CN111429161A - Feature extraction method, feature extraction device, storage medium, and electronic apparatus - Google Patents

Abstract

The embodiment of the invention relates to a feature extraction method, a feature extraction device, a storage medium and electronic equipment, and relates to the technical field of data processing. The method comprises the following steps: acquiring commodities in each edition block, wherein the edition block is a commodity set comprising at least one commodity; generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node; generating edges among the layout nodes, the commodity nodes and the user nodes based on the historical behaviors of the user on the layout and the commodity and the dependency relationship between the layout and the commodity so as to establish a topological graph; and extracting the characteristics of any node in the topological graph by utilizing a graph neural network. The invention improves the quality of the extracted features, ensures the comprehensiveness and accuracy of the features, and is beneficial to providing effective support for subsequent business analysis and business decision.

Description

Feature extraction method, feature extraction device, storage medium, and electronic apparatus

Technical Field

The embodiment of the invention relates to the technical field of data processing, in particular to a feature extraction method, a feature extraction device, a computer-readable storage medium and an electronic device.

Background

This section is intended to provide a background or context to the embodiments of the invention that are recited in the claims and the description herein is not admitted to be prior art by inclusion in this section.

For internet enterprises, data analysis is an important task, and particularly, user data and commodity data are effectively analyzed, so that corresponding support can be provided for business decision, the service quality of the enterprises can be improved, and more users can be attracted.

The feature extraction is an indispensable link in data analysis, for example, when user data is analyzed, user features are usually extracted from the data first, and then user portrayal or user classification is realized according to the user features, so that different customized services are provided for different users.

In the prior art, feature extraction is mostly realized based on dimensions set by people. For example, in order to extract user features, according to the characteristics of a data scene, a plurality of data indexes, that is, feature dimensions, such as the age, sex, login times, amount of purchased commodities and other dimensions of a user, which may be determined in an e-commerce scene, are determined in advance, and then user data is sorted according to each dimension to obtain corresponding user features.

Disclosure of Invention

However, in the prior art, the quality of the extracted features strongly depends on whether the dimension setting is reasonable: if the dimension setting is insufficient, the extracted features are insufficient, and if the dimension setting is excessive, redundant information exists in the extracted features. The quality of the extracted features is typically low due to the subjectivity of the artificially set dimensions. In addition, the user features are derived from user data, and the commodity features are derived from commodity data, so that the extracted features are relatively unilateral, and particularly in a relatively complex business scene, the features cannot reflect business relations, and therefore subsequent business analysis and business decision are not facilitated.

Therefore, an improved feature extraction method is highly needed to realize high-quality feature extraction, so that the extracted features can comprehensively reflect complex business relationships, thereby providing effective support for subsequent business analysis and business decision.

In this context, embodiments of the present invention are intended to provide a feature extraction method, a feature extraction apparatus, a computer-readable storage medium, and an electronic device.

According to a first aspect of embodiments of the present invention, there is provided a feature extraction method including: acquiring commodities in each edition block, wherein the edition block is a commodity set comprising at least one commodity; generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node; generating edges among the layout nodes, the commodity nodes and the user nodes based on the historical behaviors of the user on the layout and the commodity and the dependency relationship between the layout and the commodity so as to establish a topological graph; and extracting the characteristics of any node in the topological graph by utilizing a graph neural network.

In an optional embodiment, the user node is generated by: clustering users to obtain a plurality of user categories; and respectively generating a corresponding user node for each user category.

In an optional implementation manner, the clustering the users to obtain a plurality of user categories includes: clustering the users based on the basic information of the users to obtain a plurality of first user categories; and clustering the users based on the behavior attributes of the users to obtain a plurality of second user categories.

In an alternative embodiment, when a new user is generated, the method further comprises: if the new user has historical data, determining the behavior attribute of the new user according to the historical data of the new user, and classifying the new user into a corresponding second user category according to the behavior attribute of the new user; and if the new user does not have the historical data, classifying the new user into a corresponding first user category according to the basic information of the new user.

In an optional implementation manner, the generating, based on the historical behaviors of the user on the layout blocks and the commodities and the affiliations between the layout blocks and the commodities, edges between each of the layout block nodes, commodity nodes and user nodes to establish the topological graph includes: and counting the historical interaction behaviors of the users in each user category and the sections in the historical data to determine edges between the user nodes and the section nodes and edge weights.

In an optional implementation manner, the generating, based on the historical behaviors of the user on the layout blocks and the commodities and the affiliations between the layout blocks and the commodities, edges between each of the layout block nodes, commodity nodes and user nodes to establish the topological graph includes: and counting historical interaction behaviors of the users in each user category and the commodities in historical data to determine edges between the user nodes and the commodity nodes and edge weights.

In an optional implementation manner, the generating, based on the historical behaviors of the user on the layout blocks and the commodities and the affiliations between the layout blocks and the commodities, edges between each of the layout block nodes, commodity nodes and user nodes to establish the topological graph includes: and counting the word frequency-inverse text frequency of each commodity in each block, and determining edges and edge weights between the block nodes and the commodity nodes according to the word frequency-inverse text frequency.

In an optional implementation manner, the generating, based on the historical behaviors of the user on the layout blocks and the commodities and the affiliations between the layout blocks and the commodities, edges between each of the layout block nodes, commodity nodes and user nodes to establish the topological graph includes: and determining the relevance among the commodities according to the historical behaviors of the user aiming at the commodities, and determining the edges and the edge weight among the commodity nodes according to the relevance.

In an alternative embodiment, the graph neural network is trained by: inputting the topological graph into a graph neural network to be trained, extracting the characteristics of any two nodes in the topological graph, and calculating the characteristic similarity of the two nodes; obtaining labels aiming at the two nodes based on whether edges exist between the two nodes; and updating the parameters of the graph neural network by using the feature similarity and the label of the two nodes.

In an optional embodiment, the updating the parameters of the graph neural network by using the feature similarity and the label of the two nodes includes: substituting the feature similarity and the label of the two nodes into the following loss function:

updating parameters of the graph neural network through the loss function, wherein L represents the loss function, n is the number of samples, V represents a node set in the topological graph, u and V represent any two nodes in V, N (u) represents a set of nodes with edges to u, F (u) represents a set of nodes without edges to u, E (u | V) ═ 1 represents that when u and V have edges, the label is 1, E (u | V) ═ 0 represents that when u and V have no edges, the label is 0, and Z (u | V) ═ 0 represents that when u and V have no edges, the label is 0_uDenotes the characteristic of u, Z_vFeature of v, P (Z)_u|Z_v) Representing the feature similarity of u and v.

In an optional implementation, the extracting, by using a graph neural network, the feature of any node in the topological graph includes: determining a key neighbor node of the any node in the topological graph; and aggregating the input characteristics of the key neighbor nodes, and performing iteration of a preset number of layers through the graph neural network to obtain the characteristics of any node.

In an optional embodiment, the determining a critical neighbor node of the any node in the topology graph includes: and if any node is a block node, all commodity nodes are selected from neighbor nodes of the node to serve as key neighbor nodes of the node.

In an optional embodiment, the determining a critical neighbor node of the any node in the topology graph includes: and if any node is a commodity node, selecting a first preset number of user nodes, a second preset number of commodity nodes and all layout nodes from the neighbor nodes as key neighbor nodes of any node.

In an optional embodiment, the determining a critical neighbor node of the any node in the topology graph includes: and if any node is a user node, selecting a third preset number of nodes from neighbor nodes of the node as a key neighbor node of the node.

In an optional embodiment, the input features of the section nodes are obtained by: and extracting characteristics of at least one of title texts, category information and poster images of the sections, and coding to obtain input characteristics of section nodes corresponding to the sections.

