A feature extraction method, device and storage medium

Abstract

The application discloses a feature extraction method, device and equipment. The feature extraction method comprises the following steps: acquiring a first feature map, wherein the first feature map comprises a plurality of nodes and an association relationship between the plurality of nodes, and the nodes are used for representing object attribute information involved in a recommendation processing procedure of an object; performing subgraph reconstruction on the plurality of nodes of the first feature map to obtain a second feature map, wherein the number of nodes in the second feature map is less than that in the first feature map; and performing feature screening on the second feature map to adjust the correlation degree between adjacent nodes with a non-zero correlation degree in the second feature map, so as to obtain a target feature map.

Description

Technical Field

[0001] This application relates to the field of neural network technology, and in particular to a feature extraction method, apparatus and computer-readable storage medium. Background Technology

[0002] With the development of neural network technology, its applications are becoming more and more widespread, and users are demanding higher and higher accuracy from neural networks. Feature extraction is a key operation in neural network technology. However, during the feature extraction process, the following problems are likely to occur in the selection of neighboring nodes of the feature map: when there are many neighboring nodes in the graph, how to extract effective neighboring nodes? And when there are few neighboring nodes in the graph, the original features obtained from the neighboring nodes or embedding (dimensionality reduction) are often noisy. How to extract useful features and transform them into corresponding features will further improve the model performance.

[0003] Currently, the main methods used for feature extraction are RandomWalk sampling and Self-Attention. However, RandomWalk sampling requires random walks to relevant nodes in the graph, which introduces significant randomness and cannot guarantee the accuracy of feature extraction. While Self-Attention can effectively amplify the original features, it cannot truly remove noise. Summary of the Invention

[0004] This application provides a feature extraction method, apparatus, and device.

[0005] According to a first aspect of this application, a feature extraction method is provided, the method comprising: obtaining a first feature map, the first feature map including multiple nodes and the relationships between the multiple nodes, the nodes being used to characterize object attribute information involved in the recommendation process of an object; reconstructing a subgraph from the multiple nodes of the first feature map to obtain a second feature map, the second feature map having fewer nodes than the first feature map; and performing feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, to obtain a target feature map.

[0006] According to one embodiment of this application, the step of reconstructing a second feature map by subgraphing multiple nodes of the first feature map includes: performing weighted sampling on multiple nodes of the first feature map to obtain a third feature map, wherein the number of nodes in the third feature map is less than that in the first feature map; and performing dynamic threshold denoising on multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map that have a correlation less than a threshold value to zero, thereby obtaining the second feature map.

[0007] According to one embodiment of this application, obtaining the first feature map includes: performing feature initialization on the plurality of nodes to obtain the first feature map.

[0008] According to one embodiment of this application, the weighted sampling process for multiple nodes of the first feature map includes: sampling neighbor nodes for any node of the multiple nodes of the first feature map using a random walk method; and normalizing the number of random walk visits of the nodes.

[0009] According to one embodiment of this application, dynamic threshold denoising processing is performed on multiple nodes of the third feature map, including: determining the correlation between two adjacent nodes in the third feature map; and reassigning the correlation between the two adjacent nodes according to the relationship between the correlation and a random dynamic threshold.

[0010] According to one embodiment of this application, the step of reassigning the correlation between the two adjacent nodes based on the relationship between the correlation and the random dynamic threshold includes: when the correlation between the two adjacent nodes is less than the random dynamic threshold, assigning the correlation between the two adjacent nodes to zero.

[0011] According to one embodiment of this application, the feature filtering of the second feature map includes: for a node in the second feature map, using the neighboring nodes of the node as channels of an attention network; and sequentially adjusting the correlation between neighboring nodes of the second feature map whose correlation is not zero through channel attention and spatial attention to obtain the target feature map.

[0012] According to one embodiment of this application, before performing dynamic threshold denoising on the plurality of nodes of the third feature map, the method further includes performing feature batch normalization on the plurality of nodes of the third feature map.

