi-supervised generative adversarial network model. The semi-supervised generative adversarial network model is used to predict the attribute information of the recommended users and the behavioral data information within a preset time period to obtain the product category recommendation results.

[0175] In one possible implementation, Figure 8 A schematic diagram of a process of an information recommendation method provided in an embodiment of the present application Figure 5 ,like Figure 8 As shown, a possible implementation method of the optimization process of the long short-term memory model using the particle swarm algorithm is described. The method includes:

[0176] For any product category, taking the mth product category as an example, obtain the preprocessed historical user data in the mth product category. The historical user data includes the label corresponding to the historical user and the behavior data in the preset historical time period. The label is used to indicate the historical user's interest in the mth product category.

[0177] According to the pre-processed historical user data, the particle swarm algorithm is used to train the long short-term memory model. First, the particle swarm is initialized to determine the initial position and speed of each particle and the number of iterations. The initial position is used to represent the initial model parameters of the long short-term memory model, where the model parameters include, for example, the learning rate, batch-size, the number of neurons, etc.

[0178] For each particle, the long short-term memory model is trained according to the preprocessed historical user data and the corresponding initial model parameters, and the prediction result of the interest of the mth product category is output. According to the prediction error between the predicted result and the actual result, the fitness of each particle is calculated to determine the individual extreme value and global extreme value in the particle group.

[0179] According to the individual extreme value and the global extreme value, the speed and position of the particle are updated, and the fitness of the new particle is repeatedly calculated. The above iterative process is traversed to determine whether the preset conditions are met. If so, a long short-term memory model (PSO-LSTM) based on particle swarm parameter optimization is obtained, where the preset conditions are met, including but not limited to the maximum number of iterations for the long short-term memory model training and the loss function converges to a certain value.

[0180] Since there are multiple product categories, it is necessary to train the long short-term memory model corresponding to each product category, so as to obtain multiple long short-term memory models.

[0181] Fig. 9 A schematic diagram of the structure of an information recommendation device provided by this application, such as Fig. 9 As shown, the information recommendation device 900 provided in this embodiment includes:

[0182] The first processing module 901 is used to obtain attribute information of the user to be recommended and behavior data information within a preset time period;

[0183] The second processing module 902 is used to input the behavior data information into a plurality of long-term and short-term memory models respectively to obtain the interest of the user to be recommended in various product categories;

[0184] The third processing module 903 is used to input the attribute information into the multi-head self-attention model to obtain a feature representation, which is used to represent the probability of the recommended user selecting each product category based on the attribute information;

[0185] The fourth processing module 904 is used to input the interest and feature representation of the user to be recommended in various product categories into the fully connected neural network to obtain the product category recommendation results corresponding to the user to be recommended, wherein multiple long short-term memory models, multi-head self-attention models and fully connected neural networks constitute the discriminator in the semi-supervised generative adversarial network model.

[0186] In a possible implementation manner, the first processing module 901 is further configured to:

[0187] Build the first semi-supervised generative adversarial network model;

[0188] Acquire multiple first training samples and multiple second training samples for training the first semi-supervised generative adversarial network model; the first training samples include: input information and labels corresponding to historical users, the labels are used to indicate the target product category corresponding to the historical users; the second training samples include: input information of historical users; the input information includes attribute information and behavior data information;

[0189] The first semi-supervised generative adversarial network model is trained according to multiple first training samples, multiple second training samples and a random noise vector to obtain a trained semi-supervised generative adversarial network model, wherein the first semi-supervised generative adversarial network model includes a first generator and a first discriminator, the first generator is used to generate pseudo input information according to the random noise vector, the first discriminator is used to distinguish the authenticity of the input information, and determine the corresponding prediction result according to the input information.

[0190] In a possible implementation manner, the first processing module 901 is further configured to:

[0191] Acquire product attribute data, process the product attribute data according to a product attribute clustering model to obtain multiple product categories, and determine second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in any product category;

[0192] For any product category, the first long short-term memory model is trained according to the second behavior data information of each historical user and the interest level of any product category to obtain a long short-term memory model corresponding to any product category after training, wherein multiple product categories correspond to multiple long short-term memory models;

[0193] The first semi-supervised generative adversarial network model is constructed based on multiple long short-term memory models, the first multi-head self-attention model and the first fully connected neural network.

[0194] In a possible implementation manner, the first processing module 901 is further configured to:

[0195] Determine a training set and a test set based on the second behavior data information and the interest level of any product category;

[0196] According to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model;

[0197] The second long short-term memory model is tested according to the test set to obtain a long short-term memory model.

