Model training method, abnormal commodity identification method, electronic device and storage medium

Abstract

The embodiment of the present application provides a model training method, an abnormal commodity identification method, an electronic device and a storage medium. In the embodiment of the present application, a graph neural network model is trained based on a risk graph reflecting the risk association relationship between sample commodities and an attribute graph reflecting the attribute association relationship between sample commodities. Since the risk association relationship between sample commodities and the attribute association relationship between sample commodities are fully considered during the model training process, the target graph neural network model obtained by such training can effectively measure the risk correlation between commodities, and thus accurately identify abnormal commodities.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a model training method, an abnormal commodity identification method, an electronic device and a storage medium. Background Art

[0002] With the continuous development of technologies such as the Internet and big data, more and more e-commerce platforms that provide transaction services have emerged, such as business to business (B2B) e-commerce platforms, business to consumer (B2C) e-commerce platforms, or consumer to consumer (C2C) e-commerce platforms.

[0003] In the daily operation of major e-commerce platforms, it is necessary to accurately identify and delete various abnormal products on the e-commerce platforms. At present, traditional machine learning or deep learning methods are usually used to identify whether a product is an abnormal product based on the attribute characteristics of the product. However, the attribute characteristics of abnormal products can be easily tampered with to be very similar to legitimate products, making it difficult to accurately identify abnormal products. Summary of the invention

[0004] Multiple aspects of the present application provide a model training method, an abnormal product identification method, an electronic device and a storage medium for accurately identifying abnormal products.

[0005] An embodiment of the present application provides a model training method, including: obtaining multiple sample commodities, the multiple sample commodities including at least sample labeled commodities with real labels and sample abnormal commodities with abnormal labels, the real labels being abnormal labels or normal labels; constructing a risk map reflecting the risk association relationship between the sample commodities based on the relevant information of the multiple sample commodities, and constructing an attribute map reflecting the attribute association relationship between the sample commodities based on the attribute characteristics of the multiple sample commodities; training an initial graph neural network model based on the risk map and the attribute map to obtain a target graph neural network model that can be used for abnormal commodity identification.

[0006] An embodiment of the present application also provides a method for identifying abnormal commodities, including: obtaining multiple commodities, the multiple commodities including at least commodities to be identified and abnormal commodities with abnormal labels; constructing a target risk map reflecting the risk correlation relationship between commodities based on the risk information of the multiple commodities, and constructing a target attribute map reflecting the attribute correlation relationship between commodities based on the attribute characteristics of the multiple commodities; inputting the target risk map and the target attribute map into a target graph neural network model to identify abnormal commodities to obtain an identification result of the commodities to be identified; the target graph neural network model is trained according to the above-mentioned model training method.

[0007] An embodiment of the present application also provides an electronic device, comprising: a memory and a processor; the memory is used to store a computer program; the processor is coupled to the memory, and is used to execute the computer program to execute the steps in the model training method, or to execute the steps in the abnormal product identification method.

[0008] An embodiment of the present application also provides a computer storage medium storing a computer program. When the computer program is executed by a processor, the processor is enabled to implement the steps in the model training method, or to implement the steps in the abnormal product identification method.

[0009] In an embodiment of the present application, a graph neural network model is trained based on a risk graph reflecting the risk correlation relationship between sample commodities and an attribute graph reflecting the attribute correlation relationship between sample commodities. Since the risk correlation relationship between sample commodities and the attribute correlation relationship between sample commodities are fully considered during the model training process, the target graph neural network model obtained by such training can effectively measure the risk correlation between commodities, and thus accurately identify abnormal commodities. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0011] Figure 1 A schematic diagram of an application scenario applicable to an abnormal commodity identification method provided in an embodiment of the present application;

[0012] Figure 2 A flowchart of a model training method provided in an embodiment of the present application;

[0013] Figure 3 A schematic diagram of an application scenario applicable to a model training method provided in an embodiment of the present application;

[0014] Figure 4 A flowchart of a method for identifying abnormal commodities provided in an embodiment of the present application;

[0015] Figure 5 A schematic diagram of the structure of a model training device provided in an embodiment of the present application;

[0016] Figure 6 A schematic diagram of the structure of an abnormal commodity identification device provided in an embodiment of the present application;

[0017] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0019] In addition, in order to clearly describe the technical solutions of the embodiments of the present application, in the embodiments of the present application, "first", "second", etc. are used to distinguish the same items or similar items with substantially the same functions and effects. Those skilled in the art can understand that "first", "second", etc. do not limit the quantity and execution order, and "first", "second", etc. do not necessarily limit them to be different.

[0020] First, the terms involved in the embodiments of the present application are introduced;

[0021] Graph: A graph consisting of several nodes and edges connecting two nodes, used to depict the relationship between different nodes. In the embodiment of the present application, a risk graph is a graph that uses commodities as nodes to reflect the risk relationship between commodities; an attribute graph is a graph that uses commodities as nodes to reflect the attribute relationship between commodities.

[0022] Homogeneous graph: A graph with a single type of nodes and a single type of edges.

[0023] Heterogeneous graph: A graph that includes multiple types of nodes or multiple types of edges.

[0024] Graph Neural Networks (GNN) model: a neural network that acts directly on the graph structure. The graph neural network model has the following characteristics: 1. Ignore the input order of nodes; 2. During the calculation process, the representation of the node is affected by its surrounding neighboring nodes, while the connection of the graph itself remains unchanged; 3. The representation of the graph structure enables graph-based reasoning. Among them, the graph neural network model is divided into five categories, namely: Graph Convolution Networks (GCN) model, Graph Attention Networks (GAN) model, Graph Autoencoders (Graph Autoencoders) model, Graph Generative Networks (GGN) model and Graph Spatial-temporal Networks (Graph Spatial-temporal Networks) model.

[0025] At present, traditional machine learning or deep learning methods are usually used to identify whether a product is an abnormal product based on the attribute characteristics of the product. However, the attribute characteristics of abnormal products can be easily tampered with to be very similar to legitimate products, making it difficult to accurately identify abnormal products. To this end, an embodiment of the present application provides a model training method, an abnormal product identification method, an electronic device, and a storage medium. In an embodiment of the present application, a graph neural network model is trained based on a risk graph reflecting the risk association relationship between sample products and an attribute graph reflecting the attribute association relationship between sample products. Since the risk association relationship between sample products and the attribute association relationship between sample products are fully considered during the model training process, the target graph neural network model trained in this way can effectively measure the risk correlation between products, thereby accurately identifying abnormal products.