In an alternative embodiment, the input feature of the commodity node is obtained by: extracting characteristics of at least one of title texts, category information, price information and commodity images of the commodities, and coding to obtain input characteristics of the section nodes corresponding to the commodities.

In an alternative embodiment, the input characteristics of the user node are obtained by: and extracting characteristics of at least one of the basic information and the behavior attribute of the user, and coding to obtain the input characteristics of the user node corresponding to the user.

In an optional implementation, the extracting, by using a graph neural network, the feature of any node in the topological graph includes: extracting the characteristics of each node in the topological graph by using a graph neural network to obtain a node characteristic table; the method further comprises the following steps: when the characteristics of the object to be processed are extracted, the target node corresponding to the object to be processed is determined in the topological graph, and the characteristics of the target node are obtained by searching in the node characteristic table.

In an optional implementation manner, if the object to be processed is a target user, after obtaining the characteristics of the target user, the method further includes: and generating recommendation information which aims at the target user and is related to the layout according to the similarity between the characteristics of the target user and the characteristics of each layout node and/or the similarity between the characteristics of the target user and the characteristics of each commodity node.

According to a second aspect of embodiments of the present invention, there is provided a feature extraction device including: the system comprises an information acquisition module, a display module and a display module, wherein the information acquisition module is used for acquiring commodities in each edition block, and the edition block is a commodity set comprising at least one commodity; the node generation module is used for generating layout nodes corresponding to the layouts, commodity nodes corresponding to the commodities and user nodes; the edge generation module is used for generating edges among the layout nodes, the commodity nodes and the user nodes based on historical behaviors of the users on the layouts and the commodities and the dependency relationship between the layouts and the commodities so as to establish a topological graph; and the topological graph processing module is used for extracting the characteristics of any node in the topological graph by using the graph neural network.

In an optional embodiment, the node generating module is configured to generate the user node by: clustering users to obtain a plurality of user categories; and respectively generating a corresponding user node for each user category.

In an optional implementation manner, the node generation module is configured to cluster the users by: clustering the users based on the basic information of the users to obtain a plurality of first user categories; and clustering the users based on the behavior attributes of the users to obtain a plurality of second user categories.

In an optional embodiment, the node generation module is further configured to, when a new user is generated, determine the user category of the new user by: if the new user has historical data, determining the behavior attribute of the new user according to the historical data of the new user, and classifying the new user into a corresponding second user category according to the behavior attribute of the new user; and if the new user does not have the historical data, classifying the new user into a corresponding first user category according to the basic information of the new user.

In an optional embodiment, the edge generation module comprises a user-tile edge generation unit configured to: and counting the historical interaction behaviors of the users in each user category and the sections in the historical data to determine edges between the user nodes and the section nodes and edge weights.

In an alternative embodiment, the edge generation module comprises a user-commodity edge generation unit for: and counting historical interaction behaviors of the users in each user category and the commodities in historical data to determine edges between the user nodes and the commodity nodes and edge weights.

In an optional embodiment, the edge generating module comprises a section-commodity edge generating unit, configured to: and counting the word frequency-inverse text frequency of each commodity in each block, and determining edges and edge weights between the block nodes and the commodity nodes according to the word frequency-inverse text frequency.

In an alternative embodiment, the edge generation module comprises a commodity-commodity edge generation unit for: and determining the relevance among the commodities according to the historical behaviors of the user aiming at the commodities, and determining the edges and the edge weight among the commodity nodes according to the relevance.

In an alternative embodiment, the topology graph processing module comprises a model training unit for: inputting the topological graph into a graph neural network to be trained, extracting the characteristics of any two nodes in the topological graph, and calculating the characteristic similarity of the two nodes; obtaining labels aiming at the two nodes based on whether edges exist between the two nodes; and updating the parameters of the graph neural network by using the feature similarity and the label of the two nodes.

In an optional embodiment, the model training unit is further configured to update parameters of the graph neural network by: substituting the feature similarity and the label of the two nodes into the following loss function:

updating parameters of the graph neural network through the loss function, wherein L represents the loss function, n is the number of samples, V represents a node set in the topological graph, u and V represent any two nodes in V, N (u) represents a set of nodes with edges to u, F (u) represents a set of nodes without edges to u, E (u | V) ═ 1 represents that when u and V have edges, the label is 1, E (u | V) ═ 0 represents that when u and V have no edges, the label is 0, and Z (u | V) ═ 0 represents that when u and V have no edges, the label is 0_uDenotes the characteristic of u, Z_vFeature of v, P (Z)_u|Z_v) Representing the feature similarity of u and v.

In an optional implementation, the topology graph processing module includes: a key neighbor node determining unit, configured to determine a key neighbor node of the any node in the topology map; and the input feature aggregation unit is used for aggregating the input features of the key neighbor nodes and performing iteration of a preset number of layers through the graph neural network to obtain the features of any node.

In an optional implementation manner, the critical neighbor node determining unit is configured to: and if any node is a block node, all commodity nodes are selected from neighbor nodes of the node to serve as key neighbor nodes of the node.

In an optional implementation manner, the critical neighbor node determining unit is configured to: and if any node is a commodity node, selecting a first preset number of user nodes, a second preset number of commodity nodes and all layout nodes from the neighbor nodes as key neighbor nodes of any node.

In an optional implementation manner, the critical neighbor node determining unit is configured to: and if any node is a user node, selecting a third preset number of nodes from neighbor nodes of the node as a key neighbor node of the node.

In an optional implementation manner, the input feature aggregation unit is further configured to obtain the input features of the section nodes by: and extracting characteristics of at least one of title texts, category information and poster images of the sections, and coding to obtain input characteristics of section nodes corresponding to the sections.

In an optional implementation manner, the input feature aggregation unit is further configured to acquire the input features of the commodity node by: extracting characteristics of at least one of title texts, category information, price information and commodity images of the commodities, and coding to obtain input characteristics of the section nodes corresponding to the commodities.

In an optional implementation manner, the input feature aggregation unit is further configured to acquire the input features of the user node by: and extracting characteristics of at least one of the basic information and the behavior attribute of the user, and coding to obtain the input characteristics of the user node corresponding to the user.

In an optional implementation manner, the topology graph processing module is further configured to: extracting the characteristics of each node in the topological graph by using a graph neural network to obtain a node characteristic table; when the characteristics of the object to be processed are extracted, the target node corresponding to the object to be processed is determined in the topological graph, and the characteristics of the target node are obtained by searching in the node characteristic table.

In an optional implementation manner, the apparatus further includes a recommendation information generation module configured to: if the object to be processed is a target user, after the characteristics of the target user are obtained, generating recommendation information which is specific to the target user and related to the layout according to the similarity between the characteristics of the target user and the characteristics of each layout node and/or the similarity between the characteristics of the target user and the characteristics of each commodity node.

According to a third aspect of embodiments of the present invention, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements any one of the feature extraction methods described above.

According to a fourth aspect of the embodiments of the present invention, there is provided an electronic apparatus including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform any of the above feature extraction methods via execution of the executable instructions.

According to the feature extraction method, the feature extraction device, the computer-readable storage medium and the electronic device of the embodiment of the invention, commodities in each version block are firstly obtained; then generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node; generating edges among the nodes of each edition, the nodes of the commodities and the nodes of the users based on the historical behaviors of the users on the edition and the commodities and the dependency relationship between the edition and the commodities so as to establish a topological graph; and finally, extracting the characteristics of any node in the topological graph by using the graph neural network. On one hand, the scheme constructs a complete topological graph based on the relationship among three service objects, namely the layout, the commodity and the user, and is suitable for a complex scene with service layout division so as to cover different aspects of the scene; the characteristics of the three service objects are extracted from the topological graph by using the graph neural network, so that the comprehensiveness and the accuracy of the characteristics are ensured. On the other hand, the feature extraction does not depend on artificial setting of feature dimensions, the relation of three service objects is learned from the perspective of a machine, and the features are abstracted, so that the quality of the extracted features is improved, and effective support is provided for subsequent service analysis and service decision.

