Cross-border e-commerce data analysis method and device

Abstract

The invention discloses a cross-border e-commerce data analysis method and device, and relates to the technical field of data analysis. Obtaining first transaction data of the target commodity, determining a congeneric commodity of the target commodity, and obtaining second transaction data of the congeneric commodity; fusing the first transaction data and the second transaction data to obtain third transaction data, and preprocessing the third transaction data to obtain a target data set; determining a hidden demand of the user according to the target data set, determining an alternative commodity according to the hidden demand, and obtaining comments of the alternative commodity to obtain a comment data set; performing language meaning detection on the target data set and the comment data set to obtain an analysis result; the alternative commodity comments are analyzed and determined in combination with the hidden demands of the user, the richness of the data volume is increased, the relevance between the commodity data and the user data is effectively enhanced, and the semantic analysis result of the comments is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis, and specifically relates to a cross-border e-commerce data analysis method and device. Background Art

[0002] With the continuous development of cross-border e-commerce business, analyzing consumer reviews has become a key strategy for merchants to optimize products. By deeply mining review data, merchants can identify the advantages and disadvantages of products and understand the specific needs and preferences of consumers. By using natural language processing technology, key information and emotional tendencies in reviews can be extracted to identify the direction of product improvement.

[0003] Merchants use natural language processing technology to conduct sentiment analysis on product reviews. By deeply analyzing the positive and negative feedback in the reviews, merchants can capture consumers' opinions on product satisfaction, preferences, and dissatisfaction, and improve products based on user experience.

[0004] There are still some problems in the existing technology. For example, the correlation between product data and user data is weak, the amount of product review data is not rich enough, and the accuracy of review language analysis is poor, resulting in the inability of merchants to obtain effective improvement suggestions when analyzing reviews. Summary of the invention

[0005] The purpose of the present invention is to solve the problems that the correlation between product data and user data is weak, the data volume of product reviews is not rich enough, and the accuracy of review language analysis is poor, resulting in merchants being unable to obtain effective improvement suggestions when analyzing reviews, and to propose a cross-border e-commerce data analysis method and device.

[0006] In a first aspect of the implementation of the present invention, a cross-border e-commerce data analysis method is first proposed, the method comprising: Acquire first transaction data of a target product, determine a product of the same category as the target product, and acquire second transaction data of the product of the same category; the transaction data includes: product information and user information; a product of the same category is a product that replaces the target product in all functional tags, and both the target product and the product of the same category contain multiple functional tags; fusing the first transaction data and the second transaction data to obtain third transaction data, and preprocessing the third transaction data to obtain a target data set; Determine the user's implicit needs according to the target data set, determine the substitute product according to the implicit needs, and obtain comments of the substitute product to obtain a comment data set; the substitute product is a product that replaces the target product in part of the functional labels, and the substitute product includes a functional label; Language meaning detection is performed on the target data set and the comment data set to obtain analysis results.

[0007] Optionally, preprocessing the third transaction data to obtain a target data set includes: Preprocessing the third transaction data to obtain target data, constructing a target heterogeneous graph according to the target data, and converting the target heterogeneous graph into a two-dimensional adjacency matrix; Mapping the two-dimensional adjacency matrix to a hidden variable feature space through an encoder to obtain a hidden feature vector; the hidden feature vector includes a mean and a variance; Mapping the hidden feature vector to the original sample space through a decoder to obtain a compensation sample, and reconstructing a two-dimensional adjacency matrix according to the compensation sample to obtain a target adjacency matrix; The target adjacency matrix is ​​converted to obtain a target heterogeneous graph, and a target data set is obtained by inversely constructing the target heterogeneous graph.

[0008] Optionally, determining the implicit needs of the user according to the target data set includes: The target data set is converted into a feature vector, and the feature vector is used as an input of an xDeepFM model for implicit feature extraction; the xDeepFM model includes a deep neural network and a compressed interaction network.

[0009] Optionally, performing language meaning detection on the target dataset and the comment dataset to obtain analysis results includes: Performing part-of-speech recognition on the sentences in the target data set and the comment data set, and removing meaningless words in the sentences to obtain a sentence set; Identifying target comments and target sub-comments in the sentence set, and performing part-of-speech analysis on the target comments and the target sub-comments; Calculating the sentiment scores of the target comment and the target sub-comment, and if the sentiment scores are greater than a judgment threshold, determining the target comment as a normal comment to obtain a comment set; The comment set is classified according to the sentiment score to obtain a classification set, and the classification set is used as an analysis result.