[0026] Figure 1 This is a schematic diagram of an application scenario applicable to an abnormal commodity identification method provided in an embodiment of the present application. Figure 1 , the terminal device 10 or the terminal device 20 can interact with the server 30 of the e-commerce platform through a wired network or a wireless network. For example, a wired network may include coaxial cables, twisted pairs, and optical fibers, and a wireless network may be a 2G network, a 3G network, a 4G network, or a 5G network, a Wireless Fidelity (WIFI) network, or the like. The present application does not limit the specific type or form of interaction, as long as it can realize the function of interaction between the terminal device 10 or the terminal device 20 and the server 30. Optionally, the terminal device 10 or the terminal device 20 includes, for example, but is not limited to, a mobile phone, a tablet computer, a laptop computer, a wearable device, and a vehicle-mounted device. The server 30 includes, for example, but is not limited to, a single server or a distributed server cluster consisting of multiple servers. It should be understood that, Figure 1 The number of terminal devices 10, terminal devices 20 and servers 30 in the figure is only for illustration. In practical applications, any number of terminal devices 10, terminal devices 20 and servers 30 can be deployed according to actual needs.

[0027] In practical applications, such as Figure 1 As shown in ①, some merchants may send information about abnormal products such as wild animals and plants, counterfeit drugs, theft tools and counterfeit products to the server 30 of the e-commerce platform through the merchant-side App (application) provided by the e-commerce platform running in the terminal device 10. Figure 1 As shown in ②, the server 30 of the e-commerce platform uses the target graph neural network model to identify abnormal products from the product information of the merchant to accurately identify the information of abnormal products. Figure 1As shown in ③, after the server 30 identifies the information of abnormal commodities, it is prohibited to push the information of abnormal commodities to the user. In this way, after the user opens the shopping app provided by the e-commerce platform in the terminal device 20, the application interface of the shopping app displays the graphic information of multiple normal commodities, but does not display the graphic information of abnormal commodities, which reduces the probability of users purchasing abnormal commodities and ensures the safe operation of the e-commerce platform.

[0028] The technical solutions provided by various embodiments of the present application are described in detail below in conjunction with the accompanying drawings.

[0029] Figure 2 A flowchart of a model training method provided in an embodiment of the present application. Figure 2 , the method may include the following steps:

[0030] 201. Acquire multiple sample commodities, where the multiple sample commodities at least include sample labeled commodities with real labels and sample abnormal commodities with abnormal labels, where the real labels are abnormal labels or normal labels.

[0031] 202. Construct a risk graph reflecting the risk correlation relationship between the sample commodities based on the relevant information of the plurality of sample commodities, and construct an attribute graph reflecting the attribute correlation relationship between the sample commodities based on the attribute characteristics of the plurality of sample commodities.

[0032] 203. Train an initial graph neural network model based on the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal product identification.

[0033] In this embodiment, the sample products can be products of various product categories provided on the e-commerce platform, and the various product categories include, but are not limited to: maternal and child products, clothing, shoes and bags, festival gifts, grains, oils and seasonings, and electrical appliances and digital products.

[0034] In this embodiment, the sample label commodity refers to a sample commodity with a real label, wherein the real label is an abnormal label or a normal label. The abnormal label indicates that the sample label commodity is a sample abnormal commodity, and the normal label indicates that the sample label commodity is a sample normal commodity. Among them, the sample abnormal commodity is prohibited from being sold on the e-commerce platform, and the sample normal commodity is allowed to be sold on the e-commerce platform.

[0035] In this embodiment, multiple rounds of iterative model training are performed to obtain a target graph neural network model that can be used for abnormal commodity identification. In each round of model training, one or more training samples can be prepared, and one or more training samples can be used for this round of model training. Among them, when preparing each training sample, it is necessary to prepare multiple sample commodities including at least one sample labeled commodity and one sample abnormal commodity. In addition to sample labeled commodities and sample abnormal commodities, the multiple sample commodities may also include other sample commodities. Other sample commodities may be sample labeled commodities with known true labels (which can also be understood as sample commodities with true labels), or other sample commodities may also be sample commodities with unknown true labels (which can also be understood as sample commodities without true labels).

[0036] In this embodiment, after obtaining multiple sample commodities, relevant information of each sample commodity is obtained, and by performing big data analysis on the relevant information of the multiple sample commodities, it is possible to mine out pairs of sample commodities with risk association relationships, and construct a risk map reflecting the risk association relationship between the sample commodities based on the mined pairs of sample commodities with risk association relationships. Optionally, the relevant information of each sample commodity includes, but is not limited to: seller information, buyer information, and historical access information. Among them, the historical access information includes, but is not limited to: information of viewers who browsed the sample commodity, historical click-through rate, historical number of add-to-cart operations, and historical number of favorite operations, etc.

[0037] In this embodiment, the probability that the sample goods with risk association relationships are abnormal goods is relatively high. In practical applications, risk association relationships can be flexibly defined as needed. Risk association relationships include, but are not limited to, same seller relationships, similar seller relationships, or same click relationships. Among them, two sample goods with the same seller relationship refer to the same sellers selling the two sample goods. Two sample goods with similar seller relationships refer to similar sellers selling the two sample goods. Two sample goods with the same click relationship refer to the same viewers browsing the two sample goods. It is worth noting that if the device ID (Device Identifier) ​​or IP address (Internet Protocol Address) of the electronic devices used by multiple sellers is the same or similar, then the multiple sellers are the same sellers or similar sellers. Among them, the device ID includes, but is not limited to, IMEI (International Mobile Equipment Identity), MAC (Media Access Control Address).

[0038] In actual applications, there may be one or more risk association relationships between any two sample products. For example, if there is a same-seller relationship, a similar seller relationship, or a same-click relationship between sample product A and sample product B, then three edges connecting the nodes corresponding to sample product A and sample product B are created in the risk graph, wherein one edge represents the same-seller relationship, one edge represents the similar seller relationship, and one edge represents the same-click relationship. Therefore, in an optional implementation method, based on the relevant information of multiple sample products, constructing a risk graph reflecting the risk association relationship between sample products can be: abstracting the sample products as nodes in the risk graph; determining every two sample products with a risk association relationship based on the relevant information of multiple sample products, and creating an edge connecting the corresponding two nodes in the risk graph for every two sample products with a risk association relationship, wherein one edge is created for each risk association relationship.

[0039] In an embodiment of the present application, in addition to constructing a risk graph reflecting the risk association relationship between sample commodities, it is also necessary to construct an attribute graph reflecting the attribute association relationship between sample commodities based on the attribute characteristics of multiple sample commodities. Optionally, the attribute characteristics of the sample commodities include, but are not limited to: material, style, product name, place name, brand, usage method, model or type, etc. In an optional implementation method, constructing an attribute graph reflecting the attribute association relationship between sample commodities based on the attribute characteristics of multiple sample commodities is specifically: abstracting the sample commodities as nodes in the attribute graph; for any two sample commodities, calculating the first attribute similarity between the two sample commodities based on the attribute characteristics of the two sample commodities; determining that there is an attribute association relationship between two sample commodities whose first attribute similarity is greater than the first similarity threshold, and creating an edge connecting the corresponding two nodes in the attribute graph for the two sample commodities with the attribute association relationship.

