Commodity classification method and device based on artificial intelligence model

Abstract

The embodiment of the invention discloses a commodity classification method and device based on an artificial intelligence model, and the method comprises the steps: obtaining commodity data containing a to-be-classified commodity, and the commodity data comprises text data, image data and user behavior data; performing feature extraction and feature fusion on the commodity data to obtain a multi-modal feature vector; constructing a commodity association network, and updating the commodity association network in real time by using a dynamic network updating algorithm according to the commodity data to obtain commodity association information; constructing a semantic knowledge graph generation model, and generating a commodity classification candidate set by using the knowledge graph generation model according to the multi-modal feature vector and the commodity association information; and according to the commodity classification candidate set, using the commodity classification model to classify the to-be-classified commodity, and obtaining a classification result of the to-be-identified commodity. According to the commodity classification method, the accuracy, adaptability and real-time performance of commodity classification are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a commodity classification method and device based on an artificial intelligence model. Background Art

[0002] With the rapid development of e-commerce, there are many types of online goods. The accuracy and efficiency of product classification directly affect user experience and platform revenue. Traditional product classification methods based on manually formulated rules or simple keyword matching have been unable to cope with the efficient processing and accurate classification of massive product data. On the one hand, the randomness and diversity of information such as product names and descriptions make it difficult for manually formulated classification rules to cover them comprehensively; on the other hand, simple keyword matching is prone to a large number of misjudgments due to the limitations of semantic understanding, resulting in inaccurate product classification, which in turn affects the effectiveness of a series of business activities such as product recommendations, inventory management, and market analysis.

[0003] Therefore, there is an urgent need for a product classification method based on an intelligent model that can comprehensively utilize text, images, and user behavior data for multimodal feature fusion, dynamically update the product association network, and combine it with the knowledge graph model to achieve efficient and accurate product classification. Summary of the invention

[0004] The present application provides a commodity classification method and device based on an artificial intelligence model to improve the accuracy and efficiency of commodity classification.

[0005] This application provides the following solutions: According to a first aspect, a commodity classification method based on an artificial intelligence model is provided, the method comprising: obtaining commodity data containing commodities to be classified, the commodity data comprising: text data, image data and user behavior data; performing feature extraction and feature fusion on the commodity data to obtain a multimodal feature vector; constructing a commodity association network, and according to the commodity data, using a dynamic network update algorithm to perform real-time updates on the dynamic commodity association network to obtain commodity association information; constructing a semantic knowledge graph generation model, and using the knowledge graph generation model to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information; according to the commodity classification candidate set, using the commodity classification model to classify the commodities to be classified to obtain a classification result of the commodity to be identified.

[0006] According to an achievable method in an embodiment of the present application, the feature extraction and feature fusion of the commodity data to obtain a multimodal feature vector includes: inputting the text data into a neural network-based text classification model to perform semantic feature extraction to obtain a semantic feature vector; inputting the image data into a convolutional neural network model to perform image feature extraction to obtain an image feature vector; inputting the user behavior data into a recurrent neural network model to perform behavior feature extraction to obtain a user behavior feature vector; and performing feature fusion on the semantic feature vector, the image feature vector, and the user behavior feature vector to obtain a multimodal feature vector.

[0007] According to an achievable method in an embodiment of the present application, before performing feature extraction and feature fusion on the product data, the method also includes: preprocessing the product data, and the preprocessing includes one or more of the following: performing word segmentation, noise removal and / or word vectorization on the text data; performing cropping, scaling and / or normalization operations on the image data; deduplication, outlier filtering and / or conversion of user behavior data into a time series format.

[0008] According to an achievable method in an embodiment of the present application, the product association network is constructed, and according to the product data, the product association network is updated in real time using a dynamic network update algorithm to obtain product association information, including: taking products as nodes and the similarity or association between products as edges, and calculating the weights of the edges based on the similarity or association between products to construct a product association network; using a dynamic network update algorithm based on time series to update the edge weights and network structure of the product association network according to the product data; and extracting the corresponding product association information of the product to be classified according to the updated product association network.

[0009] According to an achievable method in an embodiment of the present application, the use of the knowledge graph generation model to generate a candidate set of commodity classification according to the multimodal feature vector and the commodity association information includes: the commodity association information is the neighbor node information of the commodity to be classified in the commodity association network; the neighbor node information and the multimodal feature vector are input into the knowledge graph generation model; the knowledge graph generation model uses a path sorting algorithm or a rule inference algorithm to perform classification reasoning to obtain the candidate set of commodity classification.

[0010] According to an achievable method in an embodiment of the present application, the commodity to be classified is classified according to the commodity classification candidate set using a commodity classification model to obtain a classification result of the commodity to be identified, including: the commodity classification model includes more than one basic classification model; the basic models classify the commodities to be classified according to the commodity classification candidate set, respectively, to obtain more than one basic classification result; the classification result of the commodity to be identified is obtained according to the weight of the basic classification model and the basic classification result.

[0011] According to an achievable method in an embodiment of the present application, the commodity classification model is obtained through pre-training, and the pre-training includes: obtaining training data including multiple training samples, the training samples including commodity data samples and classification result samples; performing feature extraction and feature fusion on the commodity data samples to obtain a multimodal feature vector; based on the commodity data samples, using a dynamic network update algorithm to perform real-time updating on the commodity association network to obtain commodity association information; using the knowledge graph generation model to generate a commodity classification candidate set based on the multimodal feature vector and the commodity association information; using the commodity classification model to classify the commodities to be classified based on the commodity classification candidate set to obtain the classification result of the commodity to be identified; wherein the training objectives include: minimizing the difference between the generated classification result and the classification result sample; and according to the training objectives, using reinforcement learning to optimize the classification strategy and adjust the parameters of the commodity classification model.

[0012] According to a second aspect, a commodity classification device based on an artificial intelligence model is provided, the device comprising: a commodity data acquisition unit, configured to acquire commodity data containing commodities to be classified, the commodity data comprising: text data, image data and user behavior data; a feature extraction and fusion unit, configured to perform feature extraction and feature fusion on the commodity data to obtain a multimodal feature vector; an association information acquisition unit, configured to construct a commodity association network, and according to the commodity data, to perform real-time update of the dynamic commodity association network using a dynamic network update algorithm to obtain commodity association information; a classification candidate set generation unit, configured to construct a semantic knowledge graph generation model, and to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information using the knowledge graph generation model; a classification result generation unit, configured to classify the commodities to be classified according to the commodity classification candidate set using a commodity classification model to obtain a classification result of the commodity to be identified.

[0013] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods described in the first aspect are implemented.

[0014] According to a fourth aspect, an electronic device is provided, comprising: one or more processors; and a memory associated with the one or more processors, the memory being used to store program instructions, and the program instructions, when read and executed by the one or more processors, execute the steps of any one of the methods described in the first aspect.

[0015] According to the specific embodiments provided in this application, this application discloses the following technical effects: (1) This application obtains multimodal feature vectors and product association information based on product data, obtains a product classification candidate set through a knowledge graph generation model based on the multimodal feature vectors and association information, and finally classifies the products to be classified using the product classification model. This application improves the accuracy, adaptability and real-time performance of product classification by combining multimodal information, knowledge graphs and product association networks to classify products.

[0016] (2) This application uses adaptive deep learning models (text classification model, convolutional neural network, recurrent neural network) to extract and fuse features for different types of product data (text, image, user behavior), which can fully explore the potential correlation between data and improve the accuracy and robustness of product classification.

[0017] (3) This application performs data preprocessing (such as text segmentation and noise removal, image normalization, and user behavior data outlier filtering) before feature extraction and fusion, which can effectively improve the quality of input data, reduce noise interference, and optimize the feature extraction effect of subsequent models.

[0018] (4) This application dynamically updates the product association network and adjusts the association between products through time series data to ensure that the network structure is always consistent with the latest product information, thereby improving the adaptability of the classification model and its ability to support scenarios with high real-time requirements.

[0019] (5) This application combines the knowledge graph generation model and the neighbor node information in the product association network, and generates a classification candidate set through a path sorting algorithm or rule reasoning, which can accurately limit the possible range of product classification, reduce the computational complexity of the classification model and improve the accuracy.

[0020] (6) This application performs classification by integrating multiple basic classification models, and uses the weights of the classification results to optimize the final classification results, thereby further improving the reliability and generalization ability of the classification model.

[0021] (7) This application constructs a classification model through a pre-training method and combines it with reinforcement learning to optimize the classification strategy, so that the model can adaptively adjust parameters, thereby improving the efficiency of the product classification model and the accuracy of the classification results.

[0022] Of course, any product implementing the present application does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A flowchart of a commodity classification method based on artificial intelligence provided in an embodiment of the present application; Figure 2 A structural block diagram of a commodity classification device based on artificial intelligence provided in an embodiment of the present application; Figure 3 A schematic block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.

[0026] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0027] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0028] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0029] At present, there are some existing technologies. These commodity classification methods mostly rely on single-mode data (such as text or image data) for analysis and cannot effectively combine multi-modal information, resulting in insufficient accuracy of classification results. In addition, traditional commodity classification methods are difficult to dynamically adapt to real-time changes in commodity information, and cannot fully explore the association information between commodities, thus limiting the intelligence and accuracy of the classification model.

[0030] In view of this, the present application provides a new idea. Figure 1 A flowchart of a commodity classification method based on artificial intelligence provided in an embodiment of the present application. Figure 1 As shown in , the method may include the following steps: Step 101: Obtain commodity data containing commodities to be classified, the commodity data including text data, image data and user behavior data.

[0031] Step 102: extracting and fusing features of the commodity data to obtain a multimodal feature vector; Step 103: construct a product association network, and update the product association network in real time according to the product data using a dynamic network update algorithm to obtain product association information.

[0032] Step 104: Construct a semantic knowledge graph generation model, and use the knowledge graph generation model to generate a candidate set of product classifications based on multimodal feature vectors and product association information.

[0033] Step 105: Based on the commodity classification candidate set, the commodity classification model is used to classify the commodities to be classified, and the classification result of the commodities to be identified is obtained.

[0034] The present application obtains multimodal feature vectors and product association information based on product data, obtains a product classification candidate set through a knowledge graph generation model based on the multimodal feature vectors and the association information, and finally classifies the products to be classified using the product classification model. The present application improves the accuracy, adaptability and real-time performance of product classification by combining multimodal information, knowledge graphs and product association networks to classify products.

[0035] First, the above step 101, namely, "obtaining commodity data containing commodities to be classified, the commodity data including: text data, image data and user behavior data" is described in detail in conjunction with the embodiment.

[0036] This application classifies the commodities to be classified according to their commodity data. The commodity data can be obtained from a variety of data sources, including e-commerce platform databases, enterprise commodity management systems, commodity information obtained by web crawlers, etc. The collected data covers multi-dimensional content such as commodity name, detailed description, image link, price, sales quantity, user evaluation, brand information, category information, etc. Specifically, commodity data includes text data, image data, and user behavior data. These three types of data provide support for commodity classification from different dimensions.

[0037] Text data refers to product data in text format, such as the product name, description, and specifications, which can directly reflect the basic information and attributes of the product. Taking an e-commerce platform as an example, the text data of a dress may include the name "New Summer Dress", the description "100% cotton, V-neck design, suitable for daily wear", and related specifications such as "Size: M".

[0038] Image data refers to pictures used to describe products. These data further enrich product information through visual features such as color, shape, material, etc. For example, the image of this dress may show that its color is light blue, the material is cotton, and the design is V-neck, and provide users with intuitive feelings.

[0039] User behavior data is data related to the user's behavior towards the product, such as the user's browsing history, purchase history, and evaluation content, which can indirectly reflect the category of the product and the user's preference. For example, if the user buys other daily casual clothing after purchasing this dress, these behavior data can be used to strengthen the classification model's judgment that the product is classified as "daily clothing".

[0040] In order to improve the accuracy of commodity classification, this application can preprocess commodity data. The specific preprocessing method is as follows: For text data, which may contain noise or redundant information, it is necessary to convert it into semantic features that can be recognized by the model through word segmentation, noise removal and word vectorization. For example, for the product description "New summer dress 100% cotton", after word segmentation, the keywords "summer, new style, dress, cotton" can be extracted, and irrelevant information such as "100%" can be removed. Subsequently, these keywords are converted into vector representations that can be used for calculations through word vectorization technology (such as Word2Vec or BERT).

[0041] For image data, the image data usually needs to be in a unified format and size to facilitate input into the model for analysis. The preprocessing steps include cropping, scaling, and normalization operations. For example, for a dress in a product image, cropping can remove the redundant part of the background to make the subject more prominent; scaling adjusts the image to a fixed size (such as 224x224 pixels); and normalization can normalize the pixel values ​​to between 0 and 1, reducing data distribution differences and improving the robustness of the model.

[0042] For user behavior data, there may be duplicate records, outliers, or inconsistent formats. Preprocessing includes deduplication, filtering outliers, and conversion to time series format. For example, for a user's purchase of a dress, if the same product is recorded multiple times, deduplication is required; outlier filtering can remove unreasonable data caused by misoperation (such as records of negative purchase quantities); after conversion to time series format, the changing trend of user behavior over time can be recorded, making it easier for the model to capture behavior patterns.

[0043] The above step 102, namely "extracting features and fusing features on commodity data to obtain a multimodal feature vector", is described in detail below in conjunction with an embodiment.

[0044] This application extracts multiple types of features from product data and fuses them into a unified multimodal feature vector for subsequent model processing.

[0045] For commodity data, multiple methods can be used for feature extraction. For example, for text data, traditional text feature extraction methods can be used to calculate the importance weights of words and represent text in the form of sparse vectors; or more advanced pre-trained language models (such as RoBERTa, GPT, etc.) can be used to obtain higher-quality semantic features. For image data, pre-trained deep convolutional networks can be used: such as Inception, VGG, EfficientNet, etc., to extract image features of different scales and complexities; or global feature extraction of images can be performed through the self-attention mechanism. For user behavior feature data, the Transformer model can be used to capture the long-term dependencies between behaviors through the global self-attention mechanism when processing behavior sequences; or deep reinforcement learning can be used to model the long-term preferences of user behaviors and use behavior trajectories to predict future decision-making tendencies.

[0046] As a preferred implementation, the present application performs feature extraction and feature fusion on the product data to obtain a multimodal feature vector, including: inputting the text data into a neural network-based text classification model to perform semantic feature extraction to obtain a semantic feature vector; inputting the image data into a convolutional neural network model to perform image feature extraction to obtain an image feature vector; inputting the user behavior data into a recurrent neural network model to perform behavior feature extraction to obtain a user behavior feature vector; and performing feature fusion on the semantic feature vector, the image feature vector, and the user behavior feature vector to obtain a multimodal feature vector.

[0047] Specifically, text data is usually extracted through semantic features through a neural network-based text classification model, such as BERT or Long Short-Term Memory (LSTM), which converts natural language descriptions into numerical semantic feature vectors. For example, after preprocessing, the product description "simple style men's business leather shoes" can be input into the BERT model to extract a vector representing the description, which contains key information such as the product's style (simple), target users (men), and purpose (business).

[0048] Image data is extracted through a convolutional neural network (CNN) to capture the appearance characteristics of the product. For example, after a picture of a pair of business leather shoes is input into the ResNet model, vector representations describing the color (black), shape (pointed toe), material (leather texture) and other visual features of the shoes can be extracted.

[0049] The user behavior data is extracted through the Recurrent Neural Network (RNN) to extract behavioral features. Specifically, models such as LSTM or Gated Recurrent Unit (GRU) can be used to capture the time series rules of user behavior. For example, if a user's recent browsing history includes products such as "men's leather shoes" and "business shirts", the LSTM model can be used to extract its behavioral feature vector, reflecting the user's preference for business men's products.

[0050] The text feature vector, image feature vector and user behavior feature vector extracted above are fused to generate a unified multimodal feature vector. The fusion method can use simple vector concatenation or more complex attention mechanism to enhance the information complementarity between different modalities. For example, for the commodity of business leather shoes, its multimodal feature vector can contain semantic information ("business men"), visual features ("black pointed leather shoes"), and behavioral preferences ("users pay attention to business products") to form a more comprehensive commodity feature representation. The fusion method can also be used to assign different weights to different modal feature vectors and then perform weighted summation to emphasize the importance of certain modalities; or different modal features can be sequentially input and fused through a recursive neural network to capture the hierarchical relationship between modalities, etc.

[0051] The above step 103, namely "constructing a product association network, and updating the product association network in real time using a dynamic network update algorithm according to the product data to obtain product association information" is described in detail below in conjunction with an embodiment.

[0052] This application can capture the interactive characteristics and potential connections of commodities in actual scenarios by establishing associations between commodities, thereby providing strong support for classification. Specifically, the commodity association network represents commodities as nodes in a graph structure, and the edges between nodes represent the similarity or association of commodities. For example, in an e-commerce platform, the association between two commodities can be calculated by the user's common purchasing behavior, the similarity of commodity descriptions, the proximity of image features, etc. The weight of each edge reflects the strength of this association, such as quantified by cosine similarity, Pearson correlation coefficient, or feature embedding model based on deep learning.

[0053] In order to ensure the real-time and dynamic nature of the network, a dynamic network update algorithm is introduced. The dynamic network update algorithm can adjust the network structure and edge weights in real time based on the newly added products and user behavior information in the product data. For example, when a new product is launched, by analyzing the semantic similarity between its description text and other products, it can be added as a new node to the product association network, and edge weights can be assigned based on the similarity. Similarly, when some products are no longer active (such as being removed from the shelves or not being visited for a long time), their edge weights can be reduced or the corresponding nodes can be removed to ensure timely updates of the network.

[0054] For example, suppose a user shows great interest in a smartphone and a phone case while browsing an e-commerce platform, and makes a purchase. The system can update the product association network based on this behavior information, adding an edge between the "smartphone" node and the "phone case" node, and the edge weight reflects the strength of the association between multiple purchases. In addition, if a brand releases a new headset, based on the similarity calculation between its feature description and the existing headset, the new headset node will automatically connect to the existing headset node and assign the corresponding edge weight.

[0055] As an implementable manner, the present application constructs a product association network, and based on the product data, uses a dynamic network update algorithm to perform real-time updates on the dynamic product association network, and when obtaining product association information, includes: taking products as nodes, and the similarity or association between products as edges, and calculating the weights of the edges based on the similarity or association between products to construct a product association network; using a dynamic network update algorithm based on time series to update the edge weights and network structure of the product association network based on the product data; and extracting the corresponding product association information of the product to be classified based on the updated product association network.

[0056] Specifically, in actual operation, the weight of the edge is determined by calculating the similarity or association between the products. For example, the similarity between two products can be evaluated based on the text description, image features, or user behavior data of the products. If there are a large number of similar keywords in the description texts of two products, or they are often selected at the same time in user purchase records, the edge weight between them will be high.

[0057] The dynamic network update algorithm is the core of this method, which is used to update the network structure according to the real-time changing commodity data. For example, when a new commodity is added to the platform or the relevance of existing commodities changes in the commodity information, the algorithm can dynamically adjust the edge weights or network structure between commodities according to the time series data. This ensures the real-time and accuracy of the commodity association network. Time series data includes the interaction behaviors between commodities recorded over time, such as user browsing, purchasing, evaluation, search and other data. For example: purchase behavior data: the frequency of users purchasing commodities A and B at the same time in a certain period of time; browsing behavior data: the time interval and frequency of users browsing commodities C and D successively; evaluation correlation data: the distribution of common evaluation keywords between commodities over time. Time series data can be obtained through user behavior data in commodity data, and the time series formed by these data records how the interaction intensity between commodities changes over time.

[0058] Driven by time series data, the network structure itself may also change, including adding edges, deleting edges, or adjusting the connection relationship between nodes: Adding edges: When a product (such as a new product) interacts significantly with other products (such as users purchasing them at the same time), but there is no corresponding edge connection in the current network, edges will be dynamically added according to the time series data; Deleting edges: When the frequency of interaction between products continues to decrease (such as a product is removed from the shelves or is no longer popular), its edge weight will gradually decrease to zero, and eventually the edge will be removed from the network; Adjusting the structure: When new types of products or user behavior patterns emerge, the network structure may need to be reorganized, such as re-clustering products or updating the centrality of nodes.

[0059] For example, on an e-commerce platform, the purchase frequency of product A (smart watch) and product B (watch strap) increased significantly over a period of time, while the sales frequency of product C (old smart watch) decreased. In this case, the dynamic network update algorithm adjusts the product association network based on the time series data. First, the edge weights are adjusted. The time series data shows that the co-purchase rate of product A and product B increased from 10% to 30%. The system calculates and updates the edge weight between A and B by using methods such as weighted moving average, which significantly increases the edge weight between A and B, thereby reflecting the enhanced association between the two products. Second, new edges may be added to the network structure. For example, the newly launched product D (smart watch charger) begins to have frequent purchase or browsing associations with products A and B. After the system detects this trend by analyzing the time series data, it dynamically adds edges between product D and product A, and between product D and product B to timely incorporate new product relationships. Finally, for products that are no longer active, the system gradually removes their associated edges. If product C is removed from the shelves or its sales popularity decreases, the frequency of interaction between it and other products will gradually decrease to zero. The system will reduce its edge weight accordingly and eventually remove the relevant edge from the product association network.

[0060] The above step 104, i.e., "constructing a semantic knowledge graph generation model, and using the knowledge graph generation model to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information" is described in detail below in conjunction with an embodiment.

[0061] The key step in this method is to build a semantic knowledge graph generation model, which is used to integrate multimodal feature vectors and product association information into a unified graph framework to generate a candidate set of product classifications. Specifically, the model extracts and represents the complex relationship between products and categories from product data through semantic reasoning and association mining, and finally provides candidate classification results for the products to be classified.

[0062] First, a semantic knowledge graph generation model is constructed. During the construction process, the product knowledge from multiple data sources is integrated to provide semantic support for product classification. These data sources include but are not limited to industry standard documents, product encyclopedia websites, and internal enterprise knowledge bases. Specifically, product entities, product attributes, and semantic relationships between product entities (such as "is a certain type of product" (is_a) relationship, "has a certain attribute" (has_attribute) relationship, and "associated with a certain entity" (related_to) relationship) are extracted from the above data. After the extraction is completed, graph database technology (such as Neo4j or JanusGraph) is used to store and manage the knowledge graph. These technologies can effectively support efficient query and operation of nodes and edges in the graph, and provide a solid foundation for subsequent semantic reasoning. For example, through the graph database, you can quickly retrieve the category or attribute of a certain product, or find other related products, thereby providing comprehensive semantic information support for generating a candidate set of product classification.

[0063] In terms of specific implementation, the semantic knowledge graph generation model is completed through the following steps: First, a preliminary association between products and categories is established based on multimodal feature vectors (including text features, image features, and user behavior features). For example, for a product A (smart watch), its text description contains the keywords "sports" and "health monitoring", the image features show a round dial, and the user behavior shows that most of its buyers are fitness enthusiasts. This information together constitutes the semantic representation of product A.

[0064] Next, combined with the product association information, the potential categories of product A are further expanded. For example, through the product association network, the neighbor node of product A, product B (smart bracelet), is found to have a high similarity with product A, and product B has been classified as "smart wearable device". Using this information, the knowledge graph generation model infers that product A may also belong to this category.

[0065] Finally, the model uses a path sorting algorithm or a rule inference algorithm to infer and sort the possible classifications of product A to form a candidate set of product classifications. For example, the model may generate the following candidate sets and their corresponding confidence scores for product A: "smart wearable devices" (85%), "electronic devices" (70%), and "sports accessories" (60%). These candidate classifications provide the basis for the subsequent final classification.

[0066] As an implementable method, when using the knowledge graph generation model to generate a product classification candidate set based on the multimodal feature vector and the product association information, the following method can be adopted: the product association information is the neighbor node information of the product to be classified in the product association network; the neighbor node information and the multimodal feature vector are input into the knowledge graph generation model; the knowledge graph generation model uses a path sorting algorithm or a rule inference algorithm to perform classification reasoning to obtain the product classification candidate set.

[0067] Specifically, the commodity association information refers to the neighbor node information of the commodity to be classified in the commodity association network. These neighbor node information reflects the characteristics of other commodities that are highly related to the commodity to be classified in terms of attributes, behaviors or semantics, such as commodities with similar purchase history or commodities with similar keywords in description. In operation, the commodity association information and the multimodal feature vector are input into the knowledge graph generation model together. The model uses the entity relationships and semantic rules constructed in the commodity semantic knowledge graph to perform classification reasoning through the path sorting algorithm or the rule reasoning algorithm. For example, the path sorting algorithm can calculate the optimal classification path based on the path weight between the commodity and the classification node, while the rule reasoning algorithm makes inferences based on the classification rules predefined in the knowledge graph. Taking the actual application as an example, if the commodity to be classified is a "wireless noise reduction headset", the system will extract its multimodal feature vector, such as the keywords "wireless" and "noise reduction" in the commodity name, the design information of the headset in the image features, and the user's purchasing behavior characteristics. Combined with the neighbor node information in the commodity association network, such as "Bluetooth headset" or "smart noise reduction device", this information is input into the knowledge graph generation model. Through the path sorting algorithm, the model may find that the path weight between "wireless noise-cancelling headphones" and the classification node "headphones / audio equipment" is the highest, thereby inferring that the product belongs to the "headphones / audio equipment" category.

[0068] The above step 105, namely "classifying the commodities to be classified by using the commodity classification model according to the commodity classification candidate set to obtain the classification result of the commodity to be identified" is described in detail below in conjunction with an embodiment.

[0069] The step of using the product classification model to classify the products to be classified based on the product classification candidate set is the key to achieving accurate classification. This process integrates the classification possibilities in the candidate set and combines the decision-making ability of the product classification model to finally generate the classification result.

[0070] Specifically, the product classification model can be obtained through training of classification models commonly used in the field of artificial intelligence, such as classification models based on support vector machines (SVM), random forests, deep neural networks (DNN) or ensemble learning.

[0071] As a preferred embodiment, the commodity classification model includes more than one basic classification model; the basic models classify the commodities to be classified according to the commodity classification candidate set to obtain more than one basic classification results; and the classification results of the commodities to be identified are obtained according to the weights of the basic classification models and the basic classification results.

[0072] Specifically, the basic classification model can be based on a variety of classification models such as support vector machine (SVM), random forest and deep neural network (DNN). These models will evaluate the products to be classified for each candidate category in the product classification candidate set, calculate the matching degree between each category and the product to be classified, and generate the corresponding basic classification results. Subsequently, the system will perform weighted fusion on these basic classification results, for example, by assigning weights according to the historical performance of each model, and finally determine the classification results of the products to be identified.

[0073] For example, suppose the product to be classified is a "smart constant temperature electric kettle", and its product classification candidate set may include "household appliances", "smart devices" and "kitchen supplies". The DNN in the product classification model may give a higher matching degree to "smart devices" based on text descriptions (such as "smart constant temperature"); SVM infers that "kitchen supplies" is the optimal classification based on user behavior data (such as the history of common purchases with other kitchen supplies); random forest combines the image features of the product (such as the shape and material of the "electric kettle") to judge that "household appliances" is the most likely. Finally, through the weight fusion strategy, the system concludes that the product best fits the classification of "kitchen supplies".

[0074] The commodity classification model of the present application is obtained through pre-training, and the pre-training includes: obtaining training data including multiple training samples, wherein the training samples include commodity data samples and classification result samples; performing feature extraction and feature fusion on the commodity data samples to obtain a multimodal feature vector; based on the commodity data samples, using a dynamic network update algorithm to perform real-time updating on the commodity association network to obtain commodity association information; using the knowledge graph generation model to generate a commodity classification candidate set based on the multimodal feature vector and the commodity association information; using the commodity classification model to classify the commodities to be classified based on the commodity classification candidate set to obtain the classification result of the commodity to be identified; wherein the training objectives include: minimizing the difference between the generated classification result and the classification result sample; and according to the training objectives, using reinforcement learning to optimize the classification strategy and adjust the parameters of the commodity classification model.

[0075] Specifically, first, pre-training requires obtaining large-scale training samples, which include product data samples (such as product text descriptions, image features, and user behavior data) and corresponding classification result samples (i.e., known classifications of each product). These data form the basis for model learning. By extracting and fusing features from product data samples, the system integrates text features, image features, and user behavior features into a multimodal feature vector, providing a unified input expression for subsequent classification.

[0076] Next, the dynamic network update algorithm is used to update the product association network in real time. During the pre-training process, the product association network regards products as nodes and the similarities or associations between products as edges, and dynamically adjusts the edge weights and network structure through time series data. For example, when certain products in the training samples frequently appear in similar purchase behaviors, the algorithm will increase the edge weights between these product nodes to more accurately reflect their association relationship.

[0077] Subsequently, the knowledge graph generation model is used to generate a candidate set of commodity classifications. By inputting the multimodal feature vector and commodity association information into the knowledge graph model, the model can combine the rule inference algorithm or the path sorting algorithm to generate the most relevant classification candidate set for the commodity to be classified. For example, for a "4K smart TV" in the training data, the knowledge graph model may generate its candidate classification set as "household appliances", "smart devices" and "display devices".

[0078] Finally, the product classification model uses the classification candidate set to classify the products in the training samples. The goal of pre-training is to minimize the error between the classification result and the classification result sample, for example, by measuring the accuracy of the classification through a loss function (such as cross entropy loss). In order to further improve the classification performance, reinforcement learning methods (such as deep Q network or proximal policy optimization algorithm) are introduced in the pre-training process to optimize the classification strategy. Reinforcement learning enables the model to dynamically adjust parameters to cope with changes in data distribution and diversified characteristics in product classification by continuously optimizing the classification strategy. Specifically, after each classification decision, the system generates a reward signal based on the accuracy and effect of the classification result, and uses the reward signal to update the policy network of the agent. Through multiple rounds of iterative training, the agent can gradually learn and optimize the classification strategy to achieve the optimal classification effect. This process not only improves the model's adaptability to complex product classification scenarios, but also effectively improves the accuracy and reliability of the classification results.

[0079] At the same time, during the training process, cross-validation, model evaluation indicators (such as accuracy, recall rate, F1 value, etc.) and integrated learning strategies (such as weighted voting, model stacking, etc.) can be used to optimize and adjust the model, determine the weights and parameters of each basic model, so that the model can accurately classify products.

[0080] The above method provided in the embodiment of the present application can be applied to a variety of application scenarios, including but not limited to: deploying the commodity classification method to run in an actual business environment. Receive the input of the commodity data to be classified, first perform data preprocessing and multimodal feature extraction, then use the classification model to perform classification and recognition, and finally output the classification results of the commodity. Regularly evaluate and monitor the performance of the model, adjust the model parameters, update the knowledge graph or optimize the classification strategy according to the evaluation results to ensure the continuous and stable operation of the system and the continuous improvement of classification performance.

[0081] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0082] According to an embodiment of another aspect, a commodity classification device based on an artificial intelligence model is provided. Figure 2 A schematic block diagram of a commodity classification device based on an artificial intelligence model according to an embodiment is shown, Figure 2 As shown, the device 200 includes: The commodity data acquisition unit 201 is configured to acquire commodity data containing commodities to be classified, wherein the commodity data includes text data, image data and user behavior data.

[0083] The feature extraction and fusion unit 202 is configured to perform feature extraction and feature fusion on the commodity data to obtain a multimodal feature vector.

[0084] The associated information acquisition unit 203 is configured to construct a commodity associated network, and update the commodity associated network in real time according to the commodity data using a dynamic network update algorithm to obtain commodity associated information.

[0085] The classification candidate set generation unit 204 is configured to construct a semantic knowledge graph generation model, and use the knowledge graph generation model to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information.

[0086] The classification result generating unit 205 is configured to classify the commodities to be classified according to the commodity classification candidate set by using a commodity classification model to obtain a classification result of the commodities to be identified.

[0087] As an implementable manner, when the feature extraction and fusion unit 202 performs feature extraction and feature fusion on the product data to obtain a multimodal feature vector, it can be configured as follows: inputting the text data into a neural network-based text classification model to perform semantic feature extraction to obtain a semantic feature vector; inputting the image data into a convolutional neural network model to perform image feature extraction to obtain an image feature vector; inputting the user behavior data into a recurrent neural network model to perform behavior feature extraction to obtain a user behavior feature vector; and performing feature fusion on the semantic feature vector, the image feature vector, and the user behavior feature vector to obtain a multimodal feature vector.

[0088] As an implementable manner, before extracting and fusing features on the commodity data, the feature extraction and fusion unit 202 can be configured to: pre-process the commodity data, wherein the pre-processing includes one or more of the following: performing word segmentation, noise removal and / or word vectorization on the text data; performing cropping, scaling and / or normalization operations on the image data; performing deduplication, outlier filtering and / or converting the user behavior data into a time series format; As an implementable manner, the association information acquisition unit 203 constructs a product association network and, based on the product data, uses a dynamic network update algorithm to perform real-time updates on the product association network. When obtaining product association information, the configuration may be as follows: taking products as nodes, and the similarity or association between products as edges, and calculating the weights of the edges based on the similarity or association between products to construct a product association network; using a dynamic network update algorithm based on a time series to update the edge weights and network structure of the product association network based on the product data; and extracting the corresponding product association information of the product to be classified based on the updated product association network.

[0089] As an implementable manner, when the classification candidate set generation unit 204 generates a commodity classification candidate set based on the multimodal feature vector and the commodity association information using the knowledge graph generation model, it can be configured as follows: the commodity association information is the neighbor node information of the commodity to be classified in the commodity association network; the neighbor node information and the multimodal feature vector are input into the knowledge graph generation model; the knowledge graph generation model uses a path sorting algorithm or a rule inference algorithm to perform classification reasoning to obtain the commodity classification candidate set.

[0090] As an implementable manner, the classification result generation unit 205 can be configured as follows when classifying the commodities to be classified according to the commodity classification candidate set using the commodity classification model to obtain the classification result of the commodities to be identified: the commodity classification model includes more than one basic classification model; the basic models classify the commodities to be classified according to the commodity classification candidate set respectively to obtain more than one basic classification results; and the classification result of the commodities to be identified is obtained according to the weight of the basic classification model and the basic classification result.

[0091] As an implementable manner, the commodity classification model is obtained through pre-training, and the classification result generating unit 205 can be configured during pre-training to: obtain training data including a plurality of training samples, wherein the training samples include commodity data samples and classification result samples; perform feature extraction and feature fusion on the commodity data samples to obtain a multimodal feature vector; According to the commodity data sample, the commodity association network is updated in real time using a dynamic network update algorithm to obtain commodity association information; the knowledge graph generation model is used to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information; the commodity classification model is used to classify the commodities to be classified according to the commodity classification candidate set to obtain the classification result of the commodity to be identified; wherein the training objectives include: minimizing the difference between the generated classification result and the classification result sample; according to the training objectives, the classification strategy is optimized using reinforcement learning to adjust the parameters of the commodity classification model.

[0092] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0093] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0094] In addition, an embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods in the aforementioned method embodiments are implemented.

[0095] And an electronic device, comprising: one or more processors; and a memory associated with the one or more processors, the memory being used to store program instructions, and the program instructions, when read and executed by the one or more processors, execute the steps of any of the methods described in the aforementioned method embodiments.

[0096] The present application also provides a computer program product, including a computer program, which implements the steps of any one of the methods in the aforementioned method embodiments when executed by a processor.

[0097] in, Figure 3 The architecture of the electronic device is shown as an example, which may include a processor 310, a video display adapter 311, a disk drive 312, an input / output interface 313, a network interface 314, and a memory 320. The processor 310, the video display adapter 311, the disk drive 312, the input / output interface 313, the network interface 314, and the memory 320 may be communicatively connected via a communication bus 330.

[0098] The processor 310 may be implemented by a general-purpose CPU, a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in this application.

[0099] The memory 320 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 320 can store an operating system 321 for controlling the operation of the electronic device 300, and a basic input and output system (BIOS) 322 for controlling the low-level operation of the electronic device 300. In addition, a web browser 323, a data storage management system 324, and a commodity classification device 325 based on an artificial intelligence model, etc. can also be stored. The above-mentioned commodity classification device 325 based on an artificial intelligence model can be an application program that specifically implements the operations of the aforementioned steps in the embodiment of the present application. In short, when the technical solution provided in the present application is implemented by software or firmware, the relevant program code is stored in the memory 320 and is called and executed by the processor 310.

[0100] The input / output interface 313 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0101] The network interface 314 is used to connect to a communication module (not shown) to achieve communication interaction between the device and other devices. The communication module can achieve communication through a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, WIFI, Bluetooth, etc.).

[0102] The bus 330 comprises a pathway for transmitting information between the various components of the device (eg, the processor 310 , the video display adapter 311 , the disk drive 312 , the input / output interface 313 , the network interface 314 , and the memory 320 ).

[0103] It should be noted that, although the above device only shows a processor 310, a video display adapter 311, a disk drive 312, an input / output interface 313, a network interface 314, a memory 320, a bus 330, etc., in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include components necessary for implementing the solution of the present application, and does not necessarily include all the components shown in the figure.

[0104] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly contributed to the prior art in the form of a computer program product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0105] The technical solution provided by the present application is described in detail above. The principle and implementation method of the present application are described in detail using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for those skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present application.

Claims

1. A commodity classification method based on an artificial intelligence model, characterized in that: The method comprises: Acquire commodity data containing commodities to be classified, wherein the commodity data includes: text data, image data and user behavior data; Performing feature extraction and feature fusion on the commodity data to obtain a multimodal feature vector; Constructing a commodity association network, and updating the commodity association network in real time using a dynamic network update algorithm according to the commodity data to obtain commodity association information; Constructing a semantic knowledge graph generation model, and using the knowledge graph generation model to generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information; According to the commodity classification candidate set, the commodities to be classified are classified using a commodity classification model to obtain a classification result of the commodities to be identified.

2. According to the method of claim 1, the step of extracting and fusing features of the commodity data to obtain a multimodal feature vector comprises: Inputting the text data into a text classification model based on a neural network to extract semantic features and obtain a semantic feature vector; Inputting the image data into a convolutional neural network model to extract image features and obtain an image feature vector; Inputting the user behavior data into a recurrent neural network model to extract behavior features and obtain a user behavior feature vector; The semantic feature vector, the image feature vector and the user's behavior feature vector are subjected to feature fusion to obtain a multimodal feature vector.

3. According to the method of claim 1, before extracting and fusing features of the commodity data, the method further comprises: Preprocessing the product data includes one or more of the following: Performing word segmentation, noise removal and / or word vectorization processing on the text data; Performing cropping, scaling and / or normalization operations on the image data; The user behavior data is deduplicated, outlier filtered, and / or converted into a time series format.

4. According to the method of claim 1, the step of constructing a product association network, updating the product association network in real time based on the product data using a dynamic network update algorithm, and obtaining product association information comprises: The products are taken as nodes, and the similarity or association between products is taken as edges. The weight of the edges is calculated based on the similarity or association between products to build a product association network. Using a dynamic network update algorithm based on time series, the edge weights and network structure of the commodity association network are updated according to the commodity data; According to the updated commodity association network, the corresponding commodity association information of the commodity to be classified is extracted.

5. According to the method of claim 1, the step of using the knowledge graph generation model to generate a candidate set of commodity classifications according to the multimodal feature vector and the commodity association information comprises: The commodity association information is the neighbor node information of the commodity to be classified in the commodity association network; Inputting the neighbor node information and the multimodal feature vector into the knowledge graph generation model; The knowledge graph generation model uses a path sorting algorithm or a rule reasoning algorithm to perform classification reasoning to obtain the commodity classification candidate set.

6. According to the method of claim 1, the step of classifying the commodities to be classified using a commodity classification model according to the commodity classification candidate set to obtain the classification result of the commodities to be identified comprises: The commodity classification model includes more than one basic classification model; The basic model classifies the commodities to be classified according to the commodity classification candidate set to obtain one or more basic classification results; The classification result of the commodity to be identified is obtained according to the weight of the basic classification model and the basic classification result.

7. According to the method of claim 1, the commodity classification model is obtained by pre-training, and the pre-training includes: Acquire training data including a plurality of training samples, wherein the training samples include commodity data samples and classification result samples; Performing feature extraction and feature fusion on the commodity data sample to obtain a multimodal feature vector; According to the commodity data sample, the commodity association network is updated in real time using a dynamic network update algorithm to obtain commodity association information; Generate a candidate set of commodity classifications according to the multimodal feature vector and the commodity association information using the knowledge graph generation model; Using the commodity classification model, classify the commodities to be classified according to the commodity classification candidate set to obtain a classification result of the commodities to be identified; The training objectives include: minimizing the difference between the generated classification result and the classification result sample; and optimizing the classification strategy by using reinforcement learning according to the training objectives to adjust the parameters of the commodity classification model.

8. A commodity classification device based on an artificial intelligence model, characterized in that: The device comprises: A commodity data acquisition unit is configured to acquire commodity data containing commodities to be classified, wherein the commodity data includes: text data, image data and user behavior data; A feature extraction and fusion unit is configured to perform feature extraction and feature fusion on the commodity data to obtain a multimodal feature vector; The associated information acquisition unit is configured to construct a commodity associated network, and update the commodity associated network in real time according to the commodity data by using a dynamic network update algorithm to obtain commodity associated information; A classification candidate set generation unit is configured to construct a semantic knowledge graph generation model, and generate a commodity classification candidate set according to the multimodal feature vector and the commodity association information by using the knowledge graph generation model; The classification result generating unit is configured to classify the commodities to be classified according to the commodity classification candidate set by using a commodity classification model to obtain a classification result of the commodities to be identified.

9. An electronic device, characterized in that: include: one or more processors; And a memory associated with the one or more processors, the memory is used to store program instructions, and when the program instructions are read and executed by the one or more processors, the steps of the method described in any one of claims 1 to 7 are executed.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images