[0010] Optionally, calculating the sentiment scores of the target comment and the target sub-comment includes: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

[0011] In a second aspect of the present invention, a cross-border e-commerce data analysis device is proposed, comprising: a transaction data acquisition module, a data preprocessing module, an implicit demand determination module and a semantic analysis module: The transaction data acquisition module is used to acquire first transaction data of a target product, determine a product of the same category as the target product, and acquire second transaction data of the product of the same category; the transaction data includes: product information and user information; a product of the same category is a product that replaces the target product in all functional tags, and both the target product and the product of the same category contain multiple functional tags; The data preprocessing module is used to merge the first transaction data and the second transaction data to obtain third transaction data, and preprocess the third transaction data to obtain a target data set; The implicit demand determination module is used to determine the implicit demand of the user according to the target data set, determine the substitute product according to the implicit demand, and obtain the comments of the substitute product to obtain the comment data set; the substitute product is a product that replaces the target product in part of the functional label, and the substitute product includes a functional label; The semantic analysis module is used to perform language meaning detection on the target data set and the comment data set to obtain analysis results.

[0012] Optionally, the data preprocessing module includes: a heterogeneous graph construction module, a matrix encoding mapping module, a matrix decoding mapping module and a heterogeneous graph inverse construction module: The heterogeneous graph construction module is used to preprocess the third transaction data to obtain target data, construct a target heterogeneous graph according to the target data, and convert the target heterogeneous graph into a two-dimensional adjacency matrix; The matrix encoding mapping module is used to map the two-dimensional adjacency matrix to the hidden variable feature space through the encoder to obtain a hidden feature vector; the hidden feature vector includes a mean and a variance; The matrix decoding and mapping module is used to map the hidden feature vector to the original sample space through the decoder to obtain a compensation sample, and reconstruct a two-dimensional adjacency matrix according to the compensation sample to obtain a target adjacency matrix; The heterogeneous graph inverse construction module is used to convert the target adjacency matrix into a target heterogeneous graph, and obtain a target data set based on the inverse construction of the target heterogeneous graph.

[0013] Optionally, the implicit demand determination module is further used for: An implicit feature extraction module is used to convert the target data set into a feature vector, and use the feature vector as an input of an xDeepFM model to perform implicit feature extraction; the xDeepFM model includes a deep neural network and a compressed interaction network.

[0014] Optionally, the semantic analysis module includes: a part-of-speech recognition module, a part-of-speech analysis module, a sentiment score determination module and an analysis result acquisition module: The part-of-speech recognition module is used to perform part-of-speech recognition on the sentences in the target data set and the comment data set, and remove meaningless words in the sentences to obtain a sentence set; The part-of-speech analysis module is used to identify the target comments and target sub-comments in the sentence set, and perform part-of-speech analysis on the target comments and the target sub-comments by using a sentiment analysis tool; The sentiment score determination module is used to calculate the sentiment scores of the target comment and the target sub-comment, and if the sentiment score is greater than a judgment threshold, the target comment is determined to be a normal comment to obtain a comment set; The analysis result acquisition module is used to classify the review set according to the sentiment score to obtain a classification set, and use the classification set as the analysis result.

[0015] Optionally, the sentiment score determination module is further used to: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

[0016] Beneficial effects of the present invention: The present invention proposes a cross-border e-commerce data analysis method, which obtains first transaction data of a target commodity, determines the same commodity as the target commodity, and obtains second transaction data of the same commodity; merges the first transaction data and the second transaction data to obtain third transaction data, and pre-processes the third transaction data to obtain a target data set; determines the user's implicit needs according to the target data set, determines alternative commodities according to the implicit needs, and obtains comments on the alternative commodities to obtain a comment data set; performs language meaning detection on the target data set and the comment data set to obtain analysis results; determines the comments on alternative commodities by combining the analysis of the user's implicit needs, increases the richness of the data volume, effectively enhances the correlation between the commodity data and the user data, and makes the comment semantic analysis results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings.

[0018] Figure 1 A flowchart of a cross-border e-commerce data analysis method is provided for an embodiment of the present invention; Figure 2A schematic structural diagram of a cross-border e-commerce data analysis device is provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the present invention, the description of "first", "second", etc. is only used for descriptive purposes, and cannot be understood as indicating or implying its relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0020] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0021] The embodiment of the present invention provides a cross-border e-commerce data analysis method. Figure 1 , Figure 1 A flow chart of a cross-border e-commerce data analysis method provided by an embodiment of the present invention. The method comprises the following steps: S101, obtaining first transaction data of a target product, determining products of the same category as the target product, and obtaining second transaction data of products of the same category; S102, fusing the first transaction data and the second transaction data to obtain third transaction data, and preprocessing the third transaction data to obtain a target data set; S103, determining the user's implicit needs according to the target data set, determining alternative products according to the implicit needs, and obtaining comments on the alternative products to obtain a comment data set; S104, performing language meaning detection on the target data set and the comment data set to obtain analysis results; Transaction data includes: product information and user information; a generic product is a product that replaces the target product on all functional tags, and both the target product and the generic product contain multiple functional tags; a substitute product is a product that replaces the target product on some functional tags, and the substitute product contains one functional tag; A cross-border e-commerce data analysis method provided by an embodiment of the present invention determines alternative product reviews by combining implicit user demand analysis, thereby increasing the richness of data volume, effectively enhancing the correlation between product data and user data, and making the review semantic analysis results more accurate.

[0022] In one implementation, comments on a target product and transaction data of similar products are obtained. Similar products are competing products of the target product, which are similar in positioning and function labels, but have differences in other product information (e.g., appearance, brand, price, etc.). For example, for mobile phones of different brands, user information includes purchase records, search keywords, access time, etc.; comments on the product obtained include target comments and target sub-comments, that is, replies to the target comments are target sub-comments; the first data set and the second data set are fused to obtain third transaction data. Fusion is the process of merging the content, features, or information of two or more data sets into a new data set, which can increase data diversity and improve analysis accuracy.

[0023] In one implementation, the user's implicit needs are determined based on the target data set, and the corresponding products in the target data set, while the implicit needs are functional substitutes for the products needed by the users, that is, substitute products. Because substitute products have high cost performance, users like to look for and buy similar substitutes, that is, products that meet or can achieve certain functional substitutions and have only one functional label. Therefore, comments on finding users' implicit needs can also serve as effective suggestions for improving target products.

[0024] In one implementation, the comments are tested for linguistic meaning. Semantic meaning detection technology analyzes the text content, identifies and understands the deep meaning of the comments, and connects them with the context. It can determine the emotional tone of the text and improve the accuracy of text comprehension. By analyzing product reviews, we can gain in-depth understanding of consumer needs and feedback, which helps merchants evaluate product quality, discover market trends and potential development tracks, and provide data support for product improvement and marketing promotion.

[0025] In one implementation, language meaning detection is performed on the target dataset and the comment dataset to obtain analysis results, and merchants can obtain and study the analysis results to upgrade or iteratively optimize the products.

[0026] In one embodiment, step S102 includes: Preprocessing the third transaction data to obtain target data, constructing a target heterogeneous graph according to the target data, and converting the target heterogeneous graph into a two-dimensional adjacency matrix; The two-dimensional adjacency matrix is ​​mapped to the hidden variable feature space through the encoder to obtain the hidden feature vector; the hidden feature vector contains the mean and variance; The hidden feature vector is mapped to the original sample space through the decoder to obtain the compensation sample, and the two-dimensional adjacency matrix is ​​reconstructed according to the compensation sample to obtain the target adjacency matrix; The target adjacency matrix is ​​transformed into a target heterogeneous graph, and the target dataset is obtained by inversely constructing the target heterogeneous graph.

[0027] In one implementation, by preprocessing the third transaction data, key product attributes and user attributes are extracted, and then a target heterogeneous graph is constructed, so that an orderly connection is established between scattered data points. By converting the heterogeneous graph into a two-dimensional adjacency matrix, the data representation can be effectively simplified. A heterogeneous graph is a two-dimensional graph used to describe the interactive relationship between users and products; users and products are respectively used as nodes of the graph, and their interactive behaviors (such as purchase records, browsing time, evaluation, etc.) form the edges of the graph; the edges can be directed or undirected, depending on the nature of the interactive behavior. For example, if user A purchases product B, a directed edge from user A to product B can be drawn in the graph.

[0028] In one implementation, the functions of the encoder and decoder are realized through a graph variational autoencoder (VGAE), which maps the two-dimensional adjacency matrix to the hidden variable feature space through the encoder and extracts the hidden feature vector. The hidden feature vector contains the mean and variance, which define the Gaussian distribution of the latent variable. New node embedding representations are obtained by sampling from the Gaussian distribution. These embedding representations can be regarded as a supplement or enhancement to the original node features. By mapping the hidden feature vector back to the original sample space through the decoder, compensating samples can be obtained, and then the target adjacency matrix is ​​reconstructed and converted to the target heterogeneous graph again. That is, the target heterogeneous graph is indirectly compensated by generating a new node embedding representation. In the above steps, the purpose is to solve the missing parts in the original data (target data) and thus reduce the sparsity of the data.

[0029] In one embodiment, step S103 includes: The target dataset is converted into a feature vector, and the feature vector is used as the input of the xDeepFM model for implicit feature extraction; the xDeepFM model contains a deep neural network and a compressed interaction network.

[0030] In one implementation, a deep neural network performs multi-layer nonlinear transformation on a feature vector to obtain an output result, and the output result is integrated into the feature vector to obtain a first feature vector. A compressed interactive network processes the feature vector through each hidden layer, and concatenates the outputs of each hidden layer to obtain a second feature vector. The second feature vector is input into a graph attention network model to obtain the user's implicit needs.

[0031] In one implementation, the xDeepFM model combines a compressed interaction network and a deep neural network to jointly learn high-order interactive features in an explicit and implicit manner. The compressed interaction network is responsible for explicitly constructing finite-order feature combinations, while the deep neural network implicitly generates bit-level feature interactions; the target data set is converted into a feature vector and used as the input of the xDeepFM model. The xDeepFM model integrates two modules, the deep neural network and the compressed interaction network, which work together on the feature vector to deeply explore the inherent laws and potential associations of the data. Through multi-layer nonlinear transformations, the deep neural network can automatically learn and extract high-order feature interactions from the original features, and can also enhance the accuracy and generalization ability of feature representation. The compressed interaction network processes the feature vector layer by layer and splices the outputs of each hidden layer to form a second feature vector. This step retains the interactive information between features and realizes efficient integration of features. The synergy of these two modules enables the xDeepFM model to solve the problems of sparsity, difficulty in extracting complex features, and mining feature crossover in traditional models, significantly improving the accuracy of recommendations.

[0032] In one implementation, the second feature vector extracted by the xDeepFM model is input into the graph attention network model, which is good at processing graph structure data and can capture the complex relationships and dynamic changes between nodes. By taking the second feature vector as input, the graph attention network model can make full use of the information in the feature vector and accurately model the mutual influence between nodes in the target heterogeneous graph through the attention mechanism. It can mine the direct connection between users and products, and the indirect relationship and potential demand between the target product function and alternative products, that is, find the user's hidden needs. During the training process, the graph attention network model takes the above-mentioned target heterogeneous graph and feature vectors (nodes, each node of the heterogeneous graph corresponds to a feature vector) as input for training, and uses the attention mechanism to aggregate information and update nodes, thereby realizing the mining and analysis of implicit data; the graph attention network model can also dynamically adjust according to the feature information of the nodes to realize real-time capture and accurate prediction of user implicit needs; users search for target products, but alternative products have certain functional substitutions for target products, and the alternative products have a higher cost-effectiveness. If users do not have additional functional requirements, then the probability of users choosing alternative products will be higher. For example: the cooling bracket of a laptop computer can be replaced by a barbecue grill to achieve certain functions and meet the cooling needs of the user's computer. At this time, the barbecue grill is a substitute product, which is an implicit demand.

[0033] In one embodiment, step S104 includes: Perform part-of-speech recognition on the sentences in the target dataset and the comment dataset, and remove meaningless words in the sentences to obtain a sentence set; Identify target comments and target sub-comments in the sentence set, and perform part-of-speech analysis on the target comments and target sub-comments; Calculate the sentiment scores of the target comment and the target sub-comment. If the sentiment score is greater than the judgment threshold, the target comment is judged as a normal comment to obtain a comment set. The review set is classified according to the sentiment scores to obtain a classification set, and the classification set is used as the analysis result.

