Commodity pushing method and commodity pushing device based on user positioning

Abstract

The invention provides a commodity pushing method and a commodity pushing device based on user positioning. The method comprises the following steps: mapping historical attribute information and historical behavior information into a first target vector and a second target vector; fusing the first target vector and the second target vector in a spatial domain to obtain a third target vector; fusing the third target vector in a time domain to obtain a fourth target vector, determining the fourth target vector as a node attribute, initializing an edge between nodes according to historical interaction information between users to obtain a graph structure, performing feature extraction on the graph structure through a graph convolutional neural network to obtain a first target feature, and obtaining a second target feature; clustering based on the first target feature to obtain user tendency location; and processing the user tendency positioning through the personalized recommendation model to obtain target commodity information, and pushing the target commodity information to the user. According to the method, the problem that the user satisfaction is reduced due to low effectiveness of the commodity pushing content in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a commodity push method based on user positioning, a commodity push device, a computer-readable storage medium, and a commodity push system. Background Art

[0002] With the accelerated development of the digital age, financial institutions have invested a lot of resources in developing personalized push strategies based on big data analysis in order to improve push efficiency and customer satisfaction. Currently, push activities are not only diverse, but also increasingly diversified in terms of contact methods, including SMS, phone calls, APPs, social media and other channels. However, this richness and diversity also brings new customer experience pain points. On the one hand, a large amount of push information is flooding users, especially in SMS and phone channels. Coupled with the frequent appearance of fraudulent SMS in society, users' trust and attention to these traditional push methods have dropped sharply. On the other hand, although the push content released by online channels such as APPs is comprehensive, there are problems such as information overload. It is difficult for users to filter push activities that meet their personal needs, resulting in low efficiency.

[0003] In the field of personalized push notifications, forming user feature representations and performing user collaborative filtering positioning are key steps to achieve accurate push notifications. Although the methods in the prior art can capture the basic characteristics of users, they lack comprehensive consideration of information in the time and space domains when integrating user historical information and multimodal features, and the constructed user portraits lack breadth and depth. For user collaborative filtering positioning, the prior art focuses on clustering using basic user information or introducing topological network graphs of user interactions for clustering, but they have failed to deeply explore the complex relationship between user attribute information and interaction information, resulting in low clustering accuracy and efficiency.

[0004] In summary, the description of user characteristics in the prior art is not comprehensive enough, and the classification process based on user characteristics is not accurate enough, resulting in inaccurate final user positioning, and the pushed content does not meet the user's personal needs, resulting in low user satisfaction and affecting the push effect. Summary of the invention

[0005] The main purpose of the present application is to provide a product push method based on user positioning, a product push device, a computer-readable storage medium and a product push system, so as to at least solve the problem in the prior art that the effectiveness of product push content is low, resulting in reduced user satisfaction.

[0006] To achieve the above object, according to one aspect of the present application, there is provided a commodity push method based on user positioning, including: extracting multiple historical attribute information and multiple historical behavior information of a user according to the user's historical consumption data at a preset period, mapping the historical attribute information into a vector to obtain a first target vector, mapping the historical behavior information into a vector to obtain a second target vector, the first target vector and the second target vector having the same dimension, the historical attribute information including user age, gender, consumption preference and credit rating, and the historical behavior information including transactions, financial management behaviors and credit records; performing weighted fusion on the first target vector and the second target vector in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector; performing weighted fusion on the third target vector in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector; determining the fourth target vector as a node attribute, and initializing the edges between nodes according to the historical interaction information between users to obtain a graph structure, performing feature extraction on the graph structure through a graph convolutional neural network to obtain a first target feature, performing clustering based on the first target feature to obtain a user preference positioning; processing the user preference positioning through a personalized recommendation model to obtain target commodity information, and pushing the target commodity information to the user, the personalized recommendation model being trained according to the feedback data of multiple users on different commodities and the different user preference positionings corresponding to the multiple users.

[0007] Optionally, performing weighted fusion on the first target vector and the second target vector in the spatial domain through a Transformer structure to obtain a third target vector includes: initializing a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head; performing weighted fusion on the empty global head, the first target vector and the second target vector according to preset weights to obtain a fifth target vector; performing feature extraction on the fifth target vector through the multi-head self-attention mechanism of the Transformer structure to obtain multiple second target features, the second target features corresponding one-to-one to the attention heads of the Transformer structure; and fusing the second target features through the fully connected layer of the Transformer structure to obtain a third target vector.

[0008] Optionally, the third target vector is weighted and fused in the time domain by a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector, including: arranging the third target vector according to the timestamp in chronological order, and adding a positional encoding to the arranged third target vector to obtain a sixth target vector, where the positional encoding is used to uniquely identify the position of the third target vector in the sixth target vector; initializing a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head; inputting the empty global head and the sixth target vector into the Transformer structure, and performing feature extraction through the multi-head self-attention mechanism of the Transformer structure to obtain multiple third target features; fusing the third target features through the fully connected layer of the Transformer structure to obtain the fourth target vector.

[0009] Optionally, the fourth target vector is determined as the node attribute, historical interaction information between users is obtained, and the edges between nodes are initialized according to the historical interaction information to obtain a graph structure, including: determining each user as a node in the graph structure, and configuring the fourth target vector as the node attribute of the node; extracting the interactions between each user according to the historical interaction information, and connecting the nodes corresponding to the users with interactions through edges; extracting the intensity and frequency of the interactions between each user according to the historical interaction information, and initializing the weights corresponding to each edge according to the intensity and frequency, and mapping them to a preset interval to obtain target weights; configuring each target weight on the corresponding edge to obtain a graph structure.

[0010] Optionally, feature extraction is performed on the graph structure through a graph convolutional neural network to obtain a first target feature, including: converting the edges and target weights of the graph structure into an adjacency matrix for representation to obtain a first target matrix; converting the nodes and node attributes of the graph structure into a diagonal matrix for representation to obtain a second target matrix; performing topological structure extraction on the first target matrix through a first graph convolutional neural network to obtain a fourth target feature; performing attribute information extraction on the first target matrix and the second target matrix through a second graph convolutional neural network to obtain a fifth target feature; performing weighted fusion on the fourth target feature and the fifth target feature to obtain the first target feature.

[0011] Optionally, before processing the user preference orientation through the personalized recommendation model to obtain the target commodity information, the method further includes: obtaining corresponding historical feedback data according to different user preference orientations, where the historical feedback data includes the push feedback of the user on different types of commodities; constructing a training data set and a test data set according to the historical feedback data and the user preference orientation; using the user preference orientation as input data and the commodity with the highest purchase rate as output data, training an alternative personalized recommendation model according to the training data set until the energy function output by the output layer of the alternative personalized recommendation model is less than a preset error; inputting the test data set into the alternative personalized recommendation model, obtaining an output result through forward propagation, and performing backpropagation on the alternative personalized recommendation model according to the loss value between the actual data and the output result until the loss value of the alternative personalized recommendation model is less than or equal to the preset loss value, and determining the alternative personalized recommendation model as the personalized recommendation model.