Drawings

The above and other objects, features and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description read in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which:

FIG. 1 shows a flow diagram of a feature extraction method according to an embodiment of the invention;

FIG. 2 illustrates a flow diagram for generating user nodes according to an embodiment of the present invention;

FIG. 3 illustrates a partial schematic diagram of a topology according to an embodiment of the present invention;

FIG. 4 shows a schematic diagram of extracting node features according to an embodiment of the invention;

FIG. 5 illustrates a flow diagram for training a graph neural network according to an embodiment of the present invention;

FIG. 6 illustrates a flow diagram for extracting node features according to an embodiment of the invention;

fig. 7 is a block diagram showing the structure of a feature extraction apparatus according to an embodiment of the present invention;

FIG. 8 illustrates a schematic diagram of a storage medium according to an embodiment of the present invention; and

fig. 9 shows a block diagram of the structure of an electronic apparatus according to an embodiment of the present invention.

In the drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

Detailed Description

The principles and spirit of the present invention will be described with reference to a number of exemplary embodiments. It is understood that these embodiments are given solely for the purpose of enabling those skilled in the art to better understand and to practice the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

As will be appreciated by one skilled in the art, embodiments of the present invention may be embodied as a system, apparatus, device, method, or computer program product. Thus, the present invention may be embodied in the form of: entirely hardware, entirely software (including firmware, resident software, micro-code, etc.), or a combination of hardware and software.

According to an embodiment of the present invention, a feature extraction method, a feature extraction device, a computer-readable storage medium, and an electronic apparatus are provided.

In this document, any number of elements in the drawings is by way of example and not by way of limitation, and any nomenclature is used solely for differentiation and not by way of limitation.

The principles and spirit of the present invention are explained in detail below with reference to several representative embodiments of the invention.

Summary of The Invention

The inventor finds that in the existing feature extraction method, the quality of the extracted features strongly depends on whether the dimension setting is reasonable or not: if the dimension setting is insufficient, the extracted features are insufficient, and if the dimension setting is excessive, redundant information exists in the extracted features. The quality of the extracted features is typically low due to the subjectivity of the artificially set dimensions. In addition, the user features are derived from user data, and the commodity features are derived from commodity data, so that the extracted features are relatively unilateral, and particularly in a relatively complex business scene, the features cannot reflect business relations, and therefore subsequent business analysis and business decision are not facilitated.

In view of the above, the basic idea of the present invention is: a feature extraction method, a feature extraction device, a computer-readable storage medium and an electronic apparatus are provided, first, commodities in each version block are obtained; then generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node; generating edges among the nodes of each edition, the nodes of the commodities and the nodes of the users based on the historical behaviors of the users on the edition and the commodities and the dependency relationship between the edition and the commodities so as to establish a topological graph; and finally, extracting the characteristics of any node in the topological graph by using the graph neural network. On one hand, the scheme constructs a complete topological graph based on the relationship among three service objects, namely the layout, the commodity and the user, and is suitable for a complex scene with service layout division so as to cover different aspects of the scene; the characteristics of the three service objects are extracted from the topological graph by using the graph neural network, so that the comprehensiveness and the accuracy of the characteristics are ensured. On the other hand, the feature extraction does not depend on artificial setting of feature dimensions, the relation of three service objects is learned from the perspective of a machine, and the features are abstracted, so that the quality of the extracted features is improved, and effective support is provided for subsequent service analysis and service decision.

Having described the general principles of the invention, various non-limiting embodiments of the invention are described in detail below.

Application scene overview

It should be noted that the following application scenarios are merely illustrated to facilitate understanding of the spirit and principles of the present invention, and the embodiments of the present invention are not limited in this respect. Rather, embodiments of the present invention may be applied to any scenario where applicable.

The invention extracts the characteristics of any business object based on the relationship among three business objects, namely the plate, the commodity and the user in the business scene, and can be applied to all scenes with the three business objects, such as: in an e-commerce scene, forming a plate block for a commodity according to certain commonality, for example, the commodities of the same type or the same brand belong to one plate block, and analyzing the data of the plate block, the commodity and a user by the method of the exemplary embodiment to extract the characteristics of the plate block, the commodity or the user; in service scenes of online music, electronic books, audio books and the like, the music, electronic books and audio books which need to be purchased by a user are used as commodities and are formed into sections, the same theme music belongs to one section, the same type of electronic books and audio books belong to one section, and the characteristics of any object are extracted by the method of the exemplary embodiment.

Exemplary method

An exemplary embodiment of the present invention first provides a feature extraction method, and a flow of the feature extraction method will be specifically described below with reference to steps S110 to S140 shown in fig. 1.

Step S110, the commodities in each block are acquired.

Wherein the plate is a commodity set comprising at least one commodity. In practical applications, the layout blocks may be various specific business objects, such as: the layout can be a service unit, for example, each service is divided into a corresponding service layout according to the service type, wherein the service layout comprises the goods related to the service; the layout block may also be a commodity display page, for example, in a web page or an App (Application program), commodities with the same type or the same brand and the like having a certain commonality are displayed in the same page, which is a layout block; the layout may also be an organization unit in the product database, for example, the product data close in time is stored in the same unit, i.e. a layout, according to the time sequence.

In the exemplary embodiment, all the layout blocks can be obtained within a certain range according to the business requirements, and then, the commodities in each of the layout blocks can be counted. For example, in an e-commerce activity, such as a festival sales promotion, the platform side organizes a certain number of venues, each venue corresponds to a special subject of a commodity, such as a digital product venue, a clothing venue, a food venue, etc., and each venue includes a certain number of commodities. With each venue in the event as a block, the merchandise in each block may be obtained.

Next, the exemplary embodiment constructs a topological graph among the layout blocks, the products, and the users, and mines and extracts features of the layout blocks, the products, and the users through the topological graph. The topology graph includes two types of elements: nodes and edges between nodes. The generation method of the node and the edge is described below by step S120 and step S130, respectively.

And step S120, generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node.

The following respectively describes the layout node, the commodity node and the user node in detail:

and (6) a block node. Generally, a layout name or a layout number can be determined for each layout, a layout node is generated corresponding to each layout, and the node is recorded through the layout name or the layout number.

And (6) commodity nodes. Each commodity has a commodity name or a commodity code, a commodity node is generated corresponding to each commodity, and the node is recorded through the commodity name or the commodity code.

And (4) a user node. Referring to fig. 2, the user node may be generated by the following steps S210 and S220:

step S210, clustering the users to obtain a plurality of user categories;

step S220 is to generate a corresponding user node for each user category.

Because the number of users is generally very large and may exceed the plate and the commodity by several orders of magnitude, if a user node is formed for each user, a great amount of redundant information may be introduced. Therefore, in the exemplary embodiment, users are clustered, each user category correspondingly generates a user node, and the node is recorded by the name or number of the user category, so that the number of the user nodes is greatly reduced, and the information consistency is improved. The user clustering method can cluster all users, for example, all registered users, and can also cluster users within a certain range according to service requirements, for example, clustering users who have purchased a behavior, or clustering users who have participated in a certain activity of an e-commerce, and the like.

The present exemplary embodiment does not specifically limit the clustering method. For example, a K-means algorithm (K-means algorithm) may be adopted to determine a suitable K value (for example, the number may be equal to the number of commodity nodes, or twice or three times the number of commodity nodes, etc.) according to the number of the section nodes and the commodity nodes; randomly selecting K users as K clustering centers; calculating the distance between other users and each clustering center, forming the data of each user into vectors, calculating Euclidean distance, cosine similarity and the like among the vectors, and dividing each user into the nearest clustering centers to obtain K user categories; updating a clustering center in each user category, and then clustering again; and (4) through multiple iterations, the clustering center in each user category tends to be stable, and clustering is finished to obtain the final K user categories. The collaborative filtering algorithm can also be adopted to count the interactive behaviors of each user on different commodities, which can include browsing behaviors, collecting behaviors, purchasing behaviors, commenting behaviors and the like, similar users are determined based on the same or similar commodity combination, and the similar users form a user category, so that a plurality of user categories are obtained.

User data typically includes two aspects: basic information of the user, such as age, gender, occupation, residential area, etc.; the behavior data of the user specifically refers to historical data corresponding to historical behaviors of the user in a website, a platform or an App, such as login behavior, browsing behavior, collection behavior, purchasing behavior and the like. In an alternative embodiment, the users may be clustered from the above two-aspect data, and step S210 may include:

clustering users based on the basic information of the users to obtain a plurality of first user categories;

and clustering the users based on the behavior attributes of the users to obtain a plurality of second user categories.

The behavior attribute is standardized data obtained by performing statistical arrangement on the behavior data of the user, and for example, a plurality of dimensions, such as the number of login times, login time, the number of collected commodities, the number of next commodities, the number of purchased commodities, the amount of consumption, and the like of the latest month may be set, and the behavior data of each user is arranged according to the dimensions, and discretized and encoded to obtain the behavior attribute in the form of a vector.

The first user category and the second user category respectively represent user categories obtained by clustering the two-aspect data. And for the same batch of users, performing first clustering on the basis of the basic information of the users to obtain a first user category, and performing second clustering on the basis of the behavior attributes of the users to obtain a second user category. Typically each user belongs to both a first user category and a second user category. Of course, in clustering, there may be a portion of users that have only basic information, no behavioral data (e.g., users that have not been active after registration), and then these users may be classified only into the corresponding first user category.

And generating a corresponding user node for each of the first user category and the second user category. The obtained user nodes have stronger pertinence, and the division of the users is more accurate.

When a new user is generated subsequently, if the new user has historical data, determining the behavior attribute of the new user according to the historical data of the new user, and dividing the new user into corresponding second user categories according to the behavior attribute of the new user; and if the new user does not have the historical data, classifying the new user into the corresponding first user category according to the basic information of the new user. For example, if the new user is actually an old user of the platform, the user has historical data since the first time the user participates in the e-commerce activity, the new user is in the e-commerce activity; and sorting the historical data to obtain the behavior attribute of the user, then calculating the distance between the behavior attribute and each second user category, and dividing the behavior attribute into the second user category with the closest distance, thereby determining the user node corresponding to the user. If the new user is a user registered on the platform, no behavior data exists, the distance between the new user and each first user category is calculated through basic information submitted by the user during registration, and the new user is divided into the first user categories closest to the first user categories, so that the user node corresponding to the user is determined. Therefore, the cold start problem of the new user can be solved, namely, the new user can be corresponding to a certain user node under the condition that the new user does not have any historical data, and the user characteristics of the new user can be extracted in the subsequent steps.

Step S130, based on the historical behaviors of the user on the layout blocks and the commodities and the dependency relationship between the layout blocks and the commodities, generating edges among the layout block nodes, the commodity nodes and the user nodes so as to establish a topological graph.

When generating edges between nodes for a topology graph, the association between nodes needs to be considered. The exemplary embodiment relates to three types of nodes, the association relationship is complex, and is determined mainly based on the historical behaviors of the user on the layout blocks and the commodities and the dependency relationship between the layout blocks and the commodities, and the association relationship related to the edges between different types of nodes is also different, so that the association relationship needs to be processed respectively.

The edges among the above layout nodes, commodity nodes and user nodes may include: the layout node comprises an edge between a user node and the layout node, an edge between the user node and the commodity node, an edge between the layout node and the commodity node, and an edge between the commodity node and the commodity node. These four sides are specifically described below:

edges between user nodes and section nodes. Historical interaction behavior of the users in each user category and the sections in the historical data can be counted to determine edges between the user nodes and the section nodes and edge weights. The historical interaction behavior of the user and the layout block refers to all behaviors which are performed by the user and related to the layout block, and includes but is not limited to: and clicking the browsing block by the user, leaving a message in the block by the user, getting a coupon in the block by the user, purchasing a commodity in the block by the user and the like. For example, all user behavior data in one historical e-commerce activity, i.e. historical data, may be obtained; counting the historical interactive behaviors of each user category and each plate in the historical data, for example, when counting the historical interactive behaviors of the user category U1 on the plate A, counting the times of clicking and browsing the plate A by all users in U1, the number of messages left in the plate A, the number of coupons picked up in the plate A, the number of purchased articles in the plate A and the like; then, normalizing and weighting the numerical values to obtain a comprehensive numerical value serving as the edge weight between the user node U1 and the layout node A; if the edge weight is less than the first threshold, it is assumed that there is no edge between the user node U1 and the chunk node A. The first threshold is a criterion for determining whether the user category is related to the section, and may be set according to experience or actual requirements, such as 0.5. Whether an edge exists between each user node and each section or not and the weight of the edge can be obtained through the method. The higher the interaction degree between the user and the layout, the closer the relationship between the user and the layout is, the higher the probability of having an edge is, and the higher the weight of the edge is.

In an alternative embodiment, a new section may exist, for example, a new section is designed in the e-commerce activity, the section does not appear in the previous activity, and therefore, the historical interaction behavior of the user with the section does not exist. For the situation, historical interaction behaviors of the user on the commodities in the section can be counted to predict the interaction degree of the user on the section, so that the edge and the weight of the edge are calculated. Thereby, the cold start problem with the block can be solved.

An edge between the user node and the commodity node. Historical interaction behaviors of the users and the commodities in each user category in the historical data can be counted to determine edges between the user nodes and the commodity nodes and edge weights. The historical interaction behavior of the user and the commodity refers to all behaviors related to the commodity, which are performed by the user, and includes but is not limited to: the user browses commodities, collects commodities, purchases commodities, reviews commodities and the like. For example, all user behavior data, i.e., historical data, on the platform over a recent period of time (e.g., a recent year) may be obtained; counting the historical interactive behaviors of each user type and each commodity in the historical data, for example, when counting the historical interactive behaviors of the user type U1 on the commodity a1, counting the times of clicking and browsing the commodity a1, the times of collecting the commodity a1, the number of purchasing the commodity a1, the times of commenting the commodity a1 and the like of all users in U1; then, normalizing and weighting the numerical values to obtain a comprehensive numerical value serving as an edge weight between the user node U1 and the commodity node a 1; if the edge weight is less than the second threshold, it is assumed that there is no edge between user node U1 and commodity node a 1. The second threshold is a criterion for determining whether the user category is related to the commodity, and may be set according to experience or actual demand, and may be 0.5. Whether an edge exists between each user node and each commodity or not and the weight of the edge can be obtained through the mode. The higher the interaction degree between the user and the commodity is, the closer the relationship between the user and the commodity is, the higher the probability of having the edge is, and the higher the weight of the edge is.