[0013] According to a second aspect of this application, a feature extraction apparatus is also provided, the apparatus comprising: an acquisition module for acquiring a first feature map, the first feature map including multiple nodes and the relationships between the multiple nodes, the nodes being used to characterize object attribute information involved in the recommendation process for objects; a reconstruction module for performing subgraph reconstruction on the multiple nodes of the first feature map to obtain a second feature map, the second feature map having fewer nodes than the first feature map; and a filtering module for performing feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, to obtain a target feature map.

[0014] According to a third aspect of this application, an apparatus is also provided, the apparatus comprising at least one processor, and at least one memory and a bus connected to the processor; wherein the processor and the memory communicate with each other via the bus; the processor is used to invoke program instructions in the memory to execute the above-described feature extraction method.

[0015] It should be understood that the teachings of this application are not required to achieve all the beneficial effects described above, but rather that a specific technical solution can achieve a specific technical effect, and other embodiments of this application can also achieve beneficial effects not mentioned above. Attached Figure Description

[0016] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, in which:

[0017] In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0018] Figure 1 A schematic diagram illustrating the implementation flow of a feature extraction method provided in an embodiment of this application is shown;

[0019] Figure 2 A schematic diagram illustrating the implementation flow of a feature extraction method provided in another embodiment of this application is shown;

[0020] Figure 3 The diagram illustrates the implementation flow of a specific application example of the feature extraction method according to an embodiment of this application;

[0021] Figure 4 A schematic diagram of the composition structure of the feature extraction device according to an embodiment of this application is shown;

[0022] Figure 5 A schematic diagram of the composition structure of the device according to an embodiment of this application is shown. Detailed Implementation

[0023] The principles and spirit of this application will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement this application, and are not intended to limit the scope of this application in any way. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0024] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Figure 1A schematic diagram illustrating the implementation flow of a feature extraction method provided in an embodiment of this application is shown.

[0026] refer to Figure 1 The feature extraction method of this application embodiment includes at least the following operation flow: Operation 101, obtaining a first feature map, the first feature map including multiple nodes and the association relationship between the multiple nodes, the nodes are used to represent the object attribute information involved in the object recommendation process; Operation 102, reconstructing the multiple nodes of the first feature map into a subgraph to obtain a second feature map, the number of nodes in the second feature map is less than that in the first feature map; Operation 103, performing feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, to obtain a target feature map.

[0027] In operation 101, a first feature map is obtained. The first feature map includes multiple nodes and the relationships between the multiple nodes. The nodes are used to represent the object attribute information involved in the recommendation process of the object.

[0028] In this embodiment of the application, features can be initialized for multiple nodes to obtain a first feature map. Here, the first feature map includes multiple nodes and the relationships between the multiple nodes, and the nodes are used to represent the object attribute information involved in the recommendation process of the object.

[0029] For example, in data processing using big data and neural network models, such as product recommendations and page recommendations, feature maps are involved. Multiple nodes in a feature map represent object attribute information related to the target recommendation object, such as product model, color, and configuration parameters. These nodes can also represent user attributes related to the target recommendation result, such as age, gender, and occupation.

[0030] In this embodiment of the application, each node can be described by a vector. The purpose of initializing the features of multiple nodes is to ensure that the vectors of multiple nodes have the same number of bits, so as to facilitate the data processing convenience and accuracy of all feature extraction operations. Here, taking node x in the first feature map as an example, node x and its adjacent nodes can be converted into corresponding initialized features, for example: x = [0.01, 002, ..., 025], where node x is a node randomly selected in the first feature map.

[0031] In operation 102, a subgraph reconstruction is performed on multiple nodes of the first feature map to obtain a second feature map, where the number of nodes in the second feature map is less than that in the first feature map.

[0032] In this embodiment of the application, in order to improve computational efficiency and save computational resources while ensuring the accuracy of data processing results, the nodes in the graph can be sampled using the PinSAGESampling (recall algorithm sampling) method to reconstruct the subgraph of multiple nodes in the first feature graph, thereby obtaining a second feature graph with fewer nodes than the first feature graph.