[0198] In a possible implementation manner, the first processing module 901 is further configured to:

[0199] Initialize the particle swarm to determine the initial position, speed, and number of iterations of each particle. The initial position is used to represent the initial model parameters of the first long short-term memory model.

[0200] For each particle, the first long short-term memory model is trained according to the initial position and the training set to obtain the prediction error, and the fitness of each particle is calculated according to the prediction error;

[0201] According to the fitness of each particle, the individual extreme value and the global extreme value are determined, and the position and speed of each particle are updated according to the individual extreme value and the global extreme value. The iterative process is repeated until the number of iterations is met and the target position of the particle corresponding to the global extreme value is obtained;

[0202] According to the model parameters corresponding to the target position, a second long short-term memory model is obtained.

[0203] In a possible implementation manner, the first processing module 901 is further configured to:

[0204] Inputting the random noise vector into the first generator in the first semi-supervised generative adversarial network model to generate pseudo input information;

[0205] Inputting the input information in the first training sample into the first discriminator in the first semi-supervised generative adversarial network model to obtain a prediction result, and determining the prediction loss according to the prediction result and the label in the first training sample;

[0206] Inputting the input information in the second training sample into the first discriminator to obtain a first probability value, and obtaining a first adversarial loss according to the first probability value, wherein the first probability value is used to indicate the probability that the input information is true;

[0207] Inputting the pseudo input information into the first discriminator to obtain a second probability value, and obtaining a second adversarial loss according to the second probability value, wherein the second probability value is used to indicate the probability that the pseudo input information is true;

[0208] The first discriminator is updated according to the prediction loss, the first adversarial loss and the second adversarial loss to obtain the second discriminator;

[0209] According to the second adversarial loss, the first generator is updated to obtain the second generator;

[0210] The second generator and the second discriminator are iteratively trained to obtain a semi-supervised generative adversarial network model.

[0211] The information recommendation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and this embodiment will not be described in detail here.

[0212] Fig.10 This is a schematic diagram of the structure of an information recommendation device provided by this application. Fig.10 As shown, the information recommendation device 1000 provided in this embodiment includes: at least one processor 1001 and a memory 1002. Optionally, the device 1000 also includes a communication component 1003. The processor 1001, the memory 1002 and the communication component 1003 are connected via a bus 1004.

[0213] In a specific implementation process, at least one processor 1001 executes the computer-executable instructions stored in the memory 1002, so that at least one processor 1001 executes the above method.

[0214] The specific implementation process of the processor 1001 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0215] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0216] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-volatile Memory, NVM), such as at least one disk memory.

[0217] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0218] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0219] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0220] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0221] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0222] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0223] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0224] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0225] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0226] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0227] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0228] In addition, this application involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.), and the use of artificial intelligence technology for automated decision-making, and a technical solution for making decisions that have a significant impact on personal rights and interests based on the results of automated decision-making. The application provides users with corresponding operation entrances for them to choose to agree or reject the results of automated decision-making; if the user chooses to reject, the expert decision-making process will be entered.

[0229] It should be noted that the terms "first", "second", etc. in the claims, the specification and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the invention described herein, for example, can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, products or devices.

[0230] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary technical means in the art not disclosed by the present invention, are not limited to the precise structure described above and shown in the drawings, and may be modified and changed in various ways without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. An information recommendation method, characterized in that: include: Obtaining attribute information of the user to be recommended and behavioral data information within a preset time period; Inputting the behavior data information into a plurality of long-short term memory models respectively to obtain the interest of the user to be recommended in various product categories; Inputting the attribute information into a multi-head self-attention model to obtain a feature representation, wherein the feature representation is used to represent the probability that the user to be recommended selects the various product categories based on the attribute information; The interest of the user to be recommended in various product categories and the feature representation are input into a fully connected neural network to obtain the product category recommendation results corresponding to the user to be recommended, wherein the multiple long short-term memory models, the multi-head self-attention model and the fully connected neural network constitute the discriminator in the semi-supervised generative adversarial network model.