Edges between the section nodes and the commodity nodes. Generally, an edge can be generated between a section node having a dependency relationship and a product node according to the dependency relationship between the section and the product, and no edge exists if there is no dependency relationship. In some scenes, the number of the layout blocks is small, each layout block comprises a large number of commodities, and thus a large number of edges are formed between the layout block nodes and the commodity nodes, so that the discrimination of different layout blocks is low, and important relationship characteristics are lost. Based on this, in one embodiment, TF-IDF (Term Frequency-Inverse text Frequency) of each commodity in each version block may be counted, and edges between the version block nodes and the commodity nodes and edge weights may be determined according to the TF-IDF. For example, assume a total of f pieces, wherein the piece a includes k items, a1, a2, …, ak; counting the historical data related to the block A, and respectively counting the number of pieces of historical data (one piece of data represents one-time user behavior) related to a1, a2, … and ak, wherein the number of pieces of historical data is respectively marked as q1, q2, … and qk, and q1 actually reflects the number of times of the user behavior related to a 1; for the commodity a1 in the block A, its TF (word frequency) is:

it should be understood that in the statistics of q1, q2, …, qk, different weights may be set for different types of user behaviors, generally speaking, browsing weight < review weight < favorite weight < purchase weight, and then the weights are accumulated for all relevant historical data of each commodity to obtain a cumulative weight value, so as to replace q1, q2, …, qk, and then formula (1) may also use the cumulative weight value to calculate TF.

Then, the number of the blocks containing the commodity a1 in all the f blocks is counted, and f is assumed to be_a1Then IDF (inverse text frequency) of the article a1 is:

note that, in IDF calculation, 1 is usually added to the denominator, i.e., 1+ f_a1To prevent the denominator from being 0. In the present exemplary embodiment, the selected product belongs to at least one section, and therefore 1 may not be added. Further, a certain coefficient value (e.g., 1) may be added to the IDF as a whole to balance the TF and IDF values, and this is also applicable to the present exemplary embodiment.

And finally, calculating TF-IDF of the commodity a1 in the block A as:

after obtaining the TF-IDF of each commodity in the section, proper normalization processing can be carried out to obtain the edge weight between the section node and the commodity node. And if the edge weight is smaller than a third threshold value, determining that no edge exists between the section node and the commodity node. The third threshold is a criterion for determining whether the plate is related to the product, and may be set according to experience or actual requirements, such as 0.6.

Commodity nodes and edges between commodity nodes. The relevance among the commodities can be determined according to the historical behaviors of the user aiming at the commodities, and the edges and the edge weight among the commodity nodes can be determined according to the relevance. Specifically, the purchasing behavior in the historical behavior may be extracted, order data corresponding to each purchasing behavior may be obtained, which commodities the user has purchased at each time may be determined, the commodities having a higher purchasing probability at the same time may be determined as the associated commodities, and an edge may be generated between commodity nodes of the associated commodities. The present exemplary embodiment may use any association analysis algorithm, taking Apriori algorithm (an association rule mining algorithm) as an example, count the number of times of purchasing behavior in historical behaviors as h, the number of times of purchasing commodity a1 as h (a1), the number of times of purchasing commodity a2 as h (a2), and the number of times of purchasing a1 and a2 as h (a1, a2), calculate the support degrees of commodities a1 and a 2:

calculate confidence for items a1 and a 2:

then, threshold values may be set for the support degree and the confidence degree, respectively, when both the support degree and the confidence degree are greater than the threshold values, it is determined that there is an edge between the commodity node a1 and the commodity node a2, and the weight of the edge is obtained through weighted calculation according to the support degree and the confidence degree. Or weighting and calculating the support degree and the confidence degree to obtain the association degree of a1 and a 2; if the association degree is greater than the fourth threshold value, an edge is considered to exist between the commodity node a1 and the commodity node a2, and the association degree is the weight of the edge; if the degree of association is less than the fourth threshold, it is deemed that there is no edge between commodity node a1 and commodity node a 2. The fourth threshold is a measure of the degree of association between the commodities, and may be set according to experience or actual requirements, such as 0.4.

The first, second, third and fourth thresholds are predetermined thresholds for different parameters, and the four thresholds are not related to each other, and may have the same or different values.

Through the above steps S120 and S130, a topological graph about the layout, the product, and the user can be established, and is shown in fig. 3 as a part of the actual topological graph. In the topological graph, the section nodes are connected with the user nodes and the commodity nodes, the user nodes are connected with the section nodes and the commodity nodes, and the commodity nodes are connected with the three nodes; the section nodes are not connected with the section nodes, and the user nodes are not connected with the user nodes. The topological graph actually comprises complex association relations among three business objects, namely the layout, the commodity and the user. Feature extraction may be performed based on the topology map as follows.

And step S140, extracting the characteristics of any node in the topological graph by using the graph neural network.

GNN (Graph Neural Network) is an algorithmic model that runs on Graph structures using deep learning. The Graph (Graph) processed by GNN generally refers to a topological Graph, including nodes and edges, rather than an image consisting of pixel points. In contrast, the pixel image mainly includes visual information, the topological graph mainly includes relationship information, and GNN is a model for processing the relationship information in the topological graph.

Fig. 4 is a schematic diagram illustrating the extraction of node features based on the Graph SAmple and aggregations (Graph SAmple and aggreGatE) algorithm in GNN. The left side in fig. 4 is a topological graph, which is input into GNN, and any one of the nodes (e.g., node a) is selected as a target node, as shown in the right side in fig. 4, and the feature of a is obtained through multilayer aggregation processing. The algorithm is represented as follows:

and inputting a topological graph G (V, E), wherein V represents a node set in the topological graph, and E represents an edge set in the topological graph. Input characteristics of any node u in the input V, in x_uRepresents; the input features represent initialization features of the externally acquired nodes, e.g. plates, quotientsThe input characteristics of the product and the user can be determined by their attributes, and the input characteristics of the three nodes will be described in detail below. Inputting a preset number K of layers and weights W of 1-K layers (the part is the main parameter of GNN and can also comprise a bias parameter)ⁱ，

The activation function σ is set, for example, a sigmoid function (S-type activation function) may be employed. Setting an aggregation function AG_i，

The set of neighbor nodes of any node u is defined as n (u), and a neighbor node refers to a node directly connected with the target node (i.e. having an edge with the target node), such as the neighbor node a in fig. 4 is B, C, D.

For any node u, there are 1-K layers of potential vectors (embedding) in the GNN

The latent vector is a representation of the intermediate features, and the algorithmic process is shown in equation (6):

where CONCAT denotes splicing.

Equation (6) shows that the processing of GNN includes two loop nestings: a large loop refers to iterating the number of layers, and a small loop refers to iterating each node in each layer. The specific treatment process comprises the following steps: starting from the first layer, performing potential vector representation on each node in each layer, taking the node a in fig. 4 as an example, obtaining a neighbor node B, C, D of the node a, aggregating the potential vector of the previous layer B, C, D, splicing the aggregation result with the potential vector of the previous layer a, and obtaining the potential vector of the node a in the current layer through weighting and activating operation (nonlinear conversion); and operating each other node at the current layer by using the same method to obtain the potential vector representation of each node at the current layer. Then enter the next layer and weighAnd repeating the operation of the previous layer to obtain the potential vector representation of each node in the next layer. Iterating layer by layer until the K layer is reached, and taking the potential vector of each node in the K layer as the output feature of each node so as to obtain the output feature Z of any node u_u。

In the present exemplary embodiment, the aggregation function AG_iCan be implemented in any form, including but not limited to the following three forms:

① average aggregation, i.e. averaging each dimension of the potential vector of the previous layer of neighboring nodes, as follows:

② L STM (L ong Short Term Memory) aggregation, using L STM network algorithm, firstly sequencing neighbor nodes (such as random sequencing, or sequencing according to actual scene characteristics), then sequentially inputting the previous layer potential vector of each neighbor node into different L STM units, sequentially processing through L STM, and outputting aggregation result at the last unit

③ Pooling aggregation first non-linear transformation of the previous layer potential vector for each neighbor node (e.g., equivalent to a single fully-connected layer, requiring the use of a weight W_poolBias b and activation function σ) and then processed by Max or Mean Pooling per dimension to obtain the aggregation results as follows:

the handling of GNNs is described above. For the target node a in fig. 4, the feature extraction process is as follows: acquiring a neighbor node B, C, D of A in the first layer, aggregating the input features of B, C, D, and carrying out nonlinear conversion on the aggregated input features and the input features of A to obtain a first-layer potential vector of A; and in the second layer, the first layer potential vectors of the neighbor node B, C, D are aggregated and are subjected to nonlinear conversion with the first layer potential vectors of A to obtain a second layer potential vector … of A, and the layer-by-layer iteration is carried out until the K layer is reached to obtain a K layer potential vector of A, so that the characteristics of A are obtained. Fig. 4 shows a case where K is 3.

By the method, the features can be extracted from the layout nodes, commodity nodes or user nodes in the topological graph to obtain the layout features, commodity features or user features.

The following describes a training process of the neural network, and as shown in fig. 5, the training process may include the following steps S510 to S530:

step S510, inputting the topological graph into a graph neural network to be trained, extracting the characteristics of any two nodes in the topological graph, and calculating the characteristic similarity of the two nodes;

step S520, obtaining labels aiming at two nodes based on whether edges exist between the two nodes;

and step S530, updating parameters of the neural network of the graph by using the feature similarity and the labels of the two nodes.

In the graph neural network to be trained, each parameter may be obtained by initialization, for example, parameters are generated by random initialization within a certain numerical range, including the above weight and bias parameters. In the exemplary embodiment, any two nodes are taken as a group of samples, for example, the node (u, v), the input sample data is the number of the two nodes, and the feature Z of the two nodes is output through the GNN_uAnd Z_vAnd calculating the feature similarity of the two nodes.

In step S130, edges between nodes are determined, and therefore labels for any two nodes can be obtained based on the existence of an edge between the two nodes, for example, the label of (u, v) is E (u | v), when an edge exists between u and v, E (u | v) is 1, and when no edge exists between u and v, E (u | v) is 0.

In general, when there is an edge between u and v, the feature similarity of u and v should be high, and when there is no edge between u and v, the feature similarity should be low. Based on this, a loss function can be constructed, as follows:

where L denotes the loss function, n is the number of samples, and illustratively, if there are m nodes in the topology, the maximum value of n is

P(Z_u|Z_v) Representing the feature similarity of u and v. The feature similarity may be calculated by various algorithms, for example, cosine similarity, euclidean distance, etc. of two feature vectors may be calculated.

It should be noted that, in the actual training, any two nodes in the topological graph are taken as a set of sample data, multiple sets of sample data can be extracted from the topological graph, and the same node can form different sample data with different nodes. The sample data includes positive samples (two nodes with edges) and negative samples (two nodes without edges). By optimizing the loss function over a batch (meaning a batch of samples), the GNN parameters are updated using gradient descent, etc.

In an alternative embodiment, step S530 may include:

substituting the feature similarity and label of the two nodes into the following loss function:

parameters of the neural network of the graph are updated by the loss function.

V represents a node set in the topological graph, and u and V represent any two nodes in V; n (u) denotes a set of nodes having an edge with u, f (u) denotes a set of nodes having no edge with u, and when E (u | v) ═ 1 denotes that u has an edge with v, the label is 1, and when E (u | v) ═ 0 denotes that u has no edge with v, the label is 0. And limiting the characteristic similarity of u and v to be between 0 and 1.

Through the loss function (10), the GNN can be optimized, so that the feature similarity of the positive sample tends to 1, and the feature similarity of the negative sample tends to 0, and the accuracy of the GNN extraction features is improved.

In an alternative embodiment, the loss function may also be constructed in the form of cross entropy, as follows:

in a complex topological graph, the connection relationship between nodes is dense, one node may have many neighbor nodes, the calculation amount is large when aggregation processing is performed, and redundant information may be introduced. Based on this, in an alternative embodiment, as shown with reference to fig. 6, step S140 may include the following steps S610 and S620:

step S610, determining a key neighbor node of any node in the topological graph;

and S620, aggregating the input characteristics of the key neighbor nodes, and performing iteration of preset layers through a graph neural network to obtain the characteristics of any node.

The key neighbor node is obtained by screening or discarding the neighbor nodes to some extent, and can be regarded as a subset of the neighbor nodes. In this exemplary embodiment, the topological graph includes three different types of nodes, so the topological graph is a heterogeneous graph, the connection relationships of the three types of nodes are different, and the key neighbor nodes may be determined in different manners, which is specifically described below.

Firstly, if any node is a block node, all commodity nodes are selected from neighbor nodes of the node and serve as key neighbor nodes of the node. In other words, the user node is discarded from the neighbor node of the block node, because the relationship between the user node and the block node is sparse, the value of the contained information may be low, and the information density can be improved and the operation can be simplified after discarding.

And secondly, if any node is a commodity node, selecting a first preset number of user nodes, a second preset number of commodity nodes and all layout nodes from the neighbor nodes as key neighbor nodes. In other words, the user nodes and the commodity nodes are properly discarded in the neighbor nodes, and all the layout nodes are reserved, because the interaction between the user and the commodity, and the interaction between the commodity and the commodity are dense and may contain some noise information. When the first preset number of user nodes and the second preset number of commodity nodes are selected, the nodes with higher edge weight can be selected, and random walk algorithm can also be adopted for selection. The first preset number and the second preset number may be set according to experience or actual requirements, for example, when the number of the user nodes is greater than that of the commodity nodes, the first preset number may be greater than the second preset number.

And thirdly, if any node is a user node, selecting a third preset number of nodes from the neighbor nodes of the user node as a key neighbor node of any node. The neighbor nodes of the user node comprise the block nodes and the commodity nodes, and generally, the edges of the user node are dense and may contain some noise information, so that the user node can be appropriately discarded. When the third preset number of nodes are selected, the nodes with higher edge weights can be selected, and random walk algorithm can also be adopted for selection. The third preset number may be set according to experience or actual requirements, and is generally related to the complexity of the topological graph.

After the key neighbor nodes are determined, the features of the key neighbor nodes are only aggregated during GNN processing, so that the calculated amount is reduced, the generalization processing capability of the GNN is improved, and the overfitting condition is reduced.

In the GNN processing process, the features are obtained through multiple layer iterations, where the initial features are input features, and then potential vectors of each layer are obtained through each layer iteration, and finally output features are obtained. It can be seen that the input features are a prerequisite for performing the iteration of the features. How to obtain the input features of the three nodes is specifically described below.

The input characteristics of the layout nodes are obtained through the following modes:

and extracting characteristics of at least one of title texts, category information and poster images of the sections, and coding to obtain input characteristics of section nodes corresponding to the sections. The title text can adopt a natural language processing model to extract text and semantic features; the category information refers to which service category the block belongs to, and can be used as an independent dimension; the poster image generally refers to an image of a plate cover or other propaganda pages, and after image features are extracted by using a convolutional neural network, the poster image and the former two features are combined or spliced through dimension reduction processing to obtain input features of corresponding plate nodes. When each feature is processed, the feature is generally subjected to digitization, discretization, normalization, vectorization and the like, that is, the feature is encoded according to a preset standard rule.

