A method and system for identifying second-hand goods

Abstract

The application relates to a second-hand commodity identification method and system, wherein the method comprises the following steps: acquiring full-amount inventory commodity information in a second-hand platform; acquiring a vector representation of a commodity according to the commodity information to obtain a commodity vector; clustering based on the commodity vector to obtain a plurality of commodity clusters; and assigning a class identifier to the commodity clusters, and the commodities in the commodity clusters share the class identifier. The commodity identification system comprises a commodity information acquisition module, a commodity vector generation module, a clustering module and a class cluster identification module. The application determines the class identifier of commodities, and after a commodity is sold, commodities with the same class identifier as the sold commodity still exist, thereby effectively solving the problem of the single-product attribute of second-hand commodities, and the class identifier of the commodities can be used to add a recall strategy in the recall stage of a recommendation system, enrich the recall system and improve the model sorting capability.

Description

Technical Field

[0001] This invention relates to the field of Internet application technology, and in particular to a method and system for identifying second-hand goods. Background Technology

[0002] With the development of internet applications and the logistics industry, more and more merchants and consumers tend to complete transactions on e-commerce platforms (hereinafter referred to as e-commerce platforms). To facilitate the management of users and goods transacting on these platforms and to streamline calculations for services such as product recommendations, e-commerce platform systems typically assign IDs, or identity identifiers, to both products and users. E-commerce platforms can be categorized based on the status of the goods traded on them: ordinary e-commerce platforms offering new product transactions and secondhand e-commerce platforms offering used goods transactions. Secondhand e-commerce platforms typically sell unique items, meaning they are sold in single stock, and thus each item corresponds to a unique product ID. Ordinary e-commerce platforms, on the other hand, have multiple stock items, with one product ID corresponding to multiple items in the inventory. For secondhand e-commerce platforms, once a product is sold, the system no longer has inventory of that ID, but the product information for that ID will still be retained in the user logs. This is known as the unique item attribute of a product. For example, a user log might record that a user clicked on and viewed the product at a certain time and ultimately placed an order for it. The uniqueness of products leads to a series of problems for e-commerce platforms: First, the recommendation model used by the recommendation system learns from user logs during the learning phase. However, when some products in the user logs have been sold, the secondhand e-commerce platform no longer has products with the same product ID in its inventory. Therefore, the model's learning of these sold products in the user logs does not help in recalling existing products on the platform. Second, product ID features are important features for recommendation systems. However, due to the single-inventory nature of secondhand products, the recommendation system in this scenario may not be able to learn accurate ID features, and the model may become very large and difficult to apply. This is because products on the platform dynamically disappear from the system inventory due to sales and enter the system inventory due to additions. Therefore, the product IDs in the system inventory are constantly changing, requiring the recommendation model to continuously learn. The consequence is that the model becomes increasingly large and inefficient over time, until it can no longer meet the needs of the application. Summary of the Invention

[0003] To address the technical problems existing in the prior art, this invention provides a method and system for identifying second-hand goods, which solves the problem of the uniqueness of second-hand goods.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, a method for identifying secondhand goods is provided, comprising the following steps: obtaining full inventory information of goods in a secondhand platform; obtaining a vector representation of the goods based on the goods information to obtain a goods vector; performing clustering based on the goods vector to obtain multiple goods clusters; and assigning class identifiers to the multiple goods clusters respectively, wherein goods in the same goods cluster share the class identifier of the goods cluster.

[0005] Preferably, the method further includes: calculating the vector distance between two products based on product vectors; comparing the vector distance between the two products with a threshold; determining that the two products are associated in response to the vector distance between the two products being less than the threshold; and constructing a product community network by connecting the associated two products with the products as nodes, wherein the node connection weight is the vector distance or the node connection weight is set to 1.

[0006] Preferably, the method further includes: obtaining one or more first associated products for each product based on product text information; obtaining one or more second associated products for each product based on user behavior information; obtaining the intersection or union of the first and second associated products of a product, wherein the products in the intersection or union are the associated products of the product; and constructing a product community network by connecting products with their associated products, wherein the node connection weight is the text vector distance, the product pair weight value, or the sum of the two, or the node connection weight is set to 1.