[0012] Optionally, clustering based on the first target feature to obtain the user preference orientation includes: clustering the first target feature according to the softmax function to obtain the user preference orientation.

[0013] According to another aspect of the present application, there is provided a commodity push device based on user positioning. The device includes: an acquisition unit, configured to extract multiple historical attribute information and multiple historical behavior information of a user according to the user's historical consumption data at a preset period, map the historical attribute information into a vector to obtain a first target vector, map the historical behavior information into a vector to obtain a second target vector, where the dimensions of the first target vector and the second target vector are the same, the historical attribute information includes user age, gender, consumption preference, and credit rating, and the historical behavior information includes transactions, financial management behaviors, and credit records; a first calculation unit, configured to perform weighted fusion of the first target vector and the second target vector in the spatial domain through a Transformer structure according to a preset weight to obtain a third target vector; a second calculation unit, configured to perform weighted fusion of the third target vector in the time domain through a Transformer structure according to the time stamp of the third target vector to obtain a fourth target vector; a construction unit, configured to determine the fourth target vector as the node attribute, initialize the edges between the nodes according to the historical interaction information between the users to obtain a graph structure, extract features of the graph structure through a graph convolutional neural network to obtain a first target feature, and perform clustering based on the first target feature to obtain the user preference orientation; a push unit, configured to process the user preference orientation through a personalized recommendation model to obtain target commodity information, and push the target commodity information to the user, where the personalized recommendation model is trained according to the feedback data of multiple users on different commodities and the different user preference orientations corresponding to the multiple users.

[0014] According to another aspect of the present application, there is provided a computer-readable storage medium, which includes a stored program. When the program runs, it controls the device where the computer-readable storage medium is located to execute any one of the methods.

[0015] According to yet another aspect of the present application, there is provided a commodity push system, including: one or more processors, a memory, and one or more programs. Among them, the one or more programs are stored in the memory and are configured to be executed by the one or more processors. The one or more programs include methods for executing any one of them.

[0016] Applying the technical solution of the present application in the above-mentioned commodity push method based on user location, first, according to the user's historical consumption data, a plurality of historical attribute information and a plurality of historical behavior information of the user are extracted according to a preset period. The historical attribute information is mapped into a vector to obtain a first target vector, and the historical behavior information is mapped into a vector to obtain a second target vector. The dimensions of the first target vector and the second target vector are the same. The historical attribute information includes user age, gender, consumption preference, and credit rating, and the historical behavior information includes transactions, financial management behaviors, and credit records. Then, according to the preset weights, the first target vector and the second target vector are weighted and fused in the spatial domain through a Transformer structure to obtain a third target vector. After that, according to the time stamp of the third target vector, the third target vector is weighted and fused in the time domain through a Transformer structure to obtain a fourth target vector. After that, the fourth target vector is determined as the node attribute, and the edges between the nodes are initialized according to the historical interaction information between the users to obtain a graph structure. The graph structure is subjected to feature extraction through a graph convolutional neural network to obtain a first target feature, and clustering is performed based on the first target feature to obtain a user preference location. Finally, the user preference location is processed through a personalized recommendation model to obtain target commodity information, and the target commodity information is pushed to the user. The personalized recommendation model is trained according to the feedback data of multiple users on different commodities and the different user preference locations corresponding to the multiple users. The present application obtains information from multiple dimensions based on a large amount of customer data, can comprehensively and completely reflect user characteristics, and converts the collaborative filtering location problem into a clustering problem of graph convolution during the user feature extraction process, can more comprehensively combine user characteristics to improve the location accuracy, and realizes the recommendation of target commodity information through a personalized recommendation model, improving user satisfaction. This method solves the problem that the effectiveness of the commodity push content in the prior art is relatively low, resulting in a decrease in user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The hardware structure block diagram of a mobile terminal showing a commodity push method based on user location provided in an embodiment of the present application is shown;

[0018] Figure 2 shows a schematic flowchart of a method for pushing commodities based on user location provided according to an embodiment of the present application;

[0019] Figure 3 shows a structural block diagram of a device for pushing commodities based on user location provided according to an embodiment of the present application.

[0020] Among them, the above-mentioned drawings include the following reference numerals:

[0021] 102, a processor; 104, a memory; 106, a transmission device; 108, an input / output device. Detailed implementation manners

[0022] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0023] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data may be interchanged under appropriate circumstances so as to describe the embodiments of the present application herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, commodity or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, commodities or devices.

[0025] For the convenience of description, some nouns or terms related to the embodiments of the present application are described below:

[0026] Graph: A topological structure consisting of a finite non-empty set of vertices V(G) and a set of edges between vertices E(G), usually expressed as: G = (V, E), where G represents the entire graph, V is the set of vertices in G, and E is the set of edges in G. The characteristic representation of a node, i.e., the characteristic vector, is called the attribute of the node. In particular, when the edge does not have directional information, it is called an undirected graph.

[0027] Graph Convolutional Networks (GCN): A convolutional neural network that can directly act on a graph and utilize its structural information. Graph convolutional networks can utilize the overall topological information by decomposing the nodes in the graph structure using Laplace features, and then mapping the obtained feature vectors as basis vectors to the spectral domain for convolution-like fusion.

[0028] Multi-head self-attention mechanism: An attention mechanism suitable for serialized data. It is the core of the Transformer structure and can parallelly calculate the similarity between all inputs (including itself) to mine global information in the sequence input. In addition to multi-head self-attention, the Transformer network also contains position encoding, residual and feedforward fully connected layers. The principle is to position encode each input and then input it into the multi-head self-attention module for fusion, and then use the feedforward fully connected layer to adjust the dimension, and finally combine the residual connection to prevent network degradation.

[0029] As introduced in the background technology, the description of user characteristics in the prior art is not comprehensive enough, and the classification process based on user characteristics is not accurate enough, resulting in inaccurate final user positioning, and the pushed content does not meet the user's personal needs, resulting in low user satisfaction and affecting the push effect. In order to solve the problem of low effectiveness of product push content in the prior art, which leads to reduced user satisfaction, the embodiments of the present application provide a product push method based on user positioning, a product push device, a computer-readable storage medium and a product push system.

[0030] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0031] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 1 is a hardware structure block diagram of a mobile terminal of a commodity push method based on user location according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1Only one processor 102 is shown (the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Among them, the above mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 The structure shown is only illustrative and does not limit the structure of the above mobile terminal. For example, the mobile terminal may further include more or fewer components than Figure 1 shown in, or have a different configuration from Figure 1 that shown.

[0032] The memory 104 can be used to store computer programs. For example, software programs and modules of application software, such as the computer program corresponding to the display method of device information in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implements the above method. The memory 104 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the above network may include a wireless network provided by a communication provider of the mobile terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus can communicate with the Internet. In one instance, the transmission device 106 may be a radio frequency (Radio Frequency, abbreviated as RF) module, which is used to communicate with the Internet wirelessly.

[0033] In this embodiment, a method for pushing products based on user location running on a mobile terminal, a computer terminal, or a similar computing device is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0034] Figure 2 is a flowchart of a method for pushing products based on user location according to an embodiment of the present application. As Figure 2 shown, the method includes the following steps:

[0035] Step S201, extract multiple historical attribute information and multiple historical behavior information of the user according to the preset period from the user's historical consumption data, map the historical attribute information into a vector to obtain a first target vector, map the historical behavior information into a vector to obtain a second target vector. The dimensions of the first target vector and the second target vector are the same. The historical attribute information includes user age, gender, consumption preference, and credit rating. The historical behavior information includes transactions, financial management behaviors, and credit records;

[0036] Specifically, taking user A as an example, assuming that the above preset period is 30 days, the system automatically extracts the consumption data of user A within multiple 30-day periods in the past year from the historical consumption data of user A. For each period, information including but not limited to user age, gender, consumption preference, and credit rating can be extracted to obtain the above historical attribute information. Similarly, behavior information such as the user's transactions, financial management behaviors, and credit records is extracted to obtain the above historical behavior information, and the above historical attribute information and the above historical behavior information are mapped into vectors of dimension X to obtain the above first target vector and the above second target vector.

[0037] Step S202, perform weighted fusion on the first target vector and the second target vector in the spatial domain through a Transformer structure according to the preset weight to obtain a third target vector;

[0038] Specifically, according to the prior weight (preset weight) provided by the expert, perform weighted fusion on the first target vector and the second target vector in the same time period. For example, assuming that the consumption preference weight is 0.6 and the age weight is 0.2, process these vectors through a Transformer structure to obtain a third target vector that combines attribute and behavior information, and the dimension of the vector remains unchanged.

[0039] It can be understood that the above operations use a Transformer structure, combined with the expert prior weight, to perform deep learning fusion on the user's historical attribute information (such as age, gender, consumption preference, credit rating) and historical behavior information (such as transactions, financial management behaviors, credit records) in the same time stage to maintain the integrity and relevance of multi-modal information.

[0040] Step S203, perform weighted fusion on the third target vector in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector;

[0041] Specifically, the third target vectors for each time stage are sorted in chronological order to form a sequence, and then processed again through the Transformer structure to consider the time information. By adding position encoding, the model can understand the order relationship of the vectors in the sequence, thereby obtaining a fourth target vector that combines the long-term and short-term behavior patterns of the user, while keeping the dimension of the vector unchanged.

[0042] It can be understood that the above operations re-use the Transformer structure for the feature vectors of different time stages, maintain the time information through position encoding, and achieve the comprehensive integration of long-term historical information and recent dynamic information.

[0043] Step S204: Determine the fourth target vector as the node attribute, initialize the edges between nodes according to the historical interaction information between users to obtain a graph structure, extract features from the graph structure through a graph convolutional neural network to obtain the first target feature, and perform clustering based on the first target feature to obtain the user preference positioning.

[0044] Specifically, each user is determined as a node, the above-mentioned fourth target vector corresponding to the user is determined as the node attribute of the node, and then the edge connection between nodes is initialized according to the interaction between users. Furthermore, the edge weights are initialized according to the frequency and intensity of the interaction to obtain the above-mentioned graph structure. Then, a graph convolutional neural network (GCN) is used to extract features from the graph to obtain the first target feature of the user nodes. Finally, clustering analysis is performed based on these features to obtain the above-mentioned user preference positioning of User A.

[0045] It can be understood that the above operations construct an interaction graph between users, extract features from the graph structure using a graph convolutional neural network (GCN), and perform clustering analysis between users from multiple dimensions of time features and spatial features by adaptively fusing node attributes and topological structure information, so as to achieve accurate positioning of user preferences.

[0046] Step S205: Process the user preference positioning through a personalized recommendation model to obtain target commodity information, and push the target commodity information to the user. The personalized recommendation model is trained based on the feedback data of multiple users on different commodities and the different user preference positionings corresponding to multiple users.

[0047] Specifically, according to the preference positioning of User A, the data is input into a pre-trained personalized recommendation model. The model will output a commodity recommendation list that best meets the needs of User A based on the characteristics of User A and the preset commodity database. And the personalized recommendation model will continuously learn online, adjust the model parameters according to the feedback of User A on the recommended content (such as clicks, purchases, etc.) to further improve the accuracy and relevance of the recommendation.

[0048] It can be understood that in the above operation, according to the user's preference positioning and combined with the online learning mechanism, the personalized recommendation model is fine-tuned to generate dynamic and personalized commodity recommendations for specific user types.

[0049] Through this embodiment, first, multiple historical attribute information and multiple historical behavior information of the user are extracted according to a preset period from the user's historical consumption data. The historical attribute information is mapped into a vector to obtain a first target vector, and the historical behavior information is mapped into a vector to obtain a second target vector. The dimensions of the first target vector and the second target vector are the same. The historical attribute information includes user age, gender, consumption preference, and credit rating, and the historical behavior information includes transaction, financial management behavior, and credit record. Then, the first target vector and the second target vector are weighted and fused in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector. After that, the third target vector is weighted and fused in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector. Then, the fourth target vector is determined as the node attribute, and the edges between the nodes are initialized according to the historical interaction information between the users to obtain a graph structure. Feature extraction is performed on the graph structure through a graph convolutional neural network to obtain a first target feature, and clustering is performed based on the first target feature to obtain user preference positioning. Finally, the user preference positioning is processed through a personalized recommendation model to obtain target commodity information, and the target commodity information is pushed to the user. The personalized recommendation model is trained based on the feedback data of multiple users on different commodities and the different user preference positionings corresponding to multiple users. This application obtains information from multiple dimensions based on a large amount of customer data, can comprehensively and completely reflect user characteristics, and converts the collaborative filtering positioning problem into a clustering problem of graph convolution during the user feature extraction process, can more comprehensively combine user characteristics to improve the positioning accuracy, and realizes the recommendation of target commodity information through a personalized recommendation model to improve user satisfaction. This method solves the problem that the effectiveness of the commodity push content in the prior art is relatively low, resulting in a decrease in user satisfaction.

[0050] In an optional implementation manner, in order to extract user feature information in the spatial domain, the above step S202 includes:

[0051] Step S2021, initialize a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head;

[0052] Specifically, according to the dimension X of the above first target vector and the above second target vector, a vector with dimension X is initialized, and all elements of the vector are set to 0 to obtain the above empty global head.

[0053] It can be understood that the empty global head plays the role of a global feature in the multi-head self-attention mechanism of the Transformer, which is used to further capture the global information between vectors.

[0054] Step S2022, perform weighted fusion on the empty global head, the first target vector, and the second target vector according to a preset weight to obtain a fifth target vector;

[0055] Specifically, according to expert prior knowledge, a preset weight matrix is determined, which contains the relative importance of different attribute information and behavior information. This matrix is used to perform weighted fusion on the empty global head vector, the first target vector, and the second target vector. Specifically, the first target vector and the second target vector are multiplied by their corresponding preset weights respectively, and then the results are added to the empty global head vector to obtain the fifth target vector.

[0056] It can be understood that the above process ensures the dominant position of key attribute and behavior information in the fusion process, and at the same time allows the supplementation of other information to form a comprehensive feature representation.

[0057] Step S2023, perform feature extraction on the fifth target vector through the multi-head self-attention mechanism of the Transformer structure to obtain multiple second target features, and the second target features correspond one by one to the attention heads of the Transformer structure;

[0058] Specifically, the fifth target vector is input into the multi-head self-attention module of the Transformer for further feature extraction. The multi-head self-attention mechanism decomposes the input vector into multiple sub-vectors, calculates the attention for each sub-vector separately, and then recombines the results. This method can capture the complex relationships between different parts of the input vector, especially suitable for processing multi-modal data.

[0059] In this application, the multi-head self-attention mechanism weights the features extracted for each dimension of the fifth target vector to generate multiple second target features, and each feature corresponds to an attention head.

[0060] Step S2024, fuse the second target features through the fully connected layer of the Transformer structure to obtain a third target vector.

[0061] Specifically, the second target features are input into the feed-forward fully connected layer of the Transformer to adjust the dimension of the vector and fuse the features generated by different attention heads, so as to obtain a more comprehensive unified vector, that is, the above-mentioned third target vector.

[0062] In the above embodiments, it can be understood that the Transformer structure is a deep learning model that processes sequential data through the multi-head self-attention mechanism and the feed-forward network, and is particularly suitable for processing text and time series information. In this application, the Transformer is used to fuse the attribute information and behavior information of the user at a specific time point, capture the correlation between different information through the self-attention mechanism, and at the same time adjust the importance of different features through preset weights to form a comprehensive vector including multi-dimensional information.

[0063] Through the above embodiments, the comprehensiveness and accuracy of user feature description can be significantly improved. Specifically, the introduction of the empty global head enhances the model's ability to capture global information; the use of preset weights ensures the dominant position of key features in the fusion process; the combination of the multi-head self-attention mechanism and the fully connected layer not only effectively processes multi-modal information, but also can automatically adjust the importance of different features to generate higher-quality user feature vectors.

[0064] In order to extract the user feature information in the time domain, in an alternative embodiment, the above step S203 includes:

[0065] Step S2031: Arrange the third target vector according to the time sequence based on the time stamp, and add positional encoding to the arranged third target vector to obtain a sixth target vector, where the positional encoding is used to uniquely identify the position of the third target vector in the sixth target vector;

[0066] Specifically, according to the time stamp of the above third target vector, arrange the above third vector in chronological order to form a sequence. Furthermore, in order to facilitate the Transformer structure to understand the sequential data, positional encoding is added to the serialized third target vector to ensure that the model can distinguish the features at different time points of the sequence.

[0067] In one embodiment, the above positional encoding can be a combination of sine and cosine functions, or other encoding methods that can displace and identify the third target vector in the sequence.

[0068] Step S2032: Initialize a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head;

[0069] Specifically, similarly, initialize a vector with the same dimension as the first target vector and the vector elements being 0 to obtain an empty global head.

[0070] Step S2033: Input the empty global header and the sixth target vector into the Transformer structure, and perform feature extraction through the multi-head self-attention mechanism of the Transformer structure to obtain multiple third target features;

[0071] Specifically, input the initialized empty global header vector and the sixth target vector into the Transformer structure. In the multi-head self-attention module, each head independently calculates the attention weights between each sequence vector, and these weights reflect the degree of correlation between the vectors. And position encoding is introduced to enable each head to adjust its attention weights according to the position information of the vectors, so as to capture the chronological information.

[0072] It can be understood that through the above operations, each head will generate a new third target feature, and these vectors can reflect the user's behavior patterns at different time points and their mutual influences.

[0073] Step S2034: Fuse the third target features through the fully connected layer of the Transformer structure to obtain the fourth target vector.

[0074] Specifically, the generated third target features are input into the fully connected layer of the Transformer, the dimension of the vector is adjusted, and the features obtained from multiple attention heads are fused to generate a more comprehensive unified vector, that is, the above-mentioned fourth target vector.

[0075] It can be understood that the fully connected layer ensures that the dimension of the output feature vector is consistent with the input through a series of linear transformations and non-linear activation functions, and at the same time, by fusing the features obtained from multi-head self-attention, the model's ability to capture the user's behavior patterns is enhanced.

[0076] Through the above embodiments, the present application introduces position encoding to ensure that the model can capture chronological information, which is convenient for extracting the dynamic changes of user behavior; and the combination of the multi-head self-attention mechanism and the fully connected layer enables the model to not only understand the user features at each time point, but also capture the long-term dependence relationships between features, and generate a fourth target vector that more comprehensively reflects the user's behavior pattern.

[0077] In order to construct the above graph structure, in an optional implementation manner, the above step S204 includes:

[0078] Step S2041: Determine each user as a node in the graph structure, and configure the fourth target vector as the node attribute of the node;

[0079] Specifically, taking user A as an example, in the case where the above-mentioned fourth target vector corresponding to user A has been extracted, determine A as a node in the graph structure, and configure the node attribute of the node according to the above-mentioned fourth target vector.

[0080] It is understandable that the above operations ensure that the nodes not only represent customers but also carry the characteristic information comprehensively obtained by the user in the time and space dimensions for subsequent operations to analyze and cluster.

[0081] Step S2042: Extract the interactions between users according to the historical interaction information, and connect the nodes corresponding to the users with interactions through edges;

[0082] Specifically, taking users A and B as examples, obtain the transaction records, interaction frequencies, etc. between users A and B, which can characterize the association between the two users, to determine whether there is an interaction relationship. If so, then connect nodes A and B through an edge.

[0083] It is understandable that the historical interaction information can be a transaction initiated by user A to user B, a comment or like on user A by user B on social media, etc. In this way, the graph structure not only contains the attribute information of the users themselves but also reflects the relationship network between users.

[0084] Step S2043: Extract the intensity and frequency of interactions between users according to the historical interaction information, initialize the weights corresponding to each edge according to the intensity and frequency, and map them to a preset interval to obtain the target weights;

[0085] Specifically, initialize the weight of the edge according to the intensity and frequency of the interaction between users A and B. For example, if users A and B have conducted multiple transactions with a large transaction amount in the past month, the weight of this edge may be higher. The quantified values of the interaction intensity and frequency will be mapped to a preset interval, such as (0, 1).

[0086] It is understandable that the above operations more intuitively reflect the degree of closeness of the relationship between users by introducing the weights of the edges.

[0087] Step S2044: Configure each target weight on the corresponding edge to obtain a graph structure.

[0088] Specifically, configure the initialized weight on the edge to complete the construction of the graph structure.

[0089] It can be understood that the graph structure can flexibly capture the relationships and attributes between entities. In this application, the graph structure is applied to user positioning and commodity recommendation to model the interaction network between users, as well as the behavioral characteristics and user positioning of users. By regarding users as nodes in the graph and user feature vectors as node attributes, the characteristics of users in different dimensions can be intuitively represented. At the same time, the interaction relationship between users is represented by the edges of the graph, and the weight of the edge reflects the intensity and frequency of the interaction. Graph Convolutional Network (GCN) and related algorithms can effectively extract features from such graph structures for in-depth analysis and prediction.

[0090] Through the above embodiments, by configuring the user dynamic feature vector as the node attribute and initializing the weight of the edge according to the interaction intensity and frequency between users, the constructed graph structure can more accurately reflect the dynamic behavior of users and the relationship between users. When the Graph Convolutional Network (GCN) performs feature extraction and user clustering, this structured information can help the algorithm capture deeper user preferences and behavioral patterns, improving the accuracy of clustering and the accuracy of personalized recommendation.

[0091] In order to analyze the graph structure to extract the comprehensive feature information of users, in an optional implementation manner, the above step S204 further includes:

[0092] Step S2045, converting the edges and target weights of the graph structure into an adjacency matrix for representation to obtain the first target matrix;

[0093] Specifically, taking users A and B as an example, the weight information of the edges in the graph structure is converted into an adjacency matrix, denoted as A (Adjacency Matrix). For the edge between users A and B, if there is a weight, then the corresponding position in the A matrix will store this weight value; otherwise, if there is no connection, 0 will be stored.

[0094] Step S2046, converting the nodes and node attributes of the graph structure into a diagonal matrix for representation to obtain the second target matrix;

[0095] Specifically, taking users A and B as an example, the attribute information in the graph structure is represented as a diagonal matrix, denoted as D (Diagonal Matrix). The non-diagonal elements in the diagonal matrix D are 0, and the diagonal elements contain the information of the node attributes (i.e., the fourth target vector). For example, the element in the first row and first column of the D matrix will be the value of the first element in the attribute vector of user A, and so on.

[0096] Step S2047, extracting the topological structure of the first target matrix through the first graph convolutional neural network to obtain the fourth target feature;

[0097] Specifically, the first graph convolutional neural network (GCN1) is used to process the adjacency matrix A to extract the topological structure features of the graph. GCN1 propagates the attribute information of nodes on the graph structure through convolutional operations, and each node will receive the information of its neighboring nodes and aggregate it. Repeat this process multiple times to gradually fuse the information of more distant nodes, thereby obtaining a fourth target feature representing the topological structure features.

[0098] Step S2048, the second graph convolutional neural network is used to extract the attribute information from the first target matrix and the second target matrix to obtain a fifth target feature;

[0099] Specifically, the second graph convolutional neural network (GCN2) is used to process both the adjacency matrix A and the diagonal matrix D simultaneously to extract the attribute information features of the nodes. GCN2 also uses convolutional operations to fuse node attributes and neighborhood information, but it focuses more on the attribute information of the nodes themselves. By taking the adjacency matrix A and the diagonal matrix D as inputs, GCN2 can capture how the attribute information of nodes affects their interactions with other nodes in the graph, generating a fifth target feature.

[0100] Step S2049, the fourth target feature and the fifth target feature are weighted and fused to obtain the first target feature.

[0101] Specifically, the extracted fourth target feature (topological structure feature) and fifth target feature (attribute information feature) are weighted and fused to obtain the first target feature.

[0102] It can be understood that the purpose of weighted fusion is to integrate the feature vectors reflecting the node topological structure and node attribute information together to form a more comprehensive and all - round representation. Weighted fusion can be achieved by learning a weight parameter λ, multiplying the fourth target feature by λ and adding it to the fifth target feature multiplied by (1 - λ), where the value of λ can be optimized through training to achieve the best fusion effect.

[0103] Through the above embodiments, by applying the graph convolutional neural network on the graph structure, it is possible to effectively extract the comprehensive features containing topological structure and attribute information. The extraction of topological structure features helps the model in collaborative filtering recommendation to understand the direct and indirect connections between users, so as to capture the popular trends and mutual influences in the user group. The extraction of attribute information helps the model to understand the preferences of individual users to achieve refined recommendation of personal attributes. Feature weighted fusion ensures comprehensive consideration of the user's social network relationship and individual attribute information, optimizing the extraction of user positioning.

[0104] To train the above - mentioned personalized recommendation model, in an alternative embodiment, before the personalized recommendation model processes the user preference positioning to obtain the target commodity information, the method further includes:

[0105] Step S301: Obtain corresponding historical feedback data according to different user preference positions. The historical feedback data includes the push feedback of users on different types of products.

[0106] Specifically, taking User A and User B as examples, they belong to different user preference positions respectively. For each user, obtain their push feedback on different types of products from the historical records, such as purchase behavior, click-through rate, collection, etc. These feedback data will be used to construct the training and test data sets for the personalized recommendation model.

[0107] Step S302: Construct a training data set and a test data set according to the historical feedback data and the user preference position.

[0108] Specifically, construct a training data set according to the user preference position and the historical feedback data. Each data point usually includes the user preference position as the input and the user's feedback on the product as the output. Then, divide a part of the training data set as the test data set to evaluate the performance of the model.

[0109] Step S303: Use the user preference position as the input data and the product with the highest purchase rate as the output data to train the alternative personalized recommendation model according to the training data set until the energy function of the output layer of the alternative personalized recommendation model is less than the preset error.

[0110] Specifically, use the training data set to train the personalized recommendation model (alternative personalized recommendation model). During the model training process, the input is the user preference position vector (for example, the preference position vector of User A), and the output is the product with the highest purchase rate (in the training stage, this output is known based on the historical feedback data). The goal of the model is to learn a mapping function so that the input user preference position can accurately predict the output product. During the training process, the model will continuously adjust its parameters, calculate the predicted output through forward propagation, and evaluate the difference between the predicted result and the actual feedback by calculating the model loss function (such as cross-entropy loss).

[0111] Step S304: Input the test data set into the alternative personalized recommendation model, obtain the output result through forward propagation, and perform backpropagation on the alternative personalized recommendation model according to the loss value between the actual data and the output result until the loss value of the alternative personalized recommendation model is less than or equal to the preset loss value, and determine the alternative personalized recommendation model as the personalized recommendation model.

[0112] Specifically, the test data set is input into the trained alternative personalized recommendation model, and the output result is obtained again through forward propagation, and the loss value between the model prediction result and the actual user feedback is calculated. If the loss value of the model is less than or equal to the preset loss threshold, the alternative personalized recommendation model is determined as the final personalized recommendation model for actual product recommendation.

[0113] Through the above embodiments, the alternative personalized recommendation model is determined as the final personalized recommendation model for actual product recommendation. The above operations can significantly improve the accuracy and relevance of product recommendation. Users A and B will obtain completely different recommendation results according to their different tendency orientations, reducing the push of products that they are not interested in and improving user satisfaction.

[0114] In an optional implementation manner, in order to implement clustering of the first target feature, the above step S204 further includes:

[0115] Cluster the first target feature according to the softmax function to obtain the user tendency orientation.

[0116] Specifically, the softmax function is a commonly used normalization function for converting a set of numerical values into a probability distribution, especially suitable for multi-classification problems. In cluster analysis, the softmax function is used to convert the feature representation of a node (user) into the probability of being assigned to different clusters, so as to perform soft clustering. Each class represents a user tendency, such as "financial management preference", "consumption orientation", etc., and the probability that a node is assigned to each class reflects the matching degree of the node with the tendency of that class. Through the softmax function, we can generate a tendency orientation vector for each user, where each element represents the probability that the user belongs to the corresponding tendency class.

[0117] In a specific implementation, the first target feature vector of each user is input into a softmax classifier, and the output of the classifier is a tendency orientation probability vector, where each element represents the probability that the user belongs to a certain tendency class. Assume that we set the number of clusters to K, then the softmax classifier will generate a K-dimensional probability vector, indicating the possibilities that users A, B, and C belong to K kinds of tendencies respectively. According to the output of the softmax classifier, we can determine a tendency orientation for each user. This usually involves selecting the class corresponding to the maximum value in the probability vector as the final tendency orientation of the user, or adopting a soft assignment method to perform "soft" classification on the user according to the probability values, that is, the user can belong to multiple tendencies simultaneously, and the degree is determined by the probability values.

[0118] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0119] The embodiment of the present application also provides a commodity push device based on user positioning. It should be noted that the commodity push device based on user positioning in the embodiment of the present application can be used to execute the method for commodity push based on user positioning provided by the embodiment of the present application. This device is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that realizes a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0120] The following introduces the commodity push device based on user positioning provided by the embodiment of the present application.

[0121] Figure 3 is a structural block diagram of the commodity push device based on user positioning according to the embodiment of the present application. As Figure 3 shown, the device includes:

[0122] An acquisition unit 10, configured to extract multiple historical attribute information and multiple historical behavior information of a user according to the user's historical consumption data at a preset period, map the historical attribute information into a vector to obtain a first target vector, map the historical behavior information into a vector to obtain a second target vector, the first target vector and the second target vector have the same dimension, the historical attribute information includes user age, gender, consumption preference and credit rating, and the historical behavior information includes transactions, financial management behaviors and credit records;

[0123] A first calculation unit 20, configured to perform weighted fusion of the first target vector and the second target vector in the spatial domain through a Transformer structure according to a preset weight to obtain a third target vector;

[0124] A second calculation unit 30, configured to perform weighted fusion of the third target vector in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector;

[0125] A construction unit 40, configured to determine the fourth target vector as node attributes, initialize the edges between nodes according to the historical interaction information between users to obtain a graph structure, perform feature extraction on the graph structure through a graph convolutional neural network to obtain a first target feature, and perform clustering based on the first target feature to obtain a user preference positioning;

[0126] A push unit 50 is configured to process the user preference orientation through a personalized recommendation model to obtain target commodity information, and push the target commodity information to the user. The personalized recommendation model is trained based on feedback data of multiple users on different commodities and different user preference orientations corresponding to multiple users.

[0127] Through this embodiment, the acquisition unit extracts multiple historical attribute information and multiple historical behavior information of the user according to the user's historical consumption data at a preset period, maps the historical attribute information into a vector to obtain a first target vector, maps the historical behavior information into a vector to obtain a second target vector. The first target vector and the second target vector have the same dimension. The historical attribute information includes user age, gender, consumption preference and credit rating, and the historical behavior information includes transactions, financial management behaviors and credit records. The first calculation unit performs weighted fusion on the first target vector and the second target vector in the spatial domain through a Transformer structure according to a preset weight to obtain a third target vector. The second calculation unit performs weighted fusion on the third target vector in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector. The construction unit determines the fourth target vector as the node attribute, initializes the edges between the nodes according to the historical interaction information between the users to obtain a graph structure, extracts features from the graph structure through a graph convolutional neural network to obtain a first target feature, and performs clustering based on the first target feature to obtain the user preference orientation. The push unit processes the user preference orientation through a personalized recommendation model to obtain target commodity information, and pushes the target commodity information to the user. The personalized recommendation model is trained based on feedback data of multiple users on different commodities and different user preference orientations corresponding to multiple users. This application obtains information from multiple dimensions based on a large amount of customer data, can comprehensively and completely reflect user characteristics, and converts the collaborative filtering positioning problem into a clustering problem of graph convolution during the user feature extraction process, can more comprehensively combine user characteristics to improve the positioning accuracy, and realizes the recommendation of target commodity information through a personalized recommendation model to improve user satisfaction. This device solves the problem that the effectiveness of the commodity push content in the prior art is relatively low, resulting in a decrease in user satisfaction.

[0128] In an alternative embodiment, to extract user feature information in the spatial domain, the above-mentioned first calculation unit includes:

[0129] A first initialization module is configured to initialize a vector with all dimensions being 0 based on the dimension of the first target vector to obtain an empty global head;

[0130] A first calculation module is configured to perform weighted fusion on the empty global head, the first target vector and the second target vector according to a preset weight to obtain a fifth target vector;

[0131] A second computing module, configured to extract features from a fifth target vector through the multi-head self-attention mechanism of a Transformer structure, obtaining multiple second target features, where the second target features correspond one-to-one to the attention heads of the Transformer structure;

[0132] A third computing module, configured to fuse the second target features through the fully connected layer of the Transformer structure, obtaining a third target vector.

[0133] In an alternative implementation, in order to extract user feature information in the time domain, the above-mentioned second computing unit includes:

[0134] A first processing module, configured to arrange the third target vector in chronological order according to timestamps, and add positional encoding to the arranged third target vector, obtaining a sixth target vector, where the positional encoding is used to uniquely identify the position of the third target vector in the sixth target vector;

[0135] A second initialization module, configured to initialize a vector with all dimensions being 0 in the dimension of the first target vector, obtaining an empty global head;

[0136] A fourth computing module, configured to input the empty global head and the sixth target vector into the Transformer structure, and perform feature extraction through the multi-head self-attention mechanism of the Transformer structure, obtaining multiple third target features;

[0137] A fifth computing module, configured to fuse the third target features through the fully connected layer of the Transformer structure, obtaining a fourth target vector.

[0138] In an alternative implementation, in order to construct the above-mentioned graph structure, the above-mentioned construction steps include:

[0139] A determination module, configured to determine each user as a node in the graph structure, and configure the fourth target vector as the node attribute of the node;

[0140] A second processing module, configured to extract the interactions between users according to historical interaction information, and connect the nodes corresponding to the users with existing interactions through edges;

[0141] A third initialization module, configured to extract the intensity and frequency of interactions between users according to historical interaction information, and initialize the weights corresponding to each edge according to the intensity and frequency, and map them to a preset interval, obtaining target weights;

[0142] A third processing module, configured to configure each target weight to the corresponding edge, obtaining the graph structure.

[0143] In an alternative implementation, to analyze the graph structure for extracting comprehensive user feature information, the above-mentioned construction unit further includes:

[0144] A fourth processing module, configured to convert the edges and target weights of the graph structure into an adjacency matrix for representation, obtaining a first target matrix;

[0145] A fifth processing module, configured to convert the nodes and node attributes of the graph structure into a diagonal matrix for representation, obtaining a second target matrix;

[0146] A sixth calculation module, configured to perform topological structure extraction on the first target matrix through a first graph convolutional neural network, obtaining a fourth target feature;

[0147] A seventh calculation module, configured to perform attribute information extraction on the first target matrix and the second target matrix through a second graph convolutional neural network, obtaining a fifth target feature;

[0148] An eighth calculation module, configured to perform weighted fusion on the fourth target feature and the fifth target feature, obtaining a first target feature.

[0149] In an alternative implementation, to train the above-mentioned personalized recommendation model, the above-mentioned device further includes:

[0150] A processing unit, configured to obtain corresponding historical feedback data according to different user preference positions before processing the user preference position through the personalized recommendation model to obtain target commodity information, where the historical feedback data includes the push feedback of the user on different types of commodities;

[0151] A first determination unit, configured to construct a training data set and a test data set according to the historical feedback data and the user preference position;

[0152] A training unit, configured to use the user preference position as input data and the commodity with the highest purchase rate as output data, and train an alternative personalized recommendation model according to the training data set until the energy function output by the output layer of the alternative personalized recommendation model is less than a preset error;

[0153] A second determination unit, configured to input the test data set into the alternative personalized recommendation model, obtain an output result through forward propagation, perform backpropagation on the alternative personalized recommendation model according to the loss value between the actual data and the output result until the loss value of the alternative personalized recommendation model is less than or equal to a preset loss value, and determine the alternative personalized recommendation model as the personalized recommendation model.

[0154] In an alternative implementation, to achieve clustering of the first target feature, the above-mentioned construction unit further includes:

[0155] A clustering module, configured to cluster the first target feature according to the softmax function to obtain user preference positioning.

[0156] The above-described commodity push device based on user positioning includes a processor and a memory. The above-described acquisition unit, first calculation unit, second calculation unit, construction unit, push unit, etc. are all stored in the memory as program units, and the processor executes the above-described program units stored in the memory to implement corresponding functions. The above-described modules are all located in the same processor; alternatively, the above-described respective modules are separately located in different processors in any combination form.

[0157] The processor includes a kernel, and the kernel retrieves the corresponding program unit from the memory. One or more kernels can be set, and the accuracy of pushing commodities can be improved by adjusting the kernel parameters.

[0158] The memory may include non-permanent memory in a computer-readable medium, forms such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one storage chip.

[0159] An embodiment of the present invention provides a computer-readable storage medium. The above-described computer-readable storage medium includes a stored program, wherein when the above-described program runs, it controls the device where the above-described computer-readable storage medium is located to execute the above-described commodity push method based on user positioning.

[0160] An embodiment of the present invention provides a processor. The above-described processor is used to run a program, wherein when the above-described program runs, it executes the above-described commodity push method based on user positioning.

[0161] An embodiment of the present invention provides a commodity push system. The communication system includes a processor, a memory, and a program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the commodity push method based on user positioning.

[0162] This application also provides a computer program product, which is adapted to execute a program initialized with at least the steps of the commodity push method based on user positioning when executed on a data processing device.

[0163] Obviously, those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed over a network composed of multiple computing devices. They can be implemented by program code executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. In this way, the present invention is not limited to any specific combination of hardware and software.

[0164] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or multiple flows and / or blocks

[0166] These computer program instructions can also be stored in a computer-readable memory capable of guiding the computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one or more of the flows Figure 1 or multiple flows and / or blocks

[0167] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable devices provide for implementing the functions in the flowFigure 1 one or more processes and / or blocks Figure 1 steps of functions specified in one or more blocks

[0168] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0169] The memory may include non-permanent memory in the computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

[0170] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology for information storage. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0171] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0172] From the above description, it can be seen that the above embodiments of the present application achieve the following technical effects:

[0173] 1) The method for pushing products based on user location in this application first extracts multiple historical attribute information and multiple historical behavior information of a user according to a preset cycle based on the user's historical consumption data. The historical attribute information is mapped into a vector to obtain a first target vector, and the historical behavior information is mapped into a vector to obtain a second target vector. The dimensions of the first target vector and the second target vector are the same. The historical attribute information includes user age, gender, consumption preference, and credit rating, and the historical behavior information includes transactions, financial management behaviors, and credit records. Then, the first target vector and the second target vector are weighted and fused in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector. After that, the third target vector is weighted and fused in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector. Then, the fourth target vector is determined as the node attribute, and the edges between nodes are initialized according to the historical interaction information between users to obtain a graph structure. Feature extraction is performed on the graph structure through a graph convolutional neural network to obtain a first target feature, and clustering is performed based on the first target feature to obtain the user preference location. Finally, the user preference location is processed through a personalized recommendation model to obtain target product information, and the target product information is pushed to the user. The personalized recommendation model is trained based on the feedback data of multiple users on different products and the different user preference locations corresponding to multiple users. This application obtains information from multiple dimensions based on a large amount of customer data, can comprehensively and completely reflect user characteristics, and converts the collaborative filtering location problem into a clustering problem of graph convolution during the user feature extraction process, can more comprehensively combine user characteristics to improve the location accuracy, and realizes the recommendation of target product information through a personalized recommendation model to improve user satisfaction. This method solves the problem that the effectiveness of the product push content in the prior art is low, resulting in a decrease in user satisfaction.

[0174] 2) The commodity push device based on user positioning in this application. The acquisition unit extracts multiple historical attribute information and multiple historical behavior information of the user according to the historical consumption data of the user at a preset cycle, maps the historical attribute information into a vector to obtain a first target vector, maps the historical behavior information into a vector to obtain a second target vector. The dimensions of the first target vector and the second target vector are the same. The historical attribute information includes user age, gender, consumption preference, and credit rating. The historical behavior information includes transactions, financial management behaviors, and credit records. The first calculation unit performs weighted fusion on the first target vector and the second target vector in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector. The second calculation unit performs weighted fusion on the third target vector in the time domain through a Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector. The construction unit determines the fourth target vector as the node attribute, initializes the edges between the nodes according to the historical interaction information between users to obtain a graph structure, extracts features from the graph structure through a graph convolutional neural network to obtain a first target feature, and performs clustering based on the first target feature to obtain the user preference positioning. The push unit processes the user preference positioning through a personalized recommendation model to obtain target commodity information and pushes the target commodity information to the user. The personalized recommendation model is trained according to the feedback data of multiple users on different commodities and the different user preference positionings corresponding to multiple users. This application obtains information from multiple dimensions based on a large amount of customer data, can comprehensively and completely reflect user characteristics, and converts the collaborative filtering positioning problem into a clustering problem of graph convolution during the user feature extraction process, can more comprehensively combine user characteristics to improve the positioning accuracy, and realizes the recommendation of target commodity information through a personalized recommendation model to improve user satisfaction. This device solves the problem that the effectiveness of the commodity push content in the prior art is relatively low, resulting in a decrease in user satisfaction.

[0175] The above are only the preferred embodiments of this application and are not used to limit this application. For those skilled in the art, this application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included in the protection scope of this application.

Claims

1. A commodity push method based on user positioning, characterized in that Including: Extracting multiple historical attribute information and multiple historical behavior information of the user according to a preset period based on the user's historical consumption data, mapping the historical attribute information into a vector to obtain a first target vector, mapping the historical behavior information into a vector to obtain a second target vector, where the dimensions of the first target vector and the second target vector are the same, the historical attribute information includes user age, gender, consumption preference, and credit rating, and the historical behavior information includes transactions, financial management behaviors, and credit records; Weightedly fusing the first target vector and the second target vector in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector; Weightedly fusing the third target vector in the time domain through the Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector; Determining the fourth target vector as node attributes, initializing the edges between nodes according to the historical interaction information between users to obtain a graph structure, extracting features from the graph structure through a graph convolutional neural network to obtain a first target feature, and performing clustering based on the first target feature to obtain user tendency positioning; Processing the user tendency positioning through a personalized recommendation model to obtain target commodity information, and pushing the target commodity information to the user, where the personalized recommendation model is trained according to the feedback data of multiple users on different commodities and different user tendency positionings corresponding to multiple users.

2. The method according to claim 1, characterized in that Weightedly fusing the first target vector and the second target vector in the spatial domain through a Transformer structure according to preset weights to obtain a third target vector, including: Initializing a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head; Weightedly fusing the empty global head, the first target vector, and the second target vector according to the preset weights to obtain a fifth target vector; Extracting features from the fifth target vector through the multi-head self-attention mechanism of the Transformer structure to obtain multiple second target features, where the second target features correspond one-to-one with the attention heads of the Transformer structure; Fusing the second target features through the fully connected layer of the Transformer structure to obtain the third target vector.

3. The method according to claim 1, wherein Weightedly fusing the third target vector in the time domain through the Transformer structure according to the timestamp of the third target vector to obtain a fourth target vector, including: Arranging the third target vector in chronological order according to the timestamp, and adding positional encoding to the arranged third target vector to obtain a sixth target vector, where the positional encoding is used to uniquely identify the position of the third target vector in the sixth target vector; Initializing a vector with all dimensions being 0 with the dimension of the first target vector to obtain an empty global head; Input the empty global header and the sixth target vector into the Transformer structure, and perform feature extraction through the multi-head self-attention mechanism of the Transformer structure to obtain multiple third target features; Fuse the third target features through the fully connected layer of the Transformer structure to obtain the fourth target vector.

4. The method according to claim 1, wherein Determine the fourth target vector as the node attribute, obtain the historical interaction information between the users, and initialize the edges between the nodes according to the historical interaction information to obtain a graph structure, including: Determine each user as a node in the graph structure, and configure the fourth target vector as the node attribute of the node; Extract the interactions between each user according to the historical interaction information, and connect the nodes corresponding to the users with interactions through edges; Extract the intensity and frequency of interactions between each user according to the historical interaction information, initialize the weights corresponding to each edge according to the intensity and frequency, and map them to a preset interval to obtain the target weights; Configure each target weight on the corresponding edge to obtain the graph structure.

5. The method according to claim 4, wherein Perform feature extraction on the graph structure through a graph convolutional neural network to obtain the first target feature, including: Convert the edges and the target weights of the graph structure into an adjacency matrix for representation to obtain the first target matrix; Convert the nodes and the node attributes of the graph structure into a diagonal matrix for representation to obtain the second target matrix; Extract the topological structure of the first target matrix through the first graph convolutional neural network to obtain the fourth target feature; Extract the attribute information of the first target matrix and the second target matrix through the second graph convolutional neural network to obtain the fifth target feature; Perform weighted fusion on the fourth target feature and the fifth target feature to obtain the first target feature.

6. The method according to claim 1, characterized in that, Before processing the user preference positioning through the personalized recommendation model to obtain the target commodity information, the method further includes: Obtain the corresponding historical feedback data according to different user preference positionings, where the historical feedback data includes the push feedback of the user on different types of commodities; Construct a training data set and a test data set according to the historical feedback data and the user preference positioning; Use the user preference positioning as the input data and the commodity with the highest purchase rate as the output data, and train the alternative personalized recommendation model according to the training data set until the energy function output by the output layer of the alternative personalized recommendation model is less than a preset error; Input the test data set into the alternative personalized recommendation model, obtain the output result through forward propagation, and perform backpropagation on the alternative personalized recommendation model according to the loss value between the actual data and the output result until the loss value of the alternative personalized recommendation model is less than or equal to the preset loss value, and determine the alternative personalized recommendation model as the personalized recommendation model.

7. The method according to claim 1, wherein Perform clustering based on the first target feature to obtain the user preference positioning, including: Cluster the first target feature according to the softmax function to obtain the user preference positioning.

8. A commodity push device based on user positioning, characterized in that, The device includes: An acquisition unit, configured to extract multiple historical attribute information and multiple historical behavior information of the user according to the user's historical consumption data at a preset period, map the historical attribute information into a vector to obtain a first target vector, map the historical behavior information into a vector to obtain a second target vector, the first target vector and the second target vector have the same dimension, the historical attribute information includes user age, gender, consumption preference and credit rating, and the historical behavior information includes transactions, financial management behaviors and credit records; A first calculation unit, configured to perform weighted fusion of the first target vector and the second target vector in the spatial domain through a Transformer structure according to a preset weight to obtain a third target vector; A second calculation unit, configured to perform weighted fusion of the third target vector in the time domain through the Transformer structure according to the time stamp of the third target vector to obtain a fourth target vector; A construction unit, configured to determine the fourth target vector as a node attribute, initialize the edges between nodes according to the historical interaction information between users to obtain a graph structure, extract features of the graph structure through a graph convolutional neural network to obtain a first target feature, and perform clustering based on the first target feature to obtain user preference positioning; A push unit, configured to process the user preference positioning through a personalized recommendation model to obtain target commodity information, and push the target commodity information to the user, where the personalized recommendation model is trained according to feedback data of multiple users on different commodities and different user preference positionings corresponding to the multiple users.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein when the program runs, it controls the device where the computer-readable storage medium is located to execute the method according to any one of claims 1 to 7.

10. A commodity push system, characterized in that, Including: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs include methods for executing the method according to any one of claims 1 to 7.

Images