[0034] In one implementation, by performing part-of-speech recognition on the sentences in the target data set and the comment data set and removing meaningless words, a set of sentences that are easier to identify and judge semantics can be constructed. The above steps not only significantly reduce the complexity of data processing, but also improve the accuracy and efficiency of subsequent analysis. The removal of meaningless words, such as function words, redundant information, and ambiguous reduplications, makes key information more prominent, that is, retaining the subject, predicate, and object of the sentence can effectively judge the core content and emotional tendency of the sentence. On this basis, in-depth analysis of the target comments and target sub-comments, that is, comprehensive emotional judgment in connection with the context, can extract the user's views and emotions contained in the comments, and provide a more reliable data basis for subsequent emotional score calculations and comment judgments.

[0035] In one implementation, a sentiment analysis tool is used to perform part-of-speech analysis on the target comment and target sub-comment, and calculate their sentiment scores. The sentiment analysis tool includes, for example: LSTM (Long Short-Term Memory): a special recurrent neural network (RNN) that can capture long-term dependencies in text and is suitable for tasks such as sentiment analysis; Emotional Analysis: a sentiment analysis tool based on natural language processing (NLP) that uses advanced machine learning algorithms and can perform in-depth analysis of various types of text data; by calculating the sentiment score of a sentence, the emotional tendency of user comments can be quantified. This allows the emotional color of the comment to be represented more accurately in digital form, and also provides an objective and scientific basis for subsequent comment judgment. When the sentiment score is greater than the judgment threshold, the target comment is judged as a normal comment and output as an analysis result. This judgment process simplifies the complexity of comment screening and improves the accuracy and credibility of the screening results. When the sentiment score is not greater than the judgment threshold, the target comment is judged as an abnormal comment and is removed to achieve effective opinion extraction; the calculation of sentiment scores and comment judgment also provide merchants with valuable information feedback, which helps merchants better understand users' satisfaction and expectations for products or services, and understand users' potential requirements.

[0036] In one embodiment, step S104 further includes: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

[0037] In one implementation, the attribute relevance score is the correlation between the comment topic in the comment area and the target product. If one of the attributes is met, the attribute relevance score is set to 1, otherwise it is -1. For example, the comment area is discussing the price, function or appearance of the target product. At this time, if one of the price, function or appearance is met, the attribute relevance score is 1; if the attribute relevance is -1, it means that the relevance of the comment is poor, that is, the comment is irrelevant to the target product and there is no value in extracting opinions; the sentiment score of the target comment and the sentiment score of the target sub-comment are both set as floating point numbers between 0 and 1, where the number of target sub-comments is the number of valid replies under the target comment. The more replies to the valid comments, the more discussion and relevance the comment has on the target product, the higher the value of the comment, and the more targeted the product improvement opinions extracted from it.

[0038] Based on the same inventive concept, the present invention also provides a cross-border e-commerce data analysis device. Figure 2 , Figure 2 A schematic diagram of the structure of a cross-border e-commerce data analysis device provided by an embodiment of the present invention includes: a transaction data acquisition module, a data preprocessing module, an implicit demand determination module and a semantic analysis module: The transaction data acquisition module is used to acquire the first transaction data of the target product, determine the same category products of the target product, and acquire the second transaction data of the same category products; the transaction data includes: product information and user information; the same category products are products that can replace the target product in all functional tags, and the target product and the same category products both contain multiple functional tags; A data preprocessing module, used for fusing the first transaction data and the second transaction data to obtain third transaction data, and preprocessing the third transaction data to obtain a target data set; The implicit demand determination module is used to determine the implicit demand of the user according to the target data set, determine the substitute product according to the implicit demand, and obtain the comments of the substitute product to obtain the comment data set; the substitute product is a product that replaces the target product in part of the functional label, and the substitute product includes a functional label; The semantic analysis module is used to detect the language meaning of the target dataset and the comment dataset to obtain the analysis results.

[0039] A cross-border e-commerce data analysis device provided based on an embodiment of the present invention determines alternative product reviews by combining implicit user demand analysis, thereby increasing the richness of data volume, effectively enhancing the correlation between product data and user data, and making the review semantic analysis results more accurate.

[0040] In one embodiment, the data preprocessing module includes: a heterogeneous graph construction module, a matrix encoding mapping module, a matrix decoding mapping module and a heterogeneous graph inverse construction module: A heterogeneous graph construction module, used for preprocessing the third transaction data to obtain target data, constructing a target heterogeneous graph according to the target data, and converting the target heterogeneous graph into a two-dimensional adjacency matrix; The matrix encoding mapping module is used to map the two-dimensional adjacency matrix to the hidden variable feature space through the encoder to obtain the hidden feature vector; the hidden feature vector includes the mean and variance; A matrix decoding and mapping module is used to map the hidden feature vector to the original sample space through a decoder to obtain a compensation sample, and reconstruct a two-dimensional adjacency matrix according to the compensation sample to obtain a target adjacency matrix; The heterogeneous graph inverse construction module is used to transform the target adjacency matrix into a target heterogeneous graph, and obtain the target data set based on the inverse construction of the target heterogeneous graph.

[0041] In one embodiment, the implicit demand determination module is further configured to: The implicit feature extraction module is used to convert the target data set into a feature vector and use the feature vector as the input of the xDeepFM model for implicit feature extraction; the xDeepFM model includes a deep neural network and a compressed interaction network.

[0042] In one embodiment, the semantic analysis module includes: a part-of-speech recognition module, a part-of-speech analysis module, a sentiment score determination module, and an analysis result acquisition module: The part-of-speech recognition module is used to perform part-of-speech recognition on the sentences in the target dataset and the comment dataset, and remove meaningless words in the sentences to obtain a sentence set; A part-of-speech analysis module is used to identify target comments and target sub-comments in a sentence set, and perform part-of-speech analysis on the target comments and target sub-comments using a sentiment analysis tool; The sentiment score determination module is used to calculate the sentiment scores of the target comment and the target sub-comment. If the sentiment score is greater than the judgment threshold, the target comment is determined to be a normal comment to obtain a comment set; The analysis result acquisition module is used to classify the comment set according to the sentiment score to obtain a classification set, and use the classification set as the analysis result.

[0043] In one embodiment, the sentiment score determination module is further used to: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

[0044] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A cross-border e-commerce data analysis method, characterized in that: The method comprises: Acquire first transaction data of a target product, determine a product of the same category as the target product, and acquire second transaction data of the product of the same category; the transaction data includes: product information and user information; a product of the same category is a product that replaces the target product in all functional tags, and both the target product and the product of the same category contain multiple functional tags; fusing the first transaction data and the second transaction data to obtain third transaction data, and preprocessing the third transaction data to obtain a target data set; Determine the user's implicit needs according to the target data set, determine the substitute product according to the implicit needs, and obtain comments of the substitute product to obtain a comment data set; the substitute product is a product that replaces the target product in part of the functional labels, and the substitute product includes a functional label; Language meaning detection is performed on the target data set and the comment data set to obtain analysis results.

2. A cross-border e-commerce data analysis method according to claim 1, characterized in that: Preprocessing the third transaction data to obtain a target data set includes: Preprocessing the third transaction data to obtain target data, constructing a target heterogeneous graph according to the target data, and converting the target heterogeneous graph into a two-dimensional adjacency matrix; Mapping the two-dimensional adjacency matrix to a hidden variable feature space through an encoder to obtain a hidden feature vector; the hidden feature vector includes a mean and a variance; Mapping the hidden feature vector to the original sample space through a decoder to obtain a compensation sample, and reconstructing a two-dimensional adjacency matrix according to the compensation sample to obtain a target adjacency matrix; The target adjacency matrix is ​​converted to obtain a target heterogeneous graph, and a target data set is obtained by inversely constructing the target heterogeneous graph.

3. A cross-border e-commerce data analysis method according to claim 1, characterized in that: Determining the implicit needs of users according to the target data set includes: The target data set is converted into a feature vector, and the feature vector is used as an input of an xDeepFM model for implicit feature extraction; the xDeepFM model includes a deep neural network and a compressed interaction network.

4. A cross-border e-commerce data analysis method according to claim 1, characterized in that: The analysis results obtained by performing language meaning detection on the target dataset and the comment dataset include: Performing part-of-speech recognition on the sentences in the target data set and the comment data set, and removing meaningless words in the sentences to obtain a sentence set; Identifying target comments and target sub-comments in the sentence set, and performing part-of-speech analysis on the target comments and the target sub-comments; Calculating the sentiment scores of the target comment and the target sub-comment, and if the sentiment scores are greater than a judgment threshold, determining the target comment as a normal comment to obtain a comment set; The review set is classified according to the sentiment score to obtain a classification set, and the classification set is used as an analysis result.

5. A cross-border e-commerce data analysis method according to claim 4, characterized in that: Calculating the sentiment scores of the target comment and the target sub-comment includes: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

6. A cross-border e-commerce data analysis device, characterized in that: The device comprises: a transaction data acquisition module, a data preprocessing module, an implicit demand determination module and a semantic analysis module: The transaction data acquisition module is used to acquire first transaction data of a target product, determine a product of the same category as the target product, and acquire second transaction data of the product of the same category; the transaction data includes: product information and user information; a product of the same category is a product that replaces the target product in all functional tags, and both the target product and the product of the same category contain multiple functional tags; The data preprocessing module is used to merge the first transaction data and the second transaction data to obtain third transaction data, and preprocess the third transaction data to obtain a target data set; The implicit demand determination module is used to determine the implicit demand of the user according to the target data set, determine the substitute product according to the implicit demand, and obtain the comments of the substitute product to obtain the comment data set; the substitute product is a product that replaces the target product in part of the functional label, and the substitute product includes a functional label; The semantic analysis module is used to perform language meaning detection on the target data set and the comment data set to obtain analysis results.

7. A cross-border e-commerce data analysis device according to claim 6, characterized in that: The data preprocessing module includes: a heterogeneous graph construction module, a matrix encoding mapping module, a matrix decoding mapping module and a heterogeneous graph inverse construction module: The heterogeneous graph construction module is used to preprocess the third transaction data to obtain target data, construct a target heterogeneous graph according to the target data, and convert the target heterogeneous graph into a two-dimensional adjacency matrix; The matrix encoding mapping module is used to map the two-dimensional adjacency matrix to the hidden variable feature space through the encoder to obtain a hidden feature vector; the hidden feature vector includes a mean and a variance; The matrix decoding and mapping module is used to map the hidden feature vector to the original sample space through the decoder to obtain a compensation sample, and reconstruct a two-dimensional adjacency matrix according to the compensation sample to obtain a target adjacency matrix; The heterogeneous graph inverse construction module is used to convert the target adjacency matrix into a target heterogeneous graph, and obtain a target data set based on the inverse construction of the target heterogeneous graph.

8. A cross-border e-commerce data analysis device according to claim 6, characterized in that: The implicit demand determination module is also used for: An implicit feature extraction module is used to convert the target data set into a feature vector, and use the feature vector as an input of an xDeepFM model to perform implicit feature extraction; the xDeepFM model includes a deep neural network and a compressed interaction network.

9. A cross-border e-commerce data analysis device according to claim 6, characterized in that: The semantic analysis module includes: a part-of-speech recognition module, a part-of-speech analysis module, a sentiment score determination module and an analysis result acquisition module: The part-of-speech recognition module is used to perform part-of-speech recognition on the sentences in the target data set and the comment data set, and remove meaningless words in the sentences to obtain a sentence set; The part-of-speech analysis module is used to identify the target comments and target sub-comments in the sentence set, and perform part-of-speech analysis on the target comments and the target sub-comments by using a sentiment analysis tool; The sentiment score determination module is used to calculate the sentiment scores of the target comment and the target sub-comment, and if the sentiment score is greater than a judgment threshold, the target comment is determined to be a normal comment to obtain a comment set; The analysis result acquisition module is used to classify the review set according to the sentiment score to obtain a classification set, and use the classification set as the analysis result.

10. A cross-border e-commerce data analysis device according to claim 9, characterized in that: The sentiment score determination module is also used for: Calculate the sentiment score using the formula Calculate the sentiment score; Among them, F is the sentiment score, is the sentiment score of the target comment, S is the attribute relevance score, is the sentiment score of the target sub-comment, α and β are constant proportional coefficients and are not zero.

Images