E-commerce commodity category adaptive identification and semantic evaluation method

Abstract

The invention provides an e-commerce commodity category adaptive recognition and semantic evaluation method, which comprises the following steps of: extracting path hops and associated edge semantic type labels between target commodities from a commodity relation graph, and extracting all shortest path sets through graph traversal to obtain a hop distance between the commodities and an edge type combination; performing multi-layer information propagation on the hop count distance between the commodities by adopting a multi-layer graph convolutional neural network through the information transmission loss value, and determining a commodity semantic embedding vector after propagation of each layer; according to the propagated commodity semantic embedding vector of each layer, the vector distance between commodity semantic embedding is extracted, the similarity is evaluated, and the optimized information propagation aggregation layer number is obtained based on the similarity convergence trend; and evaluating the coincidence degree of the commodity category inference result and the actual commodity category attribution, and identifying the category inference deviation to obtain the correction direction of the self-adaptive identification of the e-commerce commodity category.

Description

An Adaptive Recognition and Semantic Evaluation Method for E-commerce Product Categories Technical Field

[0001] This invention relates to the field of information technology, and in particular to an adaptive identification and semantic evaluation method for e-commerce product categories. Background Technology

[0002] In e-commerce platforms, the automatic identification and semantic evaluation of product categories directly determine the accuracy of recommendations, search ranking effects, and advertising efficiency, making them core elements for improving user experience and transaction conversion rates. With the rapid growth of product categories, platforms need to accurately determine whether two products belong to the same category or have a highly similar semantic relationship, thereby achieving precise classification and intelligent recommendations.

[0003] Most existing methods rely on the shortest path hop count of products in knowledge graphs or relational networks to measure category similarity; fewer hops are generally considered to indicate closer categories. However, this approach often leads to misjudgments in practical applications because the path connecting products may contain multiple layers of meaning. For example, some edges might indicate that two products are frequently purchased together, forming a complementary relationship, while others might indicate that the two products are interchangeable. Even two- or three-hop paths can carry completely opposite semantic information, leading to products with shorter path distances belonging to entirely different categories in actual business scenarios.

[0004] The core of this problem lies in the failure to effectively distinguish and utilize the semantic differences of the associated edges and the attenuation characteristics during information propagation. When products are indirectly connected through multiple edges of different semantic types, simply relying on the number of hops cannot reflect the true proximity of categories. For example, a shirt is connected to a tie through a "matching purchase" relationship, and then to a suit jacket through the same "matching purchase" relationship. Although this path only has two hops, it strongly suggests that all three belong to the formal wear category. However, if the second edge becomes "the user buys a tie and then returns it to buy a scarf instead," the semantics of this edge become a substitution relationship. At this point, even if the path still has two hops, the category affiliation of the shirt and the scarf is clearly different. If we continue to use a fixed number of hops as the basis for aggregating neighborhoods, products that should not be classified into the same category will be mistakenly brought closer together, resulting in chaotic search results or recommendations of obviously irrelevant products to the user. Therefore, how to simultaneously consider the number of hops, the semantic differences carried by the associated edges, and the natural attenuation of information when it is transmitted through different semantic edges in the product association network, so as to dynamically determine the effective neighborhood range and aggregation strength under each type of semantic path, becomes the key issue for achieving adaptive and accurate identification of product categories. Summary of the Invention

[0005] To address the problem of information propagation loss and category determination bias caused by the increased number of path hops and differences in semantic strength of edge types in existing technologies for indirectly related products in complex product relationship graphs, this invention provides an adaptive identification and semantic evaluation method for e-commerce product categories. This method achieves high-precision inference and bias correction of indirectly related product categories, improving the accuracy and robustness of category identification for products without clear target categories.