[0033] Specifically, PinSAGESampling can sample the neighboring nodes of any node in the first feature map. This is achieved by sampling the neighboring nodes of the node through a random walk, and the number of visits to the random walk nodes is normalized. In an infinite number of simulations, the number of visits to the random walk nodes is close to the PageRank score (a graph theory-based algorithm), which can fully reflect the weight of the influence of the neighboring nodes on the node. This ensures the randomness of sampling the first feature map, maintains the relationships between nodes in the first feature map, reduces the number of nodes, improves computation speed, and saves computational and storage resources.

[0034] In operation 103, feature filtering is performed on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, so as to obtain the target feature map.

[0035] In this embodiment of the application, for a node of the second feature map, the adjacent nodes of the node can be used as channels of the attention network, and the correlation between adjacent nodes with non-zero correlation of the second feature map can be adjusted sequentially through channel attention and spatial attention to obtain the target feature map, thereby realizing feature filtering of the second feature map.

[0036] In this embodiment of the application, an attention network can be used to adjust the weights of each node in the second feature map. Specifically, the weights of nodes with larger weights can be increased, while the weights of nodes with smaller weights can be decreased. This highlights the weights of key features and reduces the interference of nodes with weights below a set threshold on the final calculation result. For example, an attention algorithm including channel attention and spatial attention can be used to adjust the correlation between adjacent nodes in the second feature map that have non-zero correlation, thereby achieving feature filtering of the second feature map.

[0037] For example, for multiple nodes in the second feature map, a KCBAM (Convolutional Block Attention Module) network can be used to filter features in the second feature map. This adjusts the correlation between adjacent nodes with non-zero correlation in the second feature map, resulting in the target feature map. Specifically, when using the KCBAM network for feature filtering in the second feature map, the adjacent nodes of a given node can be treated as channels. First, the KCBAM network's channel attention is applied, and then spatial attention is used to adjust the correlation between adjacent nodes with non-zero correlation in the second feature map. Channel attention calculates the importance of each channel, while spatial attention aims to enable the model to adaptively learn the attention weights for different regions by introducing an attention module. In this way, the weights of nodes with higher weights are amplified, and the weights of nodes with lower weights are reduced, achieving feature filtering of the second feature map with greater accuracy.

[0038] Figure 2 A schematic diagram illustrating the implementation flow of a feature extraction method provided in another embodiment of this application is shown.

[0039] refer to Figure 2 The feature extraction method provided in this embodiment of the application includes at least the following operation flow: Operation 201, obtaining a first feature map, the first feature map including multiple nodes and the association relationship between the multiple nodes, the nodes being used to represent the object attribute information involved in the object recommendation process; Operation 202, performing weighted sampling processing on the multiple nodes of the first feature map to obtain a third feature map, the number of nodes in the third feature map being less than that in the first feature map; Operation 203, performing feature batch normalization processing on the multiple nodes of the third feature map; Operation 204, performing dynamic threshold denoising processing on the multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map with a correlation less than the threshold value to zero, to obtain a second feature map; Operation 205, performing feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map with a correlation of non-zero, to obtain a target feature map.

[0040] In operation 201, a first feature map is obtained. The first feature map includes multiple nodes and the relationships between the multiple nodes. The nodes are used to represent the object attribute information involved in the recommendation process of the object.

[0041] In operation 202, multiple nodes of the first feature map are subjected to weighted sampling to obtain the third feature map, which has fewer nodes than the first feature map.

[0042] In this embodiment of the application, a random walk method can be used to sample neighbor nodes for any node of multiple nodes in the first feature map, and the number of random walk visits of the nodes is normalized, thereby realizing weighted sampling of multiple nodes in the first feature map.

[0043] Here, the purpose of sampling multiple nodes is to speed up feature extraction. Therefore, the sampling process of multiple nodes must be sufficiently random to ensure the data distribution of the original data and accelerate model training, including the feature extraction process of the embodiments of this application.

[0044] In this embodiment of the application, a random walk sampling method is used, and the number of random walk visits to a node is normalized to ensure that sampling of a node does not change the data relationship between multiple nodes. Essentially, only the number of nodes is changed. That is, the essential difference between the second feature map and the first feature map is that the number of nodes is reduced. For example, the PinSAGESampling (recall algorithm sampling) method can be used to achieve random sampling of the nodes in the first feature map.

[0045] In operation 203, feature batch normalization is performed on multiple nodes of the third feature map.

[0046] In this embodiment of the application, the Batch (batch normalization) method can be used to calculate the correlation weight values ​​between each pair of nodes in the third feature map, so that in operation 204, the correlation weight values ​​that do not meet the set threshold can be reassigned.

[0047] In operation 204, dynamic threshold denoising is performed on multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map that have a correlation less than the threshold value to zero, thus obtaining the second feature map.

[0048] In this embodiment of the application, the correlation between two adjacent nodes in the third feature map determined in operation 203 can be reassigned based on the relationship between the correlation and the random dynamic threshold, thereby realizing dynamic threshold denoising processing for multiple nodes in the third feature map.

[0049] For two nodes whose correlation with each other is less than a threshold, their correlation is set to 0. However, this embodiment uses a random dynamic threshold, employing different thresholds across multiple rounds of model training to denoise multiple nodes in the third feature map. In practice, the model trained using a random dynamic threshold for feature map denoising performs better. Furthermore, to ensure the consistency of data format in the feature map,

[0050] In this embodiment of the application, the denoising process does not delete a node, but rather reassigns the correlation between nodes. Specifically, when the correlation between two adjacent nodes is less than a random dynamic threshold, the correlation between the two adjacent nodes is set to zero. Thus, based on the relationship between the correlation and the random dynamic threshold, the correlation between the two adjacent nodes is reassigned.

[0051] In operation 205, feature filtering is performed on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, so as to obtain the target feature map.

[0052] Among them, the other specific implementation processes of operations 201 and 205 are the same as those of operations 201 and 205. Figure 1 The specific implementation processes of operations 101 and 103 in the illustrated embodiments are similar and will not be described in detail here.

[0053] Figure 3 The diagram illustrates the implementation flow of a specific application example of the feature extraction method according to an embodiment of this application.

[0054] refer to Figure 3 In a specific application example of the feature extraction method in this application embodiment, for each x node ( Figure 3 Taking the "Xiaoxin" node as an example, first, the x node and its adjacent nodes are converted to the corresponding initialization features, for example: x = [0.01, 002, ..., 025].

[0055] Then, the PinSAGESampling method is used to sample neighboring nodes of node x (e.g., i7 node, metal A node, 3-in-1 node, RDD3 and thinkpad nodes, etc.). This involves sampling the neighboring nodes of node x through random walks, and then processing the number of visits to these nodes using L1 (absolute value summation normalization, also known as Manhattan distance) to obtain a normalized value. In an infinite number of simulations, this normalized value approximates the PageRank score, reflecting the weight of the influence of node x's neighboring nodes on node x.

[0056] Furthermore, Batch Normalization is used to normalize the features, and the correlation weights between node x and its neighboring nodes are calculated. For the features corresponding to the neighboring nodes of node x, the correlation weights between neighboring nodes with a value less than the random dynamic threshold and node x are assigned to zero. In this way, nodes with weak correlation to node x are removed, resulting in a new feature vector x' of the neighboring nodes of node x.

[0057] Finally, the channel attention and spatial attention of the KCBAM network are used to further filter the features in x' to obtain the purified feature vector x" (the feature diagram corresponding to the feature vector x" is not shown in the figure).

[0058] The model was trained using the feature extraction methods described above, and the experimental results are as follows:

[0059] method Hit@10 (hit rate) baseline(bs=32) 0.1015 baseline + kcbam (bs = 32) 0.1157 baseline+kcbam+λ(reduce=3,bs=32) 0.1163 baseline+kcbam+λ(reduce=15,bs=32) 0.1237 baseline+kcbam+λ(reduce=30, bs=32) 0.1229 baseline+kcbam+λ(reduce=30,bs=64) 0.1267 baseline+kcbam+λ(reduce=30, bs=128) 0.1251

[0060] Here, Baseline indicates that the model is trained using the feature extraction method of the traditional model training process, and Hit@10 (hit rate) is used to show the hit rate of the trained model. The higher the hit rate, the better the performance of the trained model.

[0061] Baseline+kcbam means that, based on the traditional feature extraction method for model training, the operation of "feature filtering on the second feature map" in this embodiment of the application is added.

[0062] Baseline+kcbam+λ represents the addition of the following operations to the traditional feature extraction method used in model training: "reconstructing a subgraph from multiple nodes of the first feature map to obtain a second feature map, and performing feature filtering on the second feature map" as described in this application embodiment. Here, reduce represents the number of sampling operations using a random dynamic threshold value.

[0063] Similarly, based on the feature extraction method described above, this application embodiment also provides a computer-readable storage medium storing a program. When the program is executed by a processor, the processor performs at least the following operation steps: Operation 101, obtaining a first feature map, the first feature map including multiple nodes and the relationships between the multiple nodes, the nodes being used to characterize the object attribute information involved in the recommendation process of the object; Operation 102, reconstructing a subgraph from the multiple nodes of the first feature map to obtain a second feature map, the number of nodes in the second feature map being less than that in the first feature map; Operation 103, performing feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, to obtain a target feature map.

[0064] Furthermore, based on the feature extraction method described above, embodiments of this application also provide a feature extraction apparatus, such as... Figure 4The device 40 includes: an acquisition module 401, used to acquire a first feature map, the first feature map including multiple nodes and the relationships between the multiple nodes, the nodes being used to characterize object attribute information involved in the recommendation process of the object; a reconstruction module 402, used to perform subgraph reconstruction on the multiple nodes of the first feature map to obtain a second feature map, the number of nodes in the second feature map being less than that in the first feature map; and a filtering module 403, used to perform feature filtering on the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, to obtain a target feature map.

[0065] In this embodiment of the application, the reconstruction module 402 includes: a sampling module, used to perform weighted sampling processing on multiple nodes of the first feature map to obtain a third feature map, wherein the number of nodes in the third feature map is less than that in the first feature map; and a denoising submodule, used to perform dynamic threshold denoising processing on multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map with a correlation less than a threshold value to zero, thereby obtaining a second feature map.

[0066] In this embodiment of the application, the acquisition module 401 includes: an initialization submodule, used to initialize the features of multiple nodes to obtain a first feature map.

[0067] In this embodiment of the application, the sampling module performs weighted sampling on multiple nodes of the first feature map, including: sampling neighbor nodes for any node of the multiple nodes of the first feature map using a random walk method; and normalizing the number of random walk visits of the nodes.

[0068] In this embodiment of the application, the denoising submodule performs dynamic threshold denoising processing on multiple nodes of the third feature map, including: determining the correlation between two adjacent nodes in the third feature map; and reassigning the correlation between the two adjacent nodes according to the relationship between the correlation and the random dynamic threshold.

[0069] In this embodiment of the application, the denoising submodule reassigns the correlation between two adjacent nodes based on the relationship between correlation and random dynamic threshold, including: when the correlation between two adjacent nodes is less than the random dynamic threshold, the correlation between the two adjacent nodes is assigned to zero.

[0070] In this embodiment of the application, the filtering module 403 includes: a channel submodule, used to treat the adjacent nodes of the nodes in the second feature map as channels of the attention network; and an adjustment submodule, used to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero through channel attention and spatial attention in turn, so as to obtain the target feature map.

[0071] In this embodiment of the present application, the apparatus 40 further includes a normalization module, which performs feature batch normalization on the multiple nodes of the third feature map before performing dynamic threshold denoising on the multiple nodes of the third feature map.

[0072] Furthermore, based on the feature extraction method described above, embodiments of this application also provide a device 50, such as... Figure 5 As shown, the device 50 includes at least one processor 501, and at least one memory 502 and bus 503 connected to the processor 501; wherein the processor 501 and the memory 502 communicate with each other through the bus 503; the processor 501 is used to call program instructions in the memory 502 to execute the above-mentioned feature extraction method.

[0073] It should be noted here that the above description of the feature extraction device and equipment embodiments is consistent with the foregoing Figures 1 to 3 The method embodiments shown are described similarly and have the same characteristics as described above. Figures 1 to 3 The beneficial effects of the methods illustrated are similar and will not be repeated here. For technical details not disclosed in the embodiments of the feature extraction apparatus and device of this application, please refer to the foregoing description of this application. Figures 1 to 3 The method embodiments shown are for understanding purposes only and will not be described in detail here for the sake of brevity.

[0074] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0076] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0077] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0078] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0079] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0080] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A feature extraction method applied to a recommendation system, the method comprising: A first feature map is obtained, which includes multiple nodes and the relationships between the nodes. The nodes are used to represent object attribute information related to the target recommendation object or user attributes related to the target recommendation result. The object attribute information includes at least one of the following: product model, color, configuration parameters. The user attributes include at least one of the following: age, gender, occupation. A second feature map is obtained by reconstructing a subgraph from multiple nodes of the first feature map, including: A third feature map is obtained by weighted sampling of multiple nodes in the first feature map, wherein the number of nodes in the third feature map is less than that in the first feature map. Dynamic threshold denoising is performed on multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map with a correlation less than the threshold value to zero, thereby obtaining the second feature map; The number of nodes in the second feature graph is less than that in the first feature graph; The second feature map is subjected to feature filtering to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, so as to obtain the target feature map.

2. The method according to claim 1, wherein obtaining the first feature map comprises: The features of the plurality of nodes are initialized to obtain the first feature map.

3. The method according to claim 1, wherein the weighted sampling process on the plurality of nodes of the first feature map comprises: A random walk approach is used to sample neighbor nodes for any node among multiple nodes in the first feature map. The number of random walk visits to a node is normalized.

4. The method according to claim 1, wherein dynamic threshold denoising processing is performed on multiple nodes of the third feature map, comprising: Determine the correlation between two adjacent nodes in the third feature graph; Based on the relationship between the correlation and the random dynamic threshold, the correlation between the two adjacent nodes is reassigned.

5. The method according to claim 4, wherein reassigning the correlation between the two adjacent nodes based on the relationship between the correlation and the random dynamic threshold comprises: When the correlation between two adjacent nodes is less than the random dynamic threshold, the correlation between the two adjacent nodes is set to zero.

6. The method according to claim 1, wherein the feature filtering of the second feature map comprises: For a node in the second feature map, the neighboring nodes of the node are used as channels of the attention network; The correlation between adjacent nodes with non-zero correlation in the second feature map is adjusted sequentially using channel attention and spatial attention to obtain the target feature map.

7. The method according to claim 1, prior to performing dynamic threshold denoising on the plurality of nodes of the third feature map, the method further includes: Feature batch normalization is performed on multiple nodes of the third feature map.

8. A feature extraction device for use in a recommendation system, the device comprising: The acquisition module is used to acquire a first feature map, which includes multiple nodes and the relationships between the multiple nodes. The nodes are used to represent object attribute information related to the target recommendation object or user attributes related to the target recommendation result. The object attribute information includes at least one of the following: product model, color, configuration parameters. The user attributes include at least one of the following: age, gender, occupation. The reconstruction module is used to reconstruct a subgraph from multiple nodes of the first feature map to obtain a second feature map, wherein the number of nodes in the second feature map is less than that in the first feature map. The reconstruction module is further configured to perform weighted sampling on multiple nodes of the first feature map to obtain a third feature map, wherein the number of nodes in the third feature map is less than that in the first feature map; and to perform dynamic threshold denoising on multiple nodes of the third feature map to assign the correlation of two adjacent nodes in the third feature map with a correlation less than a threshold value to zero, thereby obtaining a second feature map. The filtering module is used to filter features in the second feature map to adjust the correlation between adjacent nodes in the second feature map whose correlation is not zero, so as to obtain the target feature map.

9. An apparatus, the apparatus comprising at least one processor, and at least one memory and a bus connected to the processor; wherein, The processor and the memory communicate with each other via the bus; The processor is used to invoke program instructions in the memory to execute the feature extraction method according to any one of claims 1-7.

Images