2. The method according to claim 1, characterized in that Before obtaining the attribute information of the user to be recommended and the behavior data information within a preset time period, the method further includes: Build the first semi-supervised generative adversarial network model; Acquire a plurality of first training samples and a plurality of second training samples for training the first semi-supervised generative adversarial network model; the first training samples include: input information and labels corresponding to historical users, the labels are used to indicate the target product categories corresponding to the historical users; the second training samples include: input information of historical users; the input information includes attribute information and behavior data information; The first semi-supervised generative adversarial network model is trained according to the multiple first training samples, the multiple second training samples and the random noise vector to obtain the trained semi-supervised generative adversarial network model, wherein the first semi-supervised generative adversarial network model includes a first generator and a first discriminator, the first generator is used to generate pseudo input information according to the random noise vector, the first discriminator is used to distinguish the authenticity of the input information, and determine the corresponding prediction result according to the input information.

3. The method according to claim 2, characterized in that Construct the first semi-supervised generative adversarial network model, including: Acquire product attribute data, process the product attribute data according to a product attribute clustering model to obtain multiple product categories, and determine second behavior data information of all historical users corresponding to any product category in the multiple product categories within a preset historical time period and the interest of each historical user in the any product category; For any of the product categories, training the first long short-term memory model according to the second behavior data information of each historical user and the interest level of any of the product categories to obtain a trained long short-term memory model corresponding to any of the product categories, wherein multiple product categories correspond to multiple long short-term memory models; The first semi-supervised generative adversarial network model is constructed based on multiple long short-term memory models, a first multi-head self-attention model and a first fully connected neural network.

4. The method according to claim 3, characterized in that The first long short-term memory model is trained according to the second behavior data information of each historical user and the interest level of any product category to obtain a trained long short-term memory model corresponding to any product category, including: Determining a training set and a test set according to the second behavior data information and the interest level of any one of the product categories; According to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model; The second long short-term memory model is tested according to the test set to obtain the long short-term memory model.

5. The method according to claim 4, characterized in that According to the training set, the model parameters of the first long short-term memory model are optimized by using a particle swarm algorithm to obtain a second long short-term memory model, including: Initializing the particle swarm to determine the initial position, speed, and number of iterations of each particle, wherein the initial position is used to represent initial model parameters of the first long short-term memory model; For each particle, training the first long short-term memory model according to the initial position and the training set to obtain a prediction error, and calculating the fitness of each particle according to the prediction error; According to the fitness of each particle, an individual extreme value and a global extreme value are determined, and according to the individual extreme value and the global extreme value, the position and speed of each particle are updated, and the iterative process is traversed until the number of iterations is met to obtain the target position of the particle corresponding to the global extreme value; A second long short-term memory model is obtained according to the model parameters corresponding to the target position.

6. The method according to claim 2, characterized in that The first semi-supervised generative adversarial network model is trained according to the multiple first training samples, the multiple second training samples, and the random noise vector to obtain the trained semi-supervised generative adversarial network model, including: Inputting the random noise vector into the first generator in the first semi-supervised generative adversarial network model to generate the pseudo input information; Inputting the input information in the first training sample into the first discriminator in the first semi-supervised generative adversarial network model to obtain a prediction result, and determining a prediction loss according to the prediction result and the label in the first training sample; Inputting the input information in the second training sample into the first discriminator to obtain a first probability value, and obtaining a first adversarial loss according to the first probability value, wherein the first probability value is used to indicate the probability that the input information is true; Inputting the pseudo input information into the first discriminator to obtain a second probability value, and obtaining a second adversarial loss according to the second probability value, wherein the second probability value is used to indicate the probability that the pseudo input information is true; Update the first discriminator according to the prediction loss, the first adversarial loss, and the second adversarial loss to obtain a second discriminator; According to the second adversarial loss, the first generator is updated to obtain a second generator; The second generator and the second discriminator are iteratively trained to obtain the semi-supervised generative adversarial network model.

7. An information recommendation device, characterized in that: include: The first processing module is used to obtain attribute information of the user to be recommended and behavior data information within a preset time period; The second processing module is used to input the behavior data information into a plurality of long-term and short-term memory models respectively to obtain the interest of the user to be recommended in various product categories; A third processing module is used to input the attribute information into a multi-head self-attention model to obtain a feature representation, where the feature representation is used to represent the probability that the user to be recommended selects the various product categories based on the attribute information; The fourth processing module is used to input the interest of the user to be recommended in various product categories and the feature representation into the fully connected neural network to obtain the product category recommendation results corresponding to the user to be recommended, wherein the multiple long short-term memory models, the multi-head self-attention model and the fully connected neural network constitute the discriminator in the semi-supervised generative adversarial network model.

8. An information recommendation device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.

10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 6 when being executed by a processor.

Images