[0007] Preferably, the step of clustering based on the product vector includes: performing clustering calculations at one or more levels on the product community network based on the Louvain algorithm model to obtain multiple product sub-communities, wherein each product sub-community is a product cluster.

[0008] To address the aforementioned technical problems, according to another aspect of the present invention, a second-hand goods identification system is also provided, comprising a goods information acquisition module, a goods vector generation module, a clustering module, and a cluster identification module. The goods information acquisition module is configured to acquire full inventory goods information from a second-hand platform. The goods vector generation module is connected to the goods information acquisition module and configured to acquire a vector representation of a goods based on the goods information to obtain a goods vector. The clustering module is connected to the goods vector generation module and configured to perform clustering based on the goods vector to obtain multiple goods clusters. The cluster identification module is connected to the clustering module and configured to assign a class identifier to each goods cluster, wherein goods within the goods cluster share the same class identifier.

[0009] This invention determines the relationships between products through product information, such as text information and / or user behavior information, and groups similar products into the same cluster based on these relationships, giving products in the same cluster the same class identifier. When a product is sold, other products with the same class identifier still exist, thus effectively solving the problem of the uniqueness of secondhand goods. Furthermore, product class identifiers can be used to add recall strategies during the recall phase of the recommendation system, enriching the recall system. Moreover, class identifiers can be used as product features during the ranking phase of the recommendation system, thereby effectively improving the model's ranking capabilities. Attached Figure Description

[0010] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0011] Figure 1 This is a flowchart of a method for identifying second-hand goods according to an embodiment of the present invention;

[0012] Figure 2 This is a flowchart of a method for obtaining product vectors based on user behavior information according to an embodiment of the present invention;

[0013] Figure 3 This is a schematic diagram of a user clicking on a product sequence according to an embodiment of the present invention;

[0014] Figure 4 This is a schematic diagram of an undirected graph constructed based on product pairs according to an embodiment of the present invention;

[0015] Figure 5 This is a flowchart of a method for constructing a second-hand goods community network according to an embodiment of the present invention;

[0016] Figure 6 This is a schematic diagram of a commodity community network according to an embodiment of the present invention;

[0017] Figure 7 The diagram shows the clustering results of the first round of hierarchical clustering of the product community network.

[0018] Figure 8 It is based on Figure 7 The diagram shows a product community network after merging the clustering results.

[0019] Figure 9 This is a flowchart of a method for identifying second-hand goods according to another embodiment of the present invention;

[0020] Figure 10 This is a block diagram illustrating the principle of a product identification system according to an embodiment of the present invention;

[0021] Figure 11This is a block diagram illustrating the principle of a product information acquisition module according to an embodiment of the present invention.

[0022] Figure 12 This is a schematic diagram of a product identification system according to another embodiment of the present invention; and

[0023] Figure 13 This is a schematic diagram of a product identification system provided according to another embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0026] For goods sold on secondhand e-commerce platforms, each item typically has a unique identifier. Because secondhand goods are unique, once a product is sold, the platform's inventory no longer contains a product with that identifier. This invention refers to this attribute as the "unique item attribute." The secondhand goods identification method and system provided by this invention are used to solve the problem of the unique item attribute of secondhand goods.

[0027] Figure 1 This is a flowchart of a method for identifying secondhand goods according to an embodiment of the present invention. The method includes the following steps:

[0028] Step S1a: Obtain information on all inventory items on the secondhand platform. In one embodiment, the item information includes text information such as item titles and descriptions. In another embodiment, the obtained item information includes not only text information but also user behavior information, such as whether the item was clicked or favorited by the same user. This information can be obtained from user logs.

[0029] Step S2a: Obtain the vector representation of the product based on the product information to obtain the product vector. This invention can obtain product vectors based on different product information and different methods.

[0030] In one embodiment, a text vector for a product is obtained based on the product's text information. For example, the text information of each product is input into a pre-trained word2vec model to obtain word vectors in the text information. Then, a pooling operation is performed on the word vectors to obtain a text vector, which is referred to here as the first product vector to distinguish it from product vectors obtained in other ways.

[0031] In another embodiment, a second product vector can be generated based on user behavior information. For example... Figure 2 The diagram shown is a flowchart of a method for obtaining product vectors based on user behavior information according to an embodiment of the present invention. Details are as follows:

[0032] Step S200a: Obtain the sequence of items clicked by the user from user behavior information. For example, read data from a time window in each user's user log, extract the items clicked by the user, and construct a sequence of items clicked by the user in chronological order. Figure 3 The image shows a sequence of products clicked by a user: {I1, I2, I3, I4}.

[0033] Step S201a: Construct multiple co-occurring product pairs based on the product sequence. For example, pair products in the sequence according to chronological order. (As described above...) Figure 3 The sequence of products clicked by the user constitutes product pairs (I1, I2), (I2, I3), and (I3, I4).

[0034] Step S202a: Count the co-occurrence frequency or number of times product pairs appear, and construct an undirected graph based on the co-occurrence frequency or number of times, such as... Figure 4 As shown, the nodes in the undirected graph represent products, and the numbers on the edges connecting the nodes represent the number of times two nodes co-occur after standardization to between 0 and 1, also known as weights.

[0035] Step S203a: Use the Node2vec algorithm to obtain the vectorized representation of each item in the undirected graph. For example, first, use a random walk algorithm to sample the nodes in the undirected graph, and obtain a node sequence of a predetermined length based on each node. Each node in the sequence is a word, each node sequence is a document, and multiple node sequences correspond to a document set. Perform the word2vec algorithm on the document set to obtain the embedding vector representation of each word, that is, each node (item), which is the second item vector.

[0036] In another embodiment, the third product vector can be obtained based on the product text information and the user behavior information using a GNN (Graph Neural Networks) algorithm or a knowledge graph. When using the GNN algorithm, a user-product bipartite graph is first obtained using the product and user information involved in the user behavior information. Based on the user-product bipartite graph, the vector representation of each node in the graph is obtained using the GNN model, thereby obtaining the third product vector.

[0037] When using a knowledge graph approach, if the secondhand e-commerce platform stores a knowledge graph with products as nodes, the Trans series models are used to obtain the vector of each node in the knowledge graph, which is another type of third-party product vector. If there is no knowledge graph currently available, a knowledge graph can be generated using the textual information of the products and / or user behavior information, and then the Trans series models can be used to obtain the vector of each node in the knowledge graph.

[0038] Step S3a involves using a clustering algorithm to cluster the product vectors of all products obtained previously, resulting in multiple product clusters. The clustering algorithm can be, for example, K-means, EM, DBSCAN, hierarchical clustering, or graph-based community detection algorithms such as the Louvain algorithm. After clustering all products using any of the above algorithms, multiple product clusters are obtained, with each cluster containing one or more products.

[0039] Step S4a: Assign a corresponding class identifier to each product cluster. Products within a product cluster share the class identifier of that product cluster. For example, assign an ID number to each of multiple product clusters, and the products within that product cluster share the ID number.

[0040] The solution provided in this embodiment ensures that products within the same product cluster on a second-hand e-commerce platform have the same class ID. Even after a product is sold, other products with the same class ID still exist. Therefore, the recommendation model of the second-hand e-commerce platform avoids the problem of products with the same ID not existing after a product is sold, whether in the learning, product recall, or ranking phases. This improves the accuracy of product recall. Furthermore, the important feature of product class ID can be used in the ranking phase, simplifying the ranking algorithm. Moreover, when multiple products have the same class ID, the model can determine that the similarity between products with the same class ID is greater than that between products with different class IDs.

[0041] One embodiment of the present invention uses the Louvain algorithm model to cluster products on a second-hand e-commerce platform. To use the Louvain algorithm model for clustering, a product community network needs to be established. Figure 5This is a flowchart of a method for building a second-hand goods community network according to an embodiment of the present invention. The method includes the following steps:

[0042] Step S300a: Select one product as the target product.

[0043] Step S301a: Take another product as the second product and use it to perform vector calculation with the target product.

[0044] Step S302a: Calculate the vector distance between the target product and the second product. The product vectors used can be vectors obtained by any of the aforementioned methods. For example, calculating the Euclidean distance or cosine similarity between the two product vectors.