[0040] It is worth noting that the first attribute similarity reflects the similarity between the attribute characteristics of the two sample products. For any two sample products, if the corresponding first attribute similarity is greater than the first similarity threshold, the attribute characteristics of the two sample products are highly similar, and it is determined that there is an attribute association relationship between the two sample products. At this time, it is necessary to create an edge between the corresponding nodes of the two sample products in the attribute graph. If the corresponding first attribute similarity is less than or equal to the first similarity threshold, the attribute characteristics of the two sample products are less similar, and it is determined that there is no attribute association relationship between the two sample products. At this time, it is not necessary to create an edge between the corresponding nodes of the two sample products in the attribute graph. Among them, the first similarity threshold can be flexibly set according to actual application requirements.

[0041] In this embodiment, after constructing the risk graph and the attribute graph, an initial graph neural network model is trained according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal commodity identification. The initial graph neural network model is an untrained graph neural network model, and the target graph neural network model is a trained graph neural network model.

[0042] In practical applications, the risk graph and the attribute graph are input into the initial graph neural network model to predict the probability that the label of the sample label product is an abnormal label, and the loss function is calculated based on the predicted label results of the sample label product and its true label. The model parameters of the initial graph neural network model are adjusted according to the loss function. If the current round of model training meets the model training end condition, the model training is stopped, and the updated initial graph neural network model is used as the final target graph neural network model. If the current round of model training does not meet the model training end condition, steps 201 to 203 are re-executed to continue model training until the current round of model training meets the model training end condition. Among them, the model training end condition is set according to the actual situation. The model training end condition is, for example, the loss function calculated based on the predicted label results of the sample label product and its true label converges, or the number of model training times in this round reaches the maximum number of model training times.

[0043] In the embodiment of the present application, there is no restriction on the loss function, and the loss function includes, but is not limited to: log loss function, cross-entropy loss function and focal loss function for solving data imbalance problem.

[0044] The model training method provided in the embodiment of the present application performs graph neural network model training based on a risk graph reflecting the risk correlation relationship between sample commodities and an attribute graph reflecting the attribute correlation relationship between sample commodities. Since the risk correlation relationship between sample commodities and the attribute correlation relationship between sample commodities are fully considered during the model training process, the target graph neural network model obtained by such training can effectively measure the risk correlation between commodities, and thus accurately identify abnormal commodities.

[0045] In some optional embodiments of the present application, step 203 may be implemented by: inputting the risk graph and the attribute graph into the initial graph neural network model for graph structure learning to obtain a first product feature representation of the sample labeled product and a second product feature representation of the sample abnormal product; based on the first product feature representation, the second product feature representation and the true label, updating the model parameters of the initial graph neural network model to obtain the target graph neural network model.

[0046] Optionally, the risk graph and the attribute graph are input into the initial graph neural network model for graph structure learning to obtain the first product feature representation of the sample labeled product and the second product feature representation of the sample abnormal product. The implementation method may be: for any first sample product among multiple sample products, determine at least one first connection edge corresponding to the first sample product from the risk graph and the attribute graph, each first connection edge connecting a node corresponding to the first sample product and a node corresponding to a neighbor product of the first sample product; generate a vectorized representation to be spliced ​​according to the edge weight of at least one first connection edge and the corresponding vectorized representation of the attribute features of at least one neighbor product; perform vector splicing on the vectorized representation of the attribute features of the first sample product and the vectorized representation to be spliced ​​to obtain the product feature representation of the first sample product; wherein, when the first sample product is a sample labeled product, the product feature representation of the first sample product refers to the first product feature representation; when the first sample product is a sample abnormal product, the product feature representation of the first sample product refers to the second product feature representation.

[0047] In this embodiment, the first sample commodity is any sample commodity among a plurality of sample commodities, and one or more edges connecting the corresponding nodes of the first sample commodity are found from the risk graph and the attribute graph. For ease of understanding, the edges connecting the corresponding nodes of the first sample commodity are referred to as first edges. It should be understood that, in addition to connecting the node corresponding to the first sample commodity, the first edge also connects the node corresponding to a neighbor commodity of the first sample commodity. After finding one or more first edges from the risk graph and the attribute graph, a vectorized representation to be processed can be generated according to the edge weight of each first edge and the vectorized representation of the attribute characteristics of the corresponding neighbor commodity; and vector addition is performed on at least one vectorized representation to be processed to obtain a vectorized representation to be spliced; finally, the vectorized representation of the attribute characteristics of the first sample commodity and the vectorized representation to be spliced ​​are vector spliced ​​to obtain the commodity feature representation of the first sample commodity.

[0048] In an embodiment of the present application, the edge weight of any edge in a risk graph or an attribute graph can be determined based on the attribute characteristics of the commodity corresponding to the edge and the edge type to which the edge belongs. In a risk graph, one risk association relationship corresponds to one edge type, and different risk association relationships correspond to different edge types. Therefore, in a risk graph, multiple edges of different edge types may appear between two nodes, and the risk graph can be considered a heterogeneous graph; in an attribute graph, there is only one edge type that reflects an attribute association relationship, so in an attribute graph, only one edge will appear between two nodes, and no multiple edges will appear. The attribute graph can be considered a homogeneous graph.

[0049] In practical applications, for the edge weight of any link, the attribute similarity between the two commodities in the same attribute category can be calculated based on the prior weight corresponding to the edge type of the link and the attribute characteristics of the two commodities corresponding to the link in the same attribute category; at least one attribute similarity between the two commodities in at least one attribute category can be summed, averaged, or weighted averaged, and the result of the operation can be used as the edge weight of the link. The prior weight can be flexibly set according to the actual application requirements.

[0050] For ease of understanding, the following explanation is given in conjunction with formulas (1) to (4).

[0051] Assume that the first sample product is recorded as v i , v i The vectorized representation of the attribute features is denoted as h i ,in,

[0052] h i =σ(X i W+b)……(1)

[0053] Among them, X i Indicates v i The attribute characteristics of R represents the set of real numbers, d a Represents attribute feature X i dimension; σ(·) represents the activation function, optionally, σ(·) is the RELU(·) activation function; b represents the bias value; W represents X i With h i The mapping matrix between d means h i The dimension of . Among them, b and W are one of the learnable model parameters of the graph neural network model. It is worth noting that the vectorized representation of the attribute features of any product can be obtained according to formula (1).

[0054] Assume that a neighbor product of the first sample product is denoted as v j , neighbor product v j The vectorized representation of the attribute features is denoted as h j , for product v i and neighbor products v j The ψ type of edge between them, ψ type is any edge type, assuming that the product v i and neighbor products v j The edge weight of the ψ-type edge between them is denoted by Γ ψ (v i ,v j ),in,

[0055] Γ ψ (m) (vi ,v j )=f(γ ψ h i (m) W ψ (h j (m) ) T )……(2)

[0056]

[0057] Among them, Γ ψ (m) (v i ,v j ) indicates product v i and neighbor products v j The attribute similarity of the mth aspect (one aspect corresponds to one attribute category) on the ψ type edge. Γ ψ (v i ,v j ) is the product v i and neighbor products v j It is obtained by summing the attribute similarities of M aspects on the ψ type edge, where m and M are positive integers, m ranges from 1 to M, and Γ ψ (v i ,v j )∈R M ×1×1 .

[0058] Among them, the product v i The vectorized representation of the attribute features h i A tensor divided into multiple heads, h i (m) From h i The m-th tensor among the M tensors divided in .

[0059] Among them, the neighbor product v j The vectorized representation of the attribute features h j A tensor divided into multiple heads, h j (m) is the mth tensor among the M tensors divided, (h j (m) ) T Yes j (m) The transposed matrix of .

[0060] Among them, W ψ Indicates h i (m) and (h j (m) ) T The mapping matrix between W ψ It is one of the learnable model parameters of the graph neural network model.

[0061] Among them, γ ψ is the prior weight corresponding to the edge type of ψ, γ ψ The value of can be set flexibly, for example The f(·) function is used to ensure that the value range of attribute similarity is between 0 and 1. The f(·) function can be any function. For example, the f(·) function includes but is not limited to: sigmoid(·) activation function, ReLU activation function, and Tan hyperbolic tangent activation function.

[0062] In this embodiment, after determining the edge weights of each edge in the risk graph and the attribute graph, i , product v i The commodity feature representation is denoted by H i ,in,

[0063]

[0064] Among them, a i,j,ψ Represents the commodity v in the risk graph or attribute graph i and neighbor products v j The corresponding ψ type connection

[0065] The edge weight of the edge. represents the set of edge types in the risk graph, and A represents the set of edge types in the attribute graph.

[0066] in, It indicates that the ψ edge type is any one of all edge types, and all edge types come from the set of edge types in the risk graph and the set of edge types in the attribute graph.

[0067] in, Represents the neighbor product v j Is with product v i Any of all neighbor commodities with an edge of type ψ.

[0068] in, It is a mapping matrix and one of the learnable model parameters of the graph neural network model.

[0069] Among them, HeteAGG(·) is a heterogeneous aggregator. First, based on v i The edge weight of each link and the corresponding neighbor product v j The vectorized representation of the attribute features of the For all Add the vectors to get the vectorized representation to be spliced ​​to achieve neighborhood information aggregation. Then, add the vectorized representation to be spliced ​​to the product v i The vectorized representation of the attribute features h i Perform vector splicing to obtain product v i The product feature representation H i .

[0070] It is worth noting that in determining the product v i The product feature representation H i In the process, considering the aggregation of neighborhood information can effectively alleviate the problem of low same-matching rate between products (that is, the problem that most of the abnormal product associations are normal products).

[0071] In some optional embodiments of the present application, during the graph learning process, the noise edges in the risk graph or attribute graph can also be filtered out so as to perform graph learning based on the high-quality risk graph or attribute graph. In an embodiment of the present application, the noise edge can refer to an edge with a low attribute correlation, or an edge with a low risk correlation. Therefore, further optionally, before determining at least one first edge corresponding to the first sample commodity from the risk graph and the attribute graph, for any second edge in the risk graph or the attribute graph, the second edge connects the node corresponding to the second sample commodity and the node corresponding to the third sample commodity, and according to the edge type of the second edge, the attribute characteristics of the second sample commodity and the third sample commodity under at least one attribute category, the second edge is deleted to update the risk graph or the attribute graph.

[0072] In this embodiment, the second edge is any edge in the risk graph or the attribute graph, and it is determined whether the second edge is a noise edge. If so, the second edge is deleted; if not, the edge is retained in the risk graph or the attribute graph.

[0073] In an optional implementation, according to the edge type of the second edge, the attribute characteristics of the second sample product and the third sample product under at least one attribute category, the second edge is deleted specifically by: calculating the second attribute similarity between the second sample product and the third sample product under the same attribute category according to the prior weight corresponding to the edge type of the second edge and the attribute characteristics of the second sample product and the third sample product under the same attribute category; averaging at least one second attribute similarity between the second sample product and the third sample product under at least one attribute category to obtain the second attribute similarity mean; if the second attribute similarity mean is less than or equal to the second similarity threshold, deleting the second edge. Further optionally, if the second attribute similarity mean is greater than the second similarity threshold, summing at least one second attribute similarity between the second sample product and the third sample product under at least one attribute category to obtain the second attribute similarity sum, and using the second attribute similarity sum as the edge weight of the second edge.

[0074] In this embodiment, the second similarity threshold is flexibly set according to actual application requirements. If the second attribute similarity mean is less than or equal to the second similarity threshold, it means that the risk correlation or attribute correlation between the second sample product and the third sample product is poor; if the second attribute similarity mean is greater than the second similarity threshold, it means that the risk correlation or attribute correlation between the second sample product and the third sample product is good.

[0075] It is worth noting that in the process of graph structure learning, by measuring the risk correlation or attribute correlation between products, it is possible to reduce the noise edges in the risk graph or attribute graph and connect risk-related but unconnected products, thereby effectively transmitting information and representing product features on the graph.

[0076] For ease of understanding, we will explain it in conjunction with formula (5):

[0077]

[0078] Where,∈ is the second similarity threshold that can be set flexibly. refers to the second attribute similarity, refers to the mean similarity of the second attribute, Refers to the sum of the similarities of the second attribute. When the condition is ψ (v i , v j ) that is As a commodity i and neighbor products v j The edge weight of the ψ-type edge between In dissatisfaction If the condition is i and neighbor products v j The initial edge weight of the ψ-type edge between Set to 0, which means confirming the product v i and neighbor products v j The ψ-type edges between are noise edges and need to be deleted. i The corresponding edge weights of multiple edges For product v i The corresponding edge weights of multiple edges Perform normalization operation to obtain product v i The final edge weight a of the corresponding multiple edges i,j,ψ Among them, the normalized product v i The final edge weight a of the corresponding multiple edges i,j,ψ The sum of is 1.

[0079] In some embodiments of the present application, the model parameters of the initial graph neural network model are updated according to the first product feature representation, the second product feature representation and the true label to obtain the target graph neural network model. The implementation method can be: input the first product feature representation and the second product feature representation into the multi-layer perceptron layer in the initial graph neural network model for metric learning to predict the probability that the label of the sample label product is an abnormal label; according to the true label of the sample label product and its predicted label result, the model parameters of the initial graph neural network model are updated to obtain the target graph neural network model.

[0080] Specifically, a multilayer perceptron (MLP) layer is added to the initial graph neural network model, and the similarity between the sample label product and the sample abnormal product is determined based on the first product feature representation of the sample label product and the second product feature representation of the sample abnormal product by using the multilayer perceptron layer, and the probability that the label of the sample label product is an abnormal label is accurately predicted based on the similarity between the sample label product and the sample abnormal product. After obtaining the predicted label result of the sample label product, a loss function is constructed based on the predicted label result and the true label of the sample label product, and the model parameters of the initial graph neural network model are updated based on the loss function to obtain the target graph neural network model.

[0081] For example, suppose the product set including multiple sample label products is denoted as v L , from v L Select several sample abnormal commodities with abnormal labels, and record the commodity set including several sample abnormal commodities with abnormal labels as

[0082] For product collection Any sample label product v i , and product collections Any abnormal product in the sample v s , calculate the sample label product v i With sample abnormal product v s The attribute similarity between a ( i , s ),in:

[0083]

[0084] Among them, X i is the sample label product v i The attribute characteristics of X S is the sample abnormal product v s Attribute characteristics.

[0085] For any sample label product v i , according to the attribute similarity r a ( i , s ) For multiple sample abnormal products v s Sort in descending order and select multiple sample abnormal products v that are ranked first s Respectively with sample label product v i Construct training samples. Specifically, each training sample can include a sample abnormal product v s and sample label product v i In the sample label product v i When the sample is an abnormal product, the training sample is a positive sample; when the sample label product v i When the sample is a normal product, the training sample is a negative sample.

[0086] As an example, the loss function constructed based on the predicted label result and the true label of the sample label product can be defined as follows:

[0087]

[0088] Among them, Y i，s Reflects the real label of the sample label product, is the MLP based on sample abnormal products v s The product feature representation H S and sample label product v i The product feature representation H i The prediction result is, log() is a logarithmic operation; ∑ is the summation symbol.

[0089] It is worth noting that in metric learning, the commodity classification task is converted into a metric learning task between candidate commodities and abnormal commodities. Specifically, for any sample label commodity v i Select the most relevant multiple sample abnormal products v s , and for any sample label product v i Respectively with multiple sample abnormal products v s Each sample of abnormal products in the dataset is composed of paired training instances. Multiple paired training instances are used to perform paired metric learning on multiple subclasses, and then the supervision information of the graph neural network model is enhanced and enriched from the two levels of sample size and product feature representation.

[0090] It is worth noting that in each training, multiple training samples can be used for training, and the loss function for each training sample is Sum and get the final loss function Designing such a loss function can enhance and enrich the supervision information of the graph neural network model from two levels: sample size and product feature representation. In addition to ensuring effective neighbor information aggregation based on a good graph structure, the graph neural network model also uses sufficient fine-grained supervision information for parameter optimization. In this way, the trained graph neural network model can better handle large-scale graph data.

[0091] Further optionally, in order to enhance the model performance of the target graph neural network model, the model parameters of the initial graph neural network model are updated according to the true labels of the sample label products and their predicted label results to obtain the target graph neural network model. The implementation method is as follows: an initial loss function is created according to the true labels of the sample label products and their predicted label results; a regularization term is generated according to the current model parameters of the initial graph neural network model, and a target loss function is generated according to the initial loss function and the regularization term; according to the target loss function, the current model parameters of the initial graph neural network model are updated to obtain the target graph neural network model.

[0092] For example, the objective loss function is calculated as follows:

[0093]

[0094] in, Denoted as the initial loss function of the t-th model training, it can be obtained according to formula (7); is the target loss function for the t-th model training, t is a positive integer; ξ represents the weight factor of the set Θ of all model parameters to be learned, and Ω represents the L2 regularization term, which is used to ensure the generalization performance of the model and avoid overfitting.

[0095] In practical applications, the attribute features of abnormal products may be designed by sellers to be very similar to those of normal products. The original attribute graph generated based on the attribute information of the product may have low compatibility. The compatibility can reflect the attribute similarity between the products corresponding to the nodes connected by the edge to a certain extent. Such an attribute graph has a great impact on the model performance of the target graph neural network model. Considering that the product feature representation learned by the graph structure contains potential risk features, an iterative graph structure optimization mechanism is adopted to optimize the attribute graph using the learned product feature representation to ensure the smoothness of the features and the compatibility of the attribute graph during the training process.

[0096] Therefore, further optionally, after each model training, for any fourth sample product among the multiple sample products, the feature similarity between the fourth sample product and each fifth sample product among the multiple sample products is calculated based on the product feature representation of the fourth sample product and the product feature representation of each fifth sample product among the multiple sample products; if the number of fifth sample products whose feature similarity with the fourth sample product is greater than the third similarity threshold is less than or equal to the set number, then an edge is created between the nodes corresponding to the fourth sample product and the fifth sample product in the attribute graph; if the number of fifth sample products whose feature similarity with the fourth sample product is greater than the third similarity threshold is greater than the set number, then the edge between the nodes corresponding to the fourth sample product and the fifth sample product is deleted in the attribute graph. The set number and the third similarity threshold are set according to actual application requirements.

[0097] Specifically, each time the model is trained, it is trained based on the risk graph and attribute graph corresponding to multiple sample products. After the model training, the attribute graph is updated according to the product feature representations of the multiple sample products to ensure the smoothness of the features and the matching of the attribute graph during the training process. This is explained below in conjunction with formula (9).

[0098] Assume that the product feature representation H is trained for the t-1th time t-1 , a certain edge in the attribute graph corresponding to the t-th training is generated and recorded as Right now

[0099]

[0100] in, Represents the product v in the attribute graph A corresponding to the t-th training i and neighbor products v j The edges between Represents the product v based on the t-1th training i The product feature representation H t-1 and neighbor products v j The product feature representation H t-1 Calculate the cosine similarity and use the obtained cosine similarity as the product v i and neighbor products v j The feature similarity between them. i (·) indicates that the Neighboring products v after cosine similarity is sorted in descending order j Select neighbor products v whose cosine similarity is greater than the third similarity threshold j The number of, where k is the set number. When the product v i and neighbor products v j Create an edge between them. When the product v i and neighbor products v j Delete the edges between them.

[0101] For ease of understanding, combined Figure 3 The application scenario diagram shown in the figure illustrates the model training method provided in the embodiment of the present application. Figure 3 , during the model training phase, build a risk map and property graph The risk map Each circle in the graph represents a node, and the edge types between nodes include Ψ1 edge type, Ψ2 edge type, and Ψ3 edge type. Each circle in the graph represents a node, and the edge types between nodes are all of the same edge type, which can be called attribute edges. and property graph Input into the initial graph neural network model for graph structure learning, for any product v i , for product v i and any neighbor product v j Perform heterogeneous similarity measurement. When performing heterogeneous similarity measurement, calculate the product v i and neighbor products v j The edge weight Γ of the ψ-type edge between ψ (v i ,v j ),as well as like The comparison result with the controllable threshold ∈ is shown in the risk graph and property graph Delete including product v i For example, Figure 3 In, v i and v j2 The edges between them are deleted, v i and v j4 The edges between them are deleted. In graph structure learning, the heterogeneous information propagation mechanism is used to determine the product v i The current neighbor product of i The vectorized representation of the attribute features of the product v is fused with the vectorized representation of the attribute features of each current neighbor product to obtain the product v i The final product feature representation H i In the paired metric learning stage, paired metric learning is performed on the product feature representation of the sample label products in the paired training instances and the product feature representation of the sample abnormal products. Figure 3 From the tag product collection Select the product with sample label v iThe most relevant sample abnormal products v s1 、v s2 , v i With v s1 v s2 They form pairs of training instances. i With sample abnormal product v s3 The similarity is not high, v i With v s2 Do not form paired training instances.

[0102] The target graph neural network model trained by the model training method provided in the embodiment of the present application can be used to identify abnormal commodities. To this end, the embodiment of the present application also provides an abnormal commodity identification method. Figure 4 This is a flow chart of a method for identifying abnormal commodities provided in an embodiment of the present application. Figure 4 , the method may include the following steps:

[0103] 401. Acquire multiple commodities, where the multiple commodities at least include commodities to be identified and abnormal commodities with abnormal labels.

[0104] 402. Construct a target risk graph reflecting the risk correlation relationship between commodities based on the risk information of the plurality of commodities, and construct a target attribute graph reflecting the attribute correlation relationship between commodities based on the attribute characteristics of the plurality of commodities.

[0105] 403. Input the target risk graph and the target attribute graph into the target graph neural network model to identify abnormal commodities, so as to obtain the identification result of the commodity to be identified.

[0106] The identification result of the commodity to be identified may include the probability that the commodity to be identified is an abnormal commodity. In practical applications, the information of the commodity to be identified with a higher probability may also be pushed to relevant reviewers for manual review to reduce the misidentification rate of abnormal commodities.

[0107] The abnormal commodity identification method provided in the embodiment of the present application fully considers the risk correlation relationship and the attribute correlation relationship between sample commodities during the model training process in the target graph neural network model. Therefore, abnormal commodities can be accurately identified by using such a target graph neural network model.

[0108] Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of the present application. Figure 5 , the device may include:

[0109] An acquisition module 51 is used to acquire a plurality of sample commodities, wherein the plurality of sample commodities at least include sample labeled commodities with real labels and sample abnormal commodities with abnormal labels, wherein the real labels are abnormal labels or normal labels;

[0110] A construction module 52, configured to construct a risk graph reflecting the risk association relationship between the sample commodities based on the relevant information of the plurality of sample commodities, and to construct an attribute graph reflecting the attribute association relationship between the sample commodities based on the attribute characteristics of the plurality of sample commodities;

[0111] The training module 53 is used to train the initial graph neural network model according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal product identification.

[0112] Further optionally, when the construction module 52 constructs a risk map reflecting the risk association relationship between the sample commodities based on the relevant information of the plurality of sample commodities, it is specifically used to:

[0113] The sample commodities are abstracted as nodes in the risk graph; based on the relevant information of the multiple sample commodities, each two sample commodities with a risk association relationship are determined, and for each two sample commodities with a risk association relationship, an edge connecting the corresponding two nodes is created in the risk graph, wherein one edge is created for each risk association relationship.

[0114] Further optionally, when the construction module 52 constructs an attribute graph reflecting the attribute association relationship between sample products based on the attribute characteristics of multiple sample products, it is specifically used to: abstract the sample products as nodes in the attribute graph; for any two sample products, calculate the first attribute similarity between the two sample products based on their respective attribute characteristics; determine that there is an attribute association relationship between two sample products whose first attribute similarity is greater than a first similarity threshold, and create an edge connecting the corresponding two nodes in the attribute graph for the two sample products that have the attribute association relationship.

[0115] Further optionally, when the training module 53 trains the initial graph neural network model according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal product identification, it is specifically used to: input the risk graph and the attribute graph into the initial graph neural network model for graph structure learning to obtain a first product feature representation of the sample labeled product and a second product feature representation of the sample abnormal product; based on the first product feature representation, the second product feature representation and the true label, update the model parameters of the initial graph neural network model to obtain the target graph neural network model.

[0116] Further optionally, the training module 53 updates the model parameters of the initial graph neural network model according to the first product feature representation, the second product feature representation and the true label to obtain the target graph neural network model. It is specifically used to: input the first product feature representation and the second product feature representation into the multi-layer perceptron layer in the initial graph neural network model for metric learning to predict the probability that the label of the sample label product is an abnormal label; update the model parameters of the initial graph neural network model according to the true label of the sample label product and its predicted label result to obtain the target graph neural network model.

[0117] Further optionally, when the training module 53 updates the model parameters of the initial graph neural network model according to the real labels of the sample label products and the predicted label results to obtain the target graph neural network model, it is specifically used to: create an initial loss function according to the real labels of the sample label products and the predicted label results; generate a regularization term according to the current model parameters of the initial graph neural network model, and generate a target loss function according to the initial loss function and the regularization term; and update the current model parameters of the initial graph neural network model according to the target loss function to obtain the target graph neural network model.

[0118] Further optionally, when the training module 53 inputs the risk graph and the attribute graph into the initial graph neural network model for graph structure learning to obtain the first commodity feature representation of the sample label commodity and the second commodity feature representation of the sample abnormal commodity, it is specifically used to: for any first sample commodity among multiple sample commodities, determine at least one first connecting edge corresponding to the first sample commodity from the risk graph and the attribute graph, each first connecting edge connecting a node corresponding to the first sample commodity and a node corresponding to a neighbor commodity of the first sample commodity; generate a vectorized representation to be spliced ​​according to the edge weight of at least one first connecting edge and the vectorized representation of the attribute characteristics of at least one corresponding neighbor commodity; perform vector splicing on the vectorized representation of the attribute characteristics of the first sample commodity and the vectorized representation to be spliced ​​to obtain the commodity feature representation of the first sample commodity;

[0119] Among them, when the first sample product is a sample label product, the product characteristic representation of the first sample product refers to the first product characteristic representation; when the first sample product is a sample abnormal product, the product characteristic representation of the first sample product refers to the second product characteristic representation.

[0120] Further optionally, before determining at least one first edge corresponding to the first sample product from the risk graph and the attribute graph, the training module 53 is also used to: for any second edge in the risk graph or the attribute graph, delete the second edge according to the edge type of the second edge and the attribute characteristics of the second sample product and the third sample product under at least one attribute category to update the risk graph or the attribute graph; wherein the second edge connects the node corresponding to the second sample product and the node corresponding to the third sample product.

[0121] Further optionally, the training module 53 deletes the second edge according to the edge type of the second edge and the attribute characteristics of the second sample product and the third sample product under at least one attribute category to update the risk graph or the attribute graph, and is specifically used to: calculate the second attribute similarity between the second sample product and the third sample product in the same attribute category according to the prior weight corresponding to the edge type of the second edge and the attribute characteristics of the second sample product and the third sample product under the same attribute category; average at least one second attribute similarity between the second sample product and the third sample product in at least one attribute category to obtain the second attribute similarity mean; if the second attribute similarity mean is less than or equal to the second similarity threshold, delete the second edge.

[0122] Further optionally, the training module 53 is also used to: if the mean second attribute similarity is greater than the second similarity threshold, sum up at least one second attribute similarity between the second sample product and the third sample product in at least one attribute category to obtain the sum of the second attribute similarities, and use the sum of the second attribute similarities as the edge weight of the second edge.

[0123] Further optionally, the training module 53 is also used to: after each model training, for any fourth sample product among the multiple sample products, calculate the feature similarity between the fourth sample product and each fifth sample product among the multiple sample products based on the product feature representation of the fourth sample product and the product feature representation of each fifth sample product among the multiple sample products; if the number of fifth sample products whose feature similarity with the fourth sample product is greater than a third similarity threshold is less than or equal to a set number, create an edge between the nodes corresponding to the fourth sample product and the fifth sample product in the attribute graph; if the number of fifth sample products whose feature similarity with the fourth sample product is greater than the third similarity threshold is greater than a set number, delete the edge between the nodes corresponding to the fourth sample product and the fifth sample product in the attribute graph.

[0124] Figure 5 The model training apparatus shown can be executed Figure 2 The implementation principle and technical effect of the model training method in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the model training device in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0125] Figure 6 This is a schematic diagram of the structure of an abnormal commodity identification device provided in an embodiment of the present application. Figure 6 , the device may include:

[0126] An acquisition module 61 is used to acquire a plurality of commodities, wherein the plurality of commodities at least include commodities to be identified and abnormal commodities with abnormal labels;

[0127] A construction module 62 is used to construct a target risk graph reflecting the risk association relationship between commodities based on the risk information of the plurality of commodities, and to construct a target attribute graph reflecting the attribute association relationship between commodities based on the attribute characteristics of the plurality of commodities;

[0128] The processing module 63 is used to input the target risk map and the target attribute map into the target graph neural network model to identify abnormal commodities to obtain the identification results of the commodities to be identified; wherein, the target graph neural network model is trained by the model training method provided in the above-mentioned embodiment.

[0129] Figure 6 The abnormal commodity identification device shown can perform Figure 5 The implementation principle and technical effect of the abnormal commodity identification method in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the abnormal commodity identification device in the above embodiment has been described in detail in the embodiment of the method, and will not be described in detail here.

[0130] It should be noted that the execution subject of each step of the method provided in the above embodiment can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 201 to 203 can be device A; for another example, the execution subject of steps 201 and 202 can be device A, and the execution subject of step 203 can be device B; and so on.

[0131] In addition, in some of the processes described in the above embodiments and the accompanying drawings, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel, and the sequence numbers of the operations, such as 201, 202, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0132] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown, the electronic device includes: a memory 71 and a processor 72;

[0133] The memory 71 is used to store computer programs and can be configured to store various other data to support operations on the computing platform. Examples of such data include instructions for any application or method operating on the computing platform, contact data, phone book data, messages, pictures, videos, etc.

[0134] The memory 71 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0135] The processor 72 is coupled to the memory 71 and is used to execute the computer program in the memory 71, so as to: obtain a plurality of sample commodities, wherein the plurality of sample commodities at least include sample labeled commodities with real labels and sample abnormal commodities with abnormal labels, wherein the real labels are abnormal labels or normal labels; construct a risk graph reflecting the risk association relationship between the sample commodities according to the relevant information of the plurality of sample commodities, and construct an attribute graph reflecting the attribute association relationship between the sample commodities according to the attribute characteristics of the plurality of sample commodities; train an initial graph neural network model according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal commodity identification, or,

[0136] The processor 72 is coupled to the memory 71 and is used to execute the computer program in the memory 71, so as to: obtain multiple commodities, wherein the multiple commodities at least include commodities to be identified and abnormal commodities with abnormal labels; construct a target risk map reflecting the risk association relationship between commodities based on the risk information of the multiple commodities, and construct a target attribute map reflecting the attribute association relationship between commodities based on the attribute characteristics of the multiple commodities; input the target risk map and the target attribute map into the target map neural network model to identify abnormal commodities, so as to obtain the identification results of the commodities to be identified.

[0137] For the detailed implementation process of the processor executing each step, please refer to the relevant description in the aforementioned method embodiment, which will not be repeated here.

[0138] Further, if Figure 7 As shown, the electronic device also includes: a communication component 73, a display 74, a power component 75, an audio component 76 and other components. Figure 7 Only some components are shown schematically, which does not mean that the electronic device only includes Figure 7 In addition, Figure 7The components in the dashed box are optional components, not mandatory components, and the specific components depend on the product form of the electronic device. The electronic device of this embodiment can be implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, or an IOT device, or a server device such as a conventional server, a cloud server, or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, a smart phone, etc., it can include Figure 7 If the electronic device of this embodiment is implemented as a server device such as a conventional server, a cloud server or a server array, it may not include Figure 7 Components within the dashed box.

[0139] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, which, when executed, can implement each step that can be executed by an electronic device in the above method embodiment.

[0140] Accordingly, an embodiment of the present application also provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement each step in the above method embodiment that can be executed by an electronic device.

[0141] The above-mentioned communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0142] The above-mentioned display includes a screen, and the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundary of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0143] The power supply assembly provides power to various components of the device where the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply assembly is located.

[0144] The above-mentioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting an audio signal.

[0145] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0146] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0147] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0148] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0149] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0150] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0151] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0152] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0153] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A model training method, characterized in that: include: Acquire a plurality of sample commodities, wherein the plurality of sample commodities at least include sample labeled commodities with real labels and sample abnormal commodities with abnormal labels, wherein the real labels are abnormal labels or normal labels; constructing a risk graph reflecting the risk association relationship between the sample commodities based on the relevant information of the plurality of sample commodities, and constructing an attribute graph reflecting the attribute association relationship between the sample commodities based on the attribute characteristics of the plurality of sample commodities; For any second edge in the risk graph or the attribute graph, calculate the second attribute similarity between the second sample product and the third sample product in the same attribute category according to the prior weight corresponding to the edge type of the second edge and the attribute characteristics of the second sample product and the third sample product in the same attribute category; average at least one second attribute similarity between the second sample product and the third sample product in at least one attribute category to obtain a second attribute similarity mean; If the second attribute similarity mean is less than or equal to the second similarity threshold, the second edge is deleted; wherein, the second edge is an edge connecting the node corresponding to the second sample product and the node corresponding to the third sample product, the nodes in the risk graph and the attribute graph respectively represent a sample product, and the edge in the risk graph represents a risk association relationship; the edge in the attribute graph represents an attribute association relationship, and the second sample product and the third sample product are two different sample products among multiple sample products; an initial graph neural network model is trained according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal product identification.

2. The method according to claim 1, characterized in that: Constructing a risk map reflecting the risk association relationship between the sample commodities based on the relevant information of the plurality of sample commodities, including: Abstracting the sample commodities into nodes in the risk graph; Based on the relevant information of the multiple sample commodities, each two sample commodities having a risk association relationship are determined, and for each two sample commodities having a risk association relationship, an edge connecting two corresponding nodes is created in the risk graph, wherein one edge is created for each risk association relationship.

3. The method according to claim 1, characterized in that According to the attribute characteristics of the plurality of sample products, an attribute graph reflecting the attribute association relationship between the sample products is constructed, including: Abstracting the sample product into a node in the attribute graph; For any two sample products, the similarity of the first attribute between the two sample products is calculated according to the attribute characteristics of the two sample products; It is determined that an attribute association relationship exists between two sample products whose first attribute similarity is greater than a first similarity threshold, and an edge connecting two corresponding nodes is created in the attribute graph for the two sample products having the attribute association relationship.

4. The method according to claim 1, characterized in that: An initial graph neural network model is trained according to the risk graph and the attribute graph to obtain a target graph neural network model that can be used for abnormal commodity identification, including: Inputting the risk graph and the attribute graph into the initial graph neural network model for graph structure learning to obtain a first commodity feature representation of the sample labeled commodity and a second commodity feature representation of the sample abnormal commodity; According to the first product feature representation, the second product feature representation and the true label, the model parameters of the initial graph neural network model are updated to obtain the target graph neural network model.

5. The method according to claim 4, characterized in that According to the first commodity feature representation, the second commodity feature representation, and the true label, updating the model parameters of the initial graph neural network model to obtain the target graph neural network model includes: Inputting the first commodity feature representation and the second commodity feature representation into the multi-layer perceptron layer in the initial graph neural network model for metric learning to predict the probability that the label of the sample label commodity is an abnormal label; According to the real label of the sample label product and its predicted label result, the model parameters of the initial graph neural network model are updated to obtain the target graph neural network model.

6. The method according to claim 5, characterized in that According to the real label of the sample label product and its predicted label result, the model parameters of the initial graph neural network model are updated to obtain the target graph neural network model, including: Creating an initial loss function based on the real label of the sample label product and its predicted label result; Generate a regularization term according to the current model parameters of the initial graph neural network model, and generate a target loss function according to the initial loss function and the regularization term; According to the target loss function, the current model parameters of the initial graph neural network model are updated to obtain the target graph neural network model.

7. The method according to claim 4, characterized in that Inputting the risk graph and the attribute graph into the initial graph neural network model for graph structure learning to obtain a first commodity feature representation of the sample labeled commodity and a second commodity feature representation of the sample abnormal commodity, including: For any first sample product among the multiple sample products, determine at least one first connection edge corresponding to the first sample product from the risk graph and the attribute graph, each first connection edge connecting a node corresponding to the first sample product and a node corresponding to a neighboring product of the first sample product; Generate a vectorized representation to be spliced ​​according to the edge weight of at least one first connecting edge and the vectorized representation of the attribute feature of at least one corresponding neighboring product; Performing vector concatenation on the vectorized representation of the attribute feature of the first sample product and the vectorized representation to be concatenated, to obtain a product feature representation of the first sample product; Among them, when the first sample product is a sample labeled product, the product characteristic representation of the first sample product refers to the first product characteristic representation; when the first sample product is a sample abnormal product, the product characteristic representation of the first sample product refers to the second product characteristic representation.

8. The method according to claim 1, characterized in that Also includes: If the mean of the second attribute similarities is greater than the second similarity threshold, then at least one second attribute similarity between the second sample product and the third sample product in at least one attribute category is summed to obtain the sum of the second attribute similarities, and the sum of the second attribute similarities is used as the edge weight of the second edge.

9. The method according to any one of claims 2 to 8, characterized in that: Also includes: After each model training, for any fourth sample product among the multiple sample products, the feature similarity between the fourth sample product and each of the fifth sample products among the multiple sample products is calculated according to the product feature representation of the fourth sample product and the product feature representation of each of the fifth sample products among the multiple sample products; the fourth sample product and the fifth sample product are two different sample products among the multiple sample products; If the number of fifth sample products whose feature similarity with the fourth sample product is greater than the third similarity threshold is less than or equal to the set number, then creating an edge between the nodes corresponding to the fourth sample product and the fifth sample product in the attribute graph; If the number of fifth sample products whose feature similarity with the fourth sample product is greater than the third similarity threshold is greater than a set number, the edge between the nodes corresponding to the fourth sample product and the fifth sample product is deleted in the attribute graph.

10. A method for identifying abnormal commodities, characterized in that: include: Acquire multiple commodities, wherein the multiple commodities at least include commodities to be identified and abnormal commodities with abnormal labels; constructing a target risk graph reflecting the risk association relationship between the commodities according to the risk information of the plurality of commodities, and constructing a target attribute graph reflecting the attribute association relationship between the commodities according to the attribute characteristics of the plurality of commodities; Inputting the target risk graph and the target attribute graph into a target graph neural network model to identify abnormal commodities, so as to obtain an identification result of the commodity to be identified; The target graph neural network model is trained according to the training method according to any one of claims 1 to 9.

11. An electronic device, characterized in that: include: Memory and processor; The memory is used to store a computer program; the processor is coupled to the memory, and is used to execute the computer program to perform the steps in the method according to any one of claims 1 to 9, or to perform the steps in the method according to claim 10.

12. A computer storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 9, or the steps of the method according to claim 10.

Images