The input characteristics of the commodity node are obtained through the following modes:

and extracting characteristics of at least one of title texts, category information, price information and commodity images of the commodities, and coding to obtain input characteristics of the section nodes corresponding to the commodities. The processing of the title text of the commodity can refer to the processing of the title text of the plate; the category information may include a multi-level category to which the commodity belongs, and may form corresponding dimensions respectively; price information may be treated as a separate dimension; the commodity image generally refers to an appearance image in a commodity sales page, and can be combined or spliced with the first three features after dimension reduction by referring to the processing of the poster image of the plate block to obtain the input features of the corresponding commodity nodes.

The input characteristics of the user node are obtained through the following modes:

and extracting characteristics of at least one of the basic information and the behavior attribute of the user, and coding to obtain the input characteristics of the user node corresponding to the user. The basic information may include the age, sex, occupation, residential area, etc. of the user, the behavior attributes may include login behavior, browsing behavior, collection behavior, purchasing behavior, etc., and the data are sorted and encoded through a predetermined dimension to obtain the input characteristics of the corresponding user node.

It should be noted that, for the convenience of the feature aggregation process, the input features of the three nodes may have the same number of dimensions. When the input features are acquired, all the feature categories do not need to be acquired, some missing features are allowed, and 0 or other preset values may be filled in the corresponding dimensions, which is not limited in the present exemplary embodiment.

The construction of the topological graph and the training process of the GNN can be completed in an online mode. After the completion, the characteristics of each node in the topological graph can be extracted by using the graph neural network to obtain a node characteristic table. The node feature table comprises features corresponding to each section node, commodity node or user node. And when the characteristics of the object to be processed need to be extracted subsequently, determining a target node corresponding to the object to be processed in the topological graph, and then searching in the node characteristic table to obtain the characteristics of the target node. For example, when the object to be processed is a target section, finding a corresponding section node according to an identification (Identity Document, which is referred to as a unique identifier), and then obtaining the characteristics of the section node, namely the characteristics of the target section, by looking up a table; when the object to be processed is a target commodity, finding a corresponding commodity node according to the commodity ID, and then obtaining the characteristics of the commodity node, namely the characteristics of the target commodity through table lookup; when the object to be processed is a target user, if the object to be processed is an old user, the corresponding user node is found according to the user ID, if the object to be processed is a new user, the user node to which the object to be processed belongs is determined according to the basic information of the user, and then the characteristics of the user node, namely the characteristics of the target user, are obtained through table lookup.

Further, if the object to be processed is the target user, after the characteristics of the target user are obtained, recommendation information related to the layout for the target user can be generated according to the similarity between the characteristics of the target user and the characteristics of the layout nodes and/or the similarity between the characteristics of the target user and the characteristics of the commodity nodes. Specifically, the similarity of the features of the target user and the features of the nodes of the blocks is calculated respectively, the similarity of the features of the nodes of the commodity is calculated, a plurality of blocks or commodities with the highest similarity are selected, recommendation information is formed, and the recommendation information is recommended to the target user. In the internet activity, the method can predict the topic or commodity page interested by the target user and recommend the topic or commodity page to the target user, thereby making up the blank in the aspect of recommending the activity topic in the prior art and realizing high-quality customized service for the target user.

Exemplary devices

Having described the feature extraction method of the exemplary embodiment of the present invention, next, a feature extraction device of the exemplary embodiment of the present invention will be described with reference to fig. 7.

As shown in fig. 7, the feature extraction apparatus 700 may include:

the information acquisition module 710 is configured to acquire commodities in each plate, where each plate is a commodity set including at least one commodity;

a node generating module 720, configured to generate a layout node corresponding to each layout, a commodity node corresponding to each commodity, and a user node;

an edge generating module 730, configured to generate edges among the layout nodes, the commodity nodes, and the user nodes based on historical behaviors of the user on the layouts and the commodities, and an affiliation between the layouts and the commodities, so as to establish a topological graph;

and the topological graph processing module 740 is used for extracting the characteristics of any node in the topological graph by using the graph neural network.

In an alternative embodiment, the node generating module 720 is configured to generate the user node by:

clustering users to obtain a plurality of user categories; and respectively generating a corresponding user node for each user category.

In an alternative embodiment, the node generating module 720 is configured to cluster the users by:

clustering users based on the basic information of the users to obtain a plurality of first user categories;

and clustering the users based on the behavior attributes of the users to obtain a plurality of second user categories.

In an alternative embodiment, the node generating module 720 is further configured to determine the user category of the new user by:

if the new user has historical data, determining the behavior attribute of the new user according to the historical data of the new user, and dividing the new user into corresponding second user categories according to the behavior attribute of the new user;

and if the new user does not have the historical data, classifying the new user into a corresponding first user category according to the basic information of the new user.

In an alternative embodiment, the edge generation module 730 includes a user-tile edge generation unit for:

and counting the historical interaction behaviors of the users and the sections in each user category in the historical data to determine edges between the user nodes and the section nodes and edge weights.

In an alternative embodiment, the edge generation module 730 comprises a user-commodity edge generation unit for:

and counting the historical interaction behaviors of the users and the commodities in each user category in the historical data to determine edges between the user nodes and the commodity nodes and edge weights.

In an alternative embodiment, the edge generation module 730 includes a section-commodity edge generation unit for:

and counting the word frequency-inverse text frequency of each commodity in each block, and determining edges and edge weights between block nodes and commodity nodes according to the word frequency-inverse text frequency.

In an alternative embodiment, the edge generation module 730 includes an item-to-item edge generation unit for:

and determining the relevance among the commodities according to the historical behaviors of the user aiming at the commodities, and determining the edges and the edge weight among the commodity nodes according to the relevance.

In an alternative embodiment, the topology graph processing module 740 includes a model training unit for:

inputting the topological graph into a graph neural network to be trained, extracting the characteristics of any two nodes in the topological graph, and calculating the characteristic similarity of the two nodes;

obtaining labels aiming at the two nodes based on whether edges exist between the two nodes; and updating the parameters of the neural network of the graph by using the feature similarity and the labels of the two nodes.

In an alternative embodiment, the model training unit is further configured to update the parameters of the graph neural network by:

substituting the feature similarity and the label of the two nodes into the following loss function:

updating parameters of the neural network of the graph through a loss function;

wherein L denotes a loss function, n denotes the number of samples, V denotes a set of nodes in a topological graph, u and V denote any two nodes in V, n (u) denotes a set of nodes with an edge to u, f (u) denotes a set of nodes without an edge to u, E (u | V) ═ 1 denotes that u and V have an edge, the label is 1, E (u | V) ═ 0 denotes that u and V have no edge, the label is 0, Z denotes that_uDenotes the characteristic of u, Z_vFeature of v, P (Z)_u|Z_v) And (d) representing the feature similarity of the point pair (u, v).

In an alternative embodiment, the topology graph processing module 740 includes:

the key neighbor node determining unit is used for determining a key neighbor node of any node in the topological graph;

and the input feature aggregation unit is used for aggregating the input features of the key neighbor nodes and performing iteration of preset layers through the graph neural network to obtain the features of any node.

In an optional implementation manner, the critical neighbor node determining unit is configured to:

and if any node is a block node, all commodity nodes are selected from the neighbor nodes of the node to serve as key neighbor nodes of any node.

In an optional implementation manner, the critical neighbor node determining unit is configured to:

and if any node is a commodity node, selecting a first preset number of user nodes, a second preset number of commodity nodes and all layout nodes from the neighbor nodes as key neighbor nodes of any node.

In an optional implementation manner, the critical neighbor node determining unit is configured to:

and if any node is a user node, selecting a third preset number of nodes from the neighbor nodes of the user node as the key neighbor nodes of any node.

In an optional implementation manner, the input feature aggregation unit is further configured to obtain the input features of the section nodes by:

and extracting characteristics of at least one of title texts, category information and poster images of the sections, and coding to obtain input characteristics of section nodes corresponding to the sections.

In an optional implementation manner, the input feature aggregation unit is further configured to obtain the input features of the commodity node by:

and extracting characteristics of at least one of title texts, category information, price information and commodity images of the commodities, and coding to obtain input characteristics of the section nodes corresponding to the commodities.

In an optional implementation manner, the input feature aggregation unit is further configured to obtain the input features of the user node by:

and extracting characteristics of at least one of the basic information and the behavior attribute of the user, and coding to obtain the input characteristics of the user node corresponding to the user.

In an optional implementation, the topology map processing module 740 is further configured to:

extracting the characteristics of each node in the topological graph by using a graph neural network to obtain a node characteristic table;

when the characteristics of the object to be processed are extracted, the target node corresponding to the object to be processed is determined in the topological graph, and the characteristics of the target node are obtained by searching in the node characteristic table.

In an optional implementation, the feature extraction apparatus 700 further includes a recommendation information generation module 750 configured to:

and if the object to be processed is the target user, after the characteristics of the target user are obtained, generating recommendation information which is specific to the target user and related to the layout according to the similarity between the characteristics of the target user and the characteristics of the layout nodes and/or the similarity between the characteristics of the target user and the characteristics of the commodity nodes.

In addition, other specific details of the embodiments of the present invention have been described in detail in the embodiments of the present invention of the above method, and are not described herein again.

Exemplary storage Medium

A storage medium of an exemplary embodiment of the present invention is explained with reference to fig. 8.

As shown in fig. 8, a program product 800 for implementing the above method according to an embodiment of the present invention is described, which may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on a device, such as a personal computer. However, the program product of the present invention is not limited in this regard and, in the present document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

A computer readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A readable signal medium may also be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RE, etc., or any suitable combination of the foregoing.

Program code for carrying out operations of the present invention may be written in any combination of one or more programming languages, including AN object oriented programming language such as Java, C + +, or the like, as well as conventional procedural programming languages, such as the "C" language, or similar programming languages.

Exemplary electronic device

An electronic device of an exemplary embodiment of the present invention is explained with reference to fig. 9.

The electronic device 900 shown in fig. 9 is only an example and should not bring any limitations to the function and scope of use of the embodiments of the present invention.

As shown in fig. 9, the electronic device 900 is embodied in the form of a general purpose computing device. Components of electronic device 900 may include, but are not limited to: at least one processing unit 910, at least one memory unit 920, a bus 930 connecting different system components (including the memory unit 920 and the processing unit 910), a display unit 940.

Where the storage unit stores program code, which may be executed by the processing unit 910, to cause the processing unit 910 to perform the steps according to various exemplary embodiments of the present invention described in the above section "exemplary methods" of the present specification. For example, the processing unit 910 may perform the method steps as shown in fig. 1, fig. 2, fig. 5, or fig. 6, and the like.

The storage unit 920 may include volatile memory units such as a random access memory unit (RAM)921 and/or a cache memory unit 922, and may further include a read only memory unit (ROM) 923.

Storage unit 920 may also include a program/utility 924 having a set (at least one) of program modules 925, such program modules 925 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may comprise an implementation of a network environment.

The bus 930 may include a data bus, an address bus, and a control bus.

The electronic device 900 may also communicate with one or more external devices 1000 (e.g., keyboard, pointing device, bluetooth device, etc.), which may be through an input/output (I/O) interface 950. The electronic device 900 further comprises a display unit 940 connected to the input/output (I/O) interface 950 for displaying. Also, the electronic device 900 may communicate with one or more networks (e.g., a local area network (FAN), a Wide Area Network (WAN) and/or a public network, such as the Internet) via the network adapter 960. As shown, the network adapter 960 communicates with the other modules of the electronic device 900 via the bus 930. It should be appreciated that although not shown, other hardware and/or software modules may be used in conjunction with the electronic device 900, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

It should be noted that although in the above detailed description several modules or sub-modules of the apparatus are mentioned, such division is merely exemplary and not mandatory. Indeed, the features and functionality of two or more of the units/modules described above may be embodied in one unit/module according to embodiments of the invention. Conversely, the features and functions of one unit/module described above may be further divided into embodiments by a plurality of units/modules.

Moreover, while the operations of the method of the invention are depicted in the drawings in a particular order, this does not require or imply that the operations must be performed in this particular order, or that all of the illustrated operations must be performed, to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step execution, and/or one step broken down into multiple step executions.

While the spirit and principles of the invention have been described with reference to several particular embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, nor is the division of aspects, which is for convenience only as the features in such aspects may not be combined to benefit. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims (10)

1. A method of feature extraction, comprising:

acquiring commodities in each edition block, wherein the edition block is a commodity set comprising at least one commodity;

generating a layout node corresponding to each layout, a commodity node corresponding to each commodity and a user node;

generating edges among the layout nodes, the commodity nodes and the user nodes based on the historical behaviors of the user on the layout and the commodity and the dependency relationship between the layout and the commodity so as to establish a topological graph;

and extracting the characteristics of any node in the topological graph by utilizing a graph neural network.

2. The method of claim 1, wherein the user node is generated by:

clustering users to obtain a plurality of user categories;

and respectively generating a corresponding user node for each user category.

3. The method of claim 2, wherein clustering users to obtain a plurality of user categories comprises:

clustering the users based on the basic information of the users to obtain a plurality of first user categories;

and clustering the users based on the behavior attributes of the users to obtain a plurality of second user categories.

4. The method of claim 3, wherein when a new user is generated, the method further comprises:

if the new user has historical data, determining the behavior attribute of the new user according to the historical data of the new user, and classifying the new user into a corresponding second user category according to the behavior attribute of the new user;

and if the new user does not have the historical data, classifying the new user into a corresponding first user category according to the basic information of the new user.

5. The method of claim 2, wherein generating edges between each of the layout nodes, commodity nodes, and user nodes to establish a topological graph based on historical behaviors of users on the layouts, the commodities, and the affiliations between the layouts and the commodities comprises:

and counting the historical interaction behaviors of the users in each user category and the sections in the historical data to determine edges between the user nodes and the section nodes and edge weights.

6. The method of claim 2, wherein generating edges between each of the layout nodes, commodity nodes, and user nodes to establish a topological graph based on historical behaviors of users on the layouts, the commodities, and the affiliations between the layouts and the commodities comprises:

and counting historical interaction behaviors of the users in each user category and the commodities in historical data to determine edges between the user nodes and the commodity nodes and edge weights.

7. The method of claim 1, wherein generating edges between each of the layout nodes, commodity nodes, and user nodes to establish a topological graph based on historical behaviors of users on the layouts, the commodities, and the affiliations between the layouts and the commodities comprises:

and counting the word frequency-inverse text frequency of each commodity in each block, and determining edges and edge weights between the block nodes and the commodity nodes according to the word frequency-inverse text frequency.

8. A feature extraction device characterized by comprising:

the system comprises an information acquisition module, a display module and a display module, wherein the information acquisition module is used for acquiring commodities in each edition block, and the edition block is a commodity set comprising at least one commodity;

the node generation module is used for generating layout nodes corresponding to the layouts, commodity nodes corresponding to the commodities and user nodes;

the edge generation module is used for generating edges among the layout nodes, the commodity nodes and the user nodes based on historical behaviors of the users on the layouts and the commodities and the dependency relationship between the layouts and the commodities so as to establish a topological graph;

and the topological graph processing module is used for extracting the characteristics of any node in the topological graph by using the graph neural network.

9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method of any one of claims 1 to 7.

10. An electronic device, comprising:

a processor; and

a memory for storing executable instructions of the processor;

wherein the processor is configured to perform the method of any of claims 1-7 via execution of the executable instructions.

Images