[0045] In step S303a, it is determined whether the obtained product vector distance is less than a threshold. If it is less, it indicates that the target product and the second product are similar and have a high correlation. In step S304a, the second product is confirmed as a related product of the target product, and a connection is established between the two as nodes. If the product vector distance is greater than or equal to the threshold, it indicates that the target product and the second product are not similar and have a low correlation. In step S305a, it is confirmed that the second product is not a related product of the target product.

[0046] Step S306a: Determine if there are any remaining products that can be used as the second product. If yes, return to step S301a. If not, in step S307a, determine if there are any remaining products that can be used as the target product. If yes, return to step S300a. If not, in step S308a, confirm that the product community network has been successfully built and end the construction process. Here, based on the distance between products in this embodiment, a [data structure is obtained as follows]... Figure 6 The diagram illustrates a product community network. In this network, nodes represent products, and edges between nodes represent the vector distance between products. In another embodiment, to simplify the algorithm, two products are considered related when their vector distance is less than a threshold, and the weight of the edge connecting the nodes of the two related products is uniformly set to 1.

[0047] This embodiment calculates the product community network based on the Louvain algorithm to obtain multiple product sub-communities, where each product sub-community is a product cluster.

[0048] by Figure 6 The product community network shown is an example. Assume the initial product community network is as follows: Figure 6The diagram shows a total of 13 nodes and 23 edges. Initially, each node is treated as a separate class. In the first round of computation, any one of these nodes is chosen as the current computation node and merged with every node in the current community network into a single class. The modularity gain ΔQ before and after the merge is then calculated. The modularity can be calculated using the following formula 1-1:

[0049]

[0050] Where m represents the sum of the weights among nodes in the entire commodity community network; A i,j As the weight of the edge connecting node i and node j, in this embodiment, A i,j The vector distance between two items is either 1 or 0; when nodes i and j are not connected, its value is 0. i and k j Let C represent the sum of the weights of the edges connected to nodes i and j, respectively; i and C j These refer to the community identifiers of the communities where nodes i and j reside, respectively; δ(C i C j ) is a function that has a value of 1 when node i and node j are in the same community, and a value of 0 otherwise.

[0051] The modularity gain ΔQ can be calculated using the following formula 1-2:

[0052]

[0053] Where, k i,in Σtot represents the sum of the edge weights of nodes i and other nodes within the community; Σtot represents the sum of the edge weights of nodes connected to nodes within the community.

[0054] The above calculations yield multiple modularity gains. The currently calculated node and the node with the largest modularity gain greater than 0 are then merged into a new community.

[0055] After several iterations following the above process, the following is obtained: Figure 7 The first-level clustering results are shown below. The clustering results at the current level are then merged, resulting in the following: Figure 8 As shown, the merged three nodes represent three... Figure 7The three clusters are labeled C1, C2, and C3. The edges between two clusters represent the sum of the weights between nodes in the two communities. At this point, the original 13 communities have been clustered into 3 clusters. The clustering calculation can stop here if needed, or the above process can be repeated to obtain the next level of clustering results. As can be seen from the above process, this embodiment uses the Louvain algorithm to flexibly obtain clustering results of different degrees. In the actual scenario of second-hand goods, the granularity of the clustered clusters needs to be finer than the granularity of the product classification. Therefore, the number of products in the clustered clusters can be compared with the number of products contained in the corresponding category. If the former is less than or much less than the latter, clustering stops. In specific implementation, a difference threshold can be set. When the difference between the two is greater than or equal to the set difference threshold, clustering stops. In another embodiment, the number of clusters can also be used as a hyperparameter, and a suitable value can be found using methods such as grid search.

[0056] This embodiment uses the Louvain algorithm based on community network graphs to cluster products and builds community network graphs according to the similarity of products. The clustering granularity is controllable and efficient.

[0057] Figure 9 This is a flowchart of a method for identifying secondhand goods according to another embodiment of the present invention, the method comprising the following steps:

[0058] Step S1c: Obtain text information of all inventory items on the second-hand platform, such as titles and product descriptions.

[0059] Step S2c involves obtaining the text vector of the product based on its text information. For example, the product's text information is input into the word2vec model to obtain word vectors from the text information. Then, a pooling operation is performed on the word vectors to obtain the text vector, which is used as the first product vector for that product.

[0060] Step S3c: Calculate the vector distance between products based on the first product vector, such as calculating the Euclidean distance between two text vectors.

[0061] Step S4c: Based on the text vector distance between the products, obtain one or more first associated products for each product. For example, products with a vector distance less than a vector threshold are taken as first associated products, or products are sorted in ascending order of vector distance and the products ranked first are taken as first associated products.

[0062] Step S5c: Obtain user behavior information and extract the products clicked by the user to form a product sequence.

[0063] Step S6c: Obtain multiple co-occurring product pairs from the product sequence.

[0064] Step S7c: Calculate the co-occurrence frequency of the product pair and determine the weight of the product pair based on the co-occurrence frequency.

[0065] Step S8c involves determining the second associated product for each product based on the weights of the product pairs. For example, multiple product pairs containing the same product (referred to as the first product) are sorted in descending order of weight, and a certain number of product pairs are obtained. The second product in each of these product pairs is then designated as the second associated product of the first product. In another embodiment, when the second product vector is calculated based on user behavior, the second associated product can be determined based on the distance between the second product vectors of the products and a threshold.

[0066] Step S9c: Obtain the intersection or union of the first and second associated products of a product, where the products in the intersection or union are the associated products of the product. In one embodiment, the number of associated products can be increased by taking the union of the first and second associated products of a product.

[0067] Step S10c involves constructing a product community network by connecting products to their associated products, using products as nodes. The weights of the edges between nodes in this product community network are either text vector distance, the weight of a product pair, or the sum of both. To facilitate the subsequent clustering calculations, the weights of the edges between nodes can be simplified; for example, the text vector distance can be set to 1, and the weight of a product pair can be set to 1. Therefore, the weight of the edge between two nodes can be either 1 or 2. In another embodiment, the text vector distance and the weight of the product pair can be standardized to values ​​between 0 and 1, which also facilitates the subsequent clustering calculations.

[0068] Step S11c: The product community network is calculated based on the Louvain algorithm to obtain multiple product sub-communities, where each product sub-community represents a product cluster. The calculation method is as described in the previous embodiment and will not be repeated here.

[0069] Step S12c: Assign a class identifier to each product cluster, and the products in the product cluster share the same class identifier.

[0070] Since most of the products on second-hand e-commerce platforms are unique items, in order to increase the quantity of products in the same category, in this embodiment, when establishing the product community network, the associated products obtained according to text and user behavior are combined and their union is taken, which can effectively increase the quantity of products.

[0071] This invention also provides a product identification system, such as Figure 10As shown, the product identification system includes a product information acquisition module 1, a product vector generation module 2, a clustering module 3, and a cluster identification module 4. The product information acquisition module 1 is used to acquire information on all inventory products in the second-hand platform, specifically as follows: Figure 11 As shown, the system includes a text information acquisition unit 11 and a user behavior information acquisition unit 12. The text information acquisition unit 11 obtains the title and text description of the products from the second-hand platform's product database to synthesize the text information. The user behavior information acquisition unit 12 obtains product information clicked by the user from the user log according to a time window. In one embodiment, this is obtained as follows: Figure 3 The sequence of products clicked by the user is shown. The product vector generation module 2 is connected to the product information acquisition module 1 and is configured to obtain the vector representation of the products based on the product information to obtain product vectors. Specifically, this invention obtains the associations between products based on the product information, uses each product as a node, establishes an undirected graph of products based on the associations, and then uses the Node2vec algorithm to obtain the value of each product. The clustering module 3 is connected to the product vector generation module 2 and is configured to cluster the product vectors using algorithms such as K-means, EM, DBSCAN, hierarchical clustering, or the graph-based Louvain algorithm to obtain multiple product clusters. The cluster identification module 4 is connected to the clustering module 3 and is configured to assign a class identifier to each product cluster; products in each product cluster share the same class identifier.

[0072] Figure 12 This is a schematic diagram of a product identification system according to another embodiment of the present invention. In this embodiment, the product identification system includes a product information acquisition module 1a, a product vector generation module 2a, a clustering module 3a, and a cluster identification module 4a. This embodiment uses the Louvain algorithm for product clustering. Correspondingly, the system in this embodiment also includes a product community network construction module 5a for constructing a product community network.

[0073] The product vector generation module 2a comprises a first product vector generation unit 21a, a product pair construction unit 22a, an undirected graph construction unit 23a, and a node vector generation unit 24a, forming a second product vector generation unit and a third product vector generation unit 25a. The first product vector generation unit 21a uses a natural language model to obtain text vectors based on the product's text information. For example, the product's text information, consisting of a title and description, is input into a trained word2vec model to obtain word vectors. Pooling is then applied to all word vectors to obtain the product's text vector. The product pair construction unit 22a is connected to the user behavior information acquisition unit in the product information acquisition module 1a. It combines the products clicked by the user in chronological order from the user information into co-occurring product pairs and calculates the co-occurrence frequency or number of times of each pair, using this frequency or number as the weight of the product pair. The undirected graph construction unit 23a is connected to the product pair construction unit 22a, and constructs an undirected graph according to co-occurring product pairs, wherein the nodes of the undirected graph are products, and the association between nodes is the weight of the two products, such as... Figure 4 As shown. The node vector generation unit 24a is connected to the undirected graph construction unit 23a, and obtains the vectorized representation of each node in the undirected graph based on the Node2vec algorithm, thus obtaining the second product vector. The third product vector generation unit 25a receives the text information and user behavior information from the product information acquisition module 1a, and obtains the third product vector using the GNN algorithm or knowledge graph.

[0074] The product community network construction module 5a is connected to the product vector generation module 2a, and specifically includes a vector distance calculation unit 51a, a first associated product acquisition unit 52a, and a first node association unit 53a. The vector distance calculation unit 51a is connected to the node vector generation unit 24a and is used to calculate the vector distance between two products. When calculating the vector distance between products, any type of product vector can be used, such as a first product vector obtained from text information, a second product vector obtained from user behavior information, or a third product vector obtained from both text information and user behavior information. The first associated product acquisition unit 52a is connected to the vector distance calculation unit 51a and determines the associated products of a target product based on the vector distance. For example, it compares the first vector distance between the target product and a comparison product with a threshold. If the first vector distance is greater than or equal to the threshold, the comparison product is not an associated product of the target product because the vector distance is too large, indicating a significant difference between them. If the vector distance is less than the threshold, the comparison product is an associated product of the target product. Alternatively, the calculated vector distances can be sorted in ascending order, and a preset number of products at the top of the sorted list can be used as associated products of the target product. The first node association unit 53a is connected to the first associated product acquisition unit 52a. Based on the aforementioned association relationships between products, products are used as nodes, and the association relationships between products are used as connecting edges to form a product community network. The weight of the connecting edge is the vector distance between the product and its associated products or is set to 1.

[0075] Clustering module 3a is connected to the product community network construction module 5a. It performs calculations according to the Louvain algorithm. For example, in the first round of hierarchical clustering, each product is treated as a community, and one of these products is selected as the target product. This target product is then grouped into a community with other products in the network. The modularity before and after the grouping is calculated using Formula 1-1, and the difference between the two, i.e., the modularity gain ΔQ, is calculated. Alternatively, the modularity gain ΔQ can be directly calculated using Formula 1-2.

[0076]

[0077]

[0078] Where m represents the sum of the weights among nodes in the entire commodity community network; A i,j k is the weight of the edge connecting node i and node j; its value is 0 when nodes i and j are not connected. i and k j Let C represent the sum of the weights of the edges connected to nodes i and j, respectively; i and C jThese refer to the community identifiers of the communities where nodes i and j reside, respectively; δ(C i C j Let be a function, which represents a value of 1 when node i and node j are in the same community, and a value of 0 otherwise; k i,in Σtot represents the sum of edge weights between node i and other nodes within the community; Σtot represents the sum of edge weights connecting to nodes within the community. In this embodiment, the weight of the connecting edge is the vector distance between the two items or is set to 1.

[0079] After calculating the modularity gain ΔQ for all products in the current product community network, the product with the largest modularity gain greater than 0 is identified, and the target product and this product are merged into a new community.

[0080] After several iterations following the above process, several product clusters that met the requirements were finally obtained.

[0081] The cluster identification module 4a is connected to the clustering module 3a. After obtaining multiple product clusters that meet the requirements, a class identifier is set for each product cluster. The products in the product cluster share the class identifier.

[0082] Figure 13 This is a schematic diagram of a product identification system according to another embodiment of the present invention. In this embodiment, the product identification system includes a product information acquisition module 1b, a product vector generation module 2b, a clustering module 3b, and a cluster identification module 4b. Since this embodiment uses the Louvain algorithm for product clustering, it also includes a product community network construction module 5b, which is used to construct a product community network.

[0083] The product vector generation module 2b includes a first product vector generation unit 21b and a product pair construction unit 22b. The first product vector generation unit 21b and the product pair construction unit 22b are connected to... Figure 12 The first commodity vector generation unit 21a and the commodity pair construction unit 22a have the same function, which will not be described again here.

[0084] The product community network construction module 5b includes a second associated product acquisition unit 51b, a third associated product acquisition unit 52b, a fourth associated product acquisition unit 53b, and a second node association unit 54b. The second associated product acquisition unit 51b is connected to the first product vector generation unit 21b in the product vector generation module 2b, and is used to calculate the text vector distance between two products and determine the second associated product of a product based on the text vector distance. The third associated product acquisition unit 52b is connected to the product pair construction unit 22b and determines the third associated product of each product according to the product pair weights. The fourth associated product acquisition unit 53b is connected to the second associated product acquisition unit 51b and the third associated product acquisition unit 52b, and is configured to obtain the intersection or union of the second and third associated products of a product. The products in the intersection or union are the final associated products of the product. The second node association unit 54b is connected to the fourth associated product acquisition unit 53b. It uses products as nodes and product associations as edges to connect products, forming a product community network. The node edge weights are either text vector distance, product pair weights, or the sum of both, or are set to 1. The clustering module 3b calculates the product community network according to the Louvain algorithm to obtain multiple different product clusters. The node edge weights used in the calculation are the aforementioned text vector distance, product pair weights, or the sum of both.

[0085] This invention determines the association between products by using product text information and user behavior information, and groups similar products into the same cluster based on their association, giving products in the same cluster the same class identifier. When a product is sold, products with the same class identifier still exist, thus effectively solving the problem of the uniqueness of second-hand products. At the same time, the class identifier of the product can be used to add recall strategies in the recall stage of the recommendation system, enriching the recall system. In addition, the class identifier can be used as a product feature in the ranking stage of the recommendation system to improve the model's ranking ability.

[0086] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for identifying secondhand goods, comprising: Obtain information on all inventory items on secondhand platforms; The product vector is obtained by retrieving the vector representation of the product based on the product information. Clustering is performed based on the product vectors to obtain multiple product clusters; as well as Each of the multiple product clusters is assigned a class identifier, and products within the same product cluster share the class identifier of that product cluster; The step of clustering based on the product vector to obtain multiple product clusters includes: determining associated products based on the product vector; Using the total inventory of goods as nodes, related goods are connected to build a product community network; and The product community network is subjected to clustering calculations at one or more levels based on the Louvain algorithm model until the granularity of the clustered classes is finer than the granularity of the product classification, so as to obtain multiple product sub-communities, each of which is a product cluster.

2. The method according to claim 1, wherein the product information includes product text information and / or user behavior information.

3. The method according to claim 2, wherein the step of obtaining a vector representation of a product based on product information to obtain a product vector includes: Based on the product text information, the first product vector of the product is obtained using a natural language model; or Extract co-occurring product pairs from user behavior information; Calculate the co-occurrence frequency of the product pairs, and determine the weight of the product pairs based on the co-occurrence frequency; Construct an undirected graph based on product pairs and their weights; The Node2vec algorithm is used to obtain the second product vector corresponding to the product of the node based on the undirected graph; or Based on the product text information and the user behavior information, a third product vector is obtained using a GNN algorithm or knowledge graph. Wherein, the first product vector, the second product vector, or a combination thereof are used as product vectors, or the third product vector is used as a product vector.

4. The method of claim 1, further comprising: Calculate the vector distance between two products based on their product vectors; Compare the vector distance between two products with the threshold value; as well as In response to the vector distance between two products being less than a threshold, the two products are determined to be associated; wherein the node connection weight in the product community network is the vector distance or the node connection weight is set to 1.

5. The method of claim 1, further comprising: For each product, obtain one or more first-related products based on the similarity of product text information; For each product, obtain one or more second related products based on the similarity of user behavior information; as well as Obtain the intersection or union of the first and second associated products of a product, wherein the products in the intersection or union are the associated products of the product; wherein, the node connection weight in the product community network is the text information similarity, user behavior information similarity, or the sum of the two, or the node connection weight is set to 1.

6. The method according to claim 5, wherein, Further includes: Calculate the first vector distance between products based on the first product vector of the products; as well as Based on the first vector distance between products and a threshold, obtain one or more first associated products for each product.

7. The method according to claim 5, wherein, Further includes: Based on the weights and thresholds of product pairs, determine one or more second associated products for each product; or Calculate the second vector distance between products based on the second product vector of the product; as well as Based on the second vector distance between products and a threshold, obtain one or more second associated products for each product.

8. A second-hand goods identification system, comprising: The product information acquisition module, after configuration, can obtain information on all inventory products in the second-hand platform; A product vector generation module, which is connected to the product information acquisition module, is configured to obtain a vector representation of the product based on the product information to obtain a product vector; A clustering module, which is connected to the product vector generation module, is configured to perform clustering based on the product vectors to obtain multiple product clusters; as well as A cluster identifier module, which is connected to the clustering module, is configured to assign a class identifier to the product cluster, and the products in the product cluster share the class identifier; The clustering module is configured to determine associated products based on product vectors; connect associated products using the full inventory of products as nodes to construct a product community network; and perform clustering calculations on the product community network at one or more levels based on the Louvain algorithm model until the cluster granularity is finer than the product classification granularity, so as to obtain multiple product sub-communities, where each product sub-community is a product cluster.

9. The system according to claim 8, wherein, The product information acquisition module includes one or more of the following units: The text information acquisition unit is configured to retrieve the title and text description of a product from the second-hand platform's product database and synthesize them into this text information; and The user behavior information acquisition unit is configured to obtain information about products clicked by users from user logs according to time windows.

10. The system according to claim 9, wherein the product vector generation module comprises any one of the following units: The first product vector generation unit is configured to generate a text vector based on the text information of the product using a natural language model, and the text vector serves as the first product vector. The second product vector generation unit is configured to use the Node2vec algorithm to obtain a vectorized representation of each node in an undirected graph constructed based on co-occurring product pairs extracted from the user behavior information, as the second product vector; and The third product vector generation unit is configured to obtain a third product vector based on the product text information and the user behavior information using a GNN algorithm or knowledge graph. in, The first product vector, the second product vector, or a combination of the two are used as the product vector, or the third product vector is used as the product vector.

11. The system according to claim 10, further comprising a first product community network construction module connected to the product vector generation module, configured to include: A vector distance calculation unit is configured to calculate the vector distance between two products based on the product vector; The first associated product acquisition unit is connected to the vector distance calculation unit and is configured to determine the first associated product of a target product based on the vector distance. The first associated product is a product whose vector distance from the target product is less than or equal to a threshold, or a preset number of products that are sorted first by their vector distance from the target product in ascending order. as well as The first node association unit is connected to the first associated product acquisition unit. It is configured to use products as nodes and the association relationships of products as edges to form a product community network. The weight of the edges is either vector distance or the weight of the edges is set to 1.

12. The system according to claim 10, further comprising a second product community network construction module connected to the product vector generation module, and further comprising: The second associated product acquisition unit is connected to the first product vector generation unit and is configured to determine one or more second associated products for each product based on the first vector distance between products calculated from the first product vector and a threshold. The third associated product acquisition unit is connected to the second product vector generation unit and is configured to determine the third associated product of each product based on the weight of the product pair; or to determine the third associated product of each product based on the second vector distance calculated from the second product vector and a threshold between products. A fourth associated product acquisition unit, connected to the second and third associated product acquisition units, is configured to acquire the intersection or union of the second and third associated products of a given product, wherein the products in the intersection or union are the final associated products of the given product; and The second node association unit is connected to the fourth associated product acquisition unit. It is configured to use products as nodes and the proximity relationship of products as edges to form a product community network. The weight of the node edge is the text vector distance or the weight value of the product pair or the sum of the two.

Images