[0006] The technical solution of this invention is implemented as follows: An adaptive identification and semantic evaluation method for e-commerce product categories includes: extracting the path hop count and semantic type labels of associated edges between target products from a product relationship graph; extracting all shortest path sets through graph traversal to obtain the hop distance and edge type combination between products; obtaining an information transmission loss value based on the hop distance and edge type combination, according to the semantic association strength of edge types combined with the path hop count increment and decay rule; using the information transmission loss value, performing multi-layer graph convolutional neural network to propagate multi-layer information from the hop distance between products, determining the product semantic embedding vector after each layer of propagation; and extracting the direction between product semantic embeddings based on the product semantic embedding vectors after each layer of propagation. The distance is measured to assess the similarity level, and the optimized information propagation aggregation layer is obtained based on the similarity convergence trend. Based on the optimized information propagation aggregation layer, the semantic strength of the categories is aggregated, and the similarity between the neighborhood aggregation results and the standard features of each product category is evaluated to obtain a preliminary category assignment probability. Using the preliminary category assignment probability, the dominance ratio of substitute edges and complementary edges is analyzed in conjunction with the information transmission loss value. When substitute edges dominate, the sensitivity of category determination is increased; when complementary edges dominate, the sensitivity of category determination is decreased, thus determining the product category inference result. The degree of consistency between the product category inference result and the actual product category assignment is evaluated to identify category inference bias and obtain the correction direction for adaptive recognition of e-commerce product categories.

[0007] This invention extracts the hop count and edge type combinations of all shortest paths between target products through graph traversal. It quantifies information transmission loss based on the semantic association strength of edge types and the hop count increment decay rule. Then, it uses a multi-layer graph convolutional neural network with edge type awareness for hierarchical message propagation, dynamically obtains the semantic embedding of products under each hop count, and adaptively determines the optimal aggregation layer based on the convergence trend of embedding vector similarity. Finally, it adjusts the sensitivity of category determination by combining the proportion of substitute edges and complementary edges in the path, and finally achieves high-precision indirect association product category inference and bias correction, significantly improving the accuracy and robustness of category identification for products without clear category labels.

[0008] As a further optimization of the above scheme, the step of extracting the path hop count and semantic type labels of associated edges between target products from the product relationship graph, and extracting all shortest path sets through graph traversal to obtain the hop count distance and edge type combination between products includes: reading all product nodes and their direct connections from the product relationship graph, obtaining the semantic type label of each edge: a "same purchase" edge indicates that the user purchased a pair of products in the same order, a "substitute" edge indicates that the user returned and then purchased a pair of products instead, and a "complementary" edge indicates that the user purchased a pair of products that match the function in multiple transactions. A breadth-first traversal algorithm is used to expand layer by layer from the starting product node, recording the hop count and edge sequence from the starting node to each target node; based on the traversal results, the shortest path between each pair of products is extracted, the number of "same purchase" edges, "substitute" edges, and "complementary" edges on the path is counted, and an edge type sequence string is generated according to the order of edge type appearance to determine the hop count distance value and edge type combination identifier for each pair of products.

[0009] By defining three types of relationships with clear business semantics—"same-purchase edge," "alternative edge," and "complementary edge"—and using breadth-first traversal to accurately extract all shortest paths and their edge type sequences between products, not only is the structured semantic information in the original user behavior preserved, but semantic distortion caused by ignoring path diversity or misusing non-shortest paths is also avoided. This provides a high-fidelity, interpretable path foundation for subsequent information dissemination modeling, significantly improving the accuracy and traceability of multi-hop association modeling.

[0010] As a further optimization of the above scheme, the step of obtaining the information transmission loss value based on the hop distance between goods and the edge type combination, and combining the semantic association strength of the edge type with the path hop count increment decay rule, includes: based on the hop distance between goods and the edge type combination identifier, counting the number of co-purchase edges, substitute edges, and complementary edges on each path, calculating the proportion of each type of edge, and calculating the path edge type diversity index based on the distribution of the number of each edge type; based on the path edge type distribution characteristics, reading the pre-set semantic association strength value of the edge type, assigning strong association weight to co-purchase edges, medium association weight to substitute edges, and weak association weight to complementary edges, multiplying the corresponding semantic association weight of each edge for each path edge sequence, and introducing a decay factor based on the path hop count to calculate the path propagation strength value; using the path propagation strength value, using a negative exponential function to calculate the degree of information decay from the starting product to the target product, and determining the information transmission loss value between goods based on the product of the decay degree and the path hop count.

[0011] By assigning different association weights to different semantic edge types and introducing an exponential decay mechanism based on hop count, a refined quantification of the credibility of multi-hop path information is achieved. This design effectively suppresses the interference of long or weak semantic paths on category inference, while strengthening the dominant role of strong semantic short paths. This allows for a more reasonable allocation of information propagation weights in complex graph structures, reducing the risk of category drift caused by "semantic dilution."

[0012] As a further optimization of the above scheme, the step of using a multi-layer graph convolutional neural network to propagate information across multiple layers of the hop distance between products through information transmission loss values, and determining the product semantic embedding vector after each layer of propagation, includes: constructing a graph information propagation unit based on the information transmission loss values; reading the initial feature vector of the product node, including product category code, sales attributes, and user behavior features; obtaining a set of one-hop neighbor nodes for each product node; grouping them according to same-purchase edges, substitute edges, and complementary edges; and obtaining neighbor node groups distinguished by edge type by the highest sampling ratio for same-purchase edges, followed by substitute edges, and then complementary edges. The node groups and their initial feature vectors are processed using a graph convolutional neural network to aggregate features from each neighboring node group. For the same-side group, neighbor features are aggregated using mean pooling; for the alternative-side group, significantly different features are extracted using max pooling; and for the complementary-side group, weights are calculated using an attention mechanism and then summed. The three aggregation results are then concatenated to form the current layer's node representation. For this node representation, multi-layer information propagation is performed. Each layer uses the output of the previous layer as the input of the current layer to expand the neighborhood range. The feature aggregation process is repeated, and the retention rate is calculated based on the information transmission loss value. The retention rate is multiplied by the aggregation result to obtain the decayed features. After propagation through a preset number of layers, the semantic embedding vectors of each layer are output.

[0013] By grouping and sampling neighbors according to edge type, the model can adaptively capture the essential features of various relationships: co-purchase reflects co-occurrence consistency, substitution reflects functional similarity, complementation expresses functional synergy, and different feature aggregation strategies are adopted for different groups, which not only preserves the hierarchical nature of multi-hop semantics but also prevents noise accumulation. In addition, the information transmission loss value is combined to dynamically control the feature attenuation in multi-layer propagation, generating hierarchical embedding vectors that can accurately express the complex semantic relationships of multi-hop products.

[0014] As a further optimization of the above scheme, the step of extracting the vector distance between product semantic embeddings based on the product semantic embedding vectors after propagation at each layer, evaluating the similarity level, and obtaining the optimized information propagation aggregation layer based on the similarity convergence trend includes: obtaining the product semantic embedding vectors after propagation at each layer, extracting the multi-dimensional embedding representation of each product node layer by layer, calculating two metrics, Euclidean distance and cosine similarity, for each pair of products to obtain the distance matrix of product pairs at each level; according to the distance matrix, calculating the rate of change of product pair similarity with each additional layer of propagation depth, and recording the layer number as a candidate aggregation depth when the average similarity change rate of two consecutive layers is lower than a preset convergence threshold; identifying the convergence differentiation pattern of different product category levels through the convergence layer distribution and the candidate aggregation depth, calculating the aggregation weight of each layer according to the initial product category label and the distribution of neighboring node categories, and determining the optimized information propagation aggregation layer for each product category.

[0015] Abandoning the traditional approach of fixing the number of propagation layers, this method determines the number of information propagation aggregation layers by monitoring the convergence trend of the similarity of product semantic embeddings in different propagation layers, thus achieving adaptive optimization of the propagation range. This adaptive mechanism can automatically adjust the receptive field according to the connectivity density and semantic complexity of different product subgraphs, maintaining the clarity of category boundaries while ensuring sufficient information fusion, significantly improving the model's generalization ability and computational efficiency.

[0016] As a further optimization of the above scheme, the step of aggregating the semantic strength of categories based on the optimized information propagation aggregation layer, evaluating the similarity between the neighborhood aggregation result and the standard features of each product category, and obtaining a preliminary category assignment probability includes: extracting the semantic embedding vector of each product at the corresponding level according to the optimized information propagation aggregation layer, aggregating the semantic strength of each layer using a weighted average method, with the first layer having a weight of 1, and the weight of each subsequent layer decreasing by a preset decay coefficient, and obtaining the aggregated comprehensive semantic representation by summing the products of each layer's embedding vector and its corresponding weight and then dividing by the total weight; calculating the cosine similarity between the comprehensive semantic representation and the standard feature vector of each product category, where the standard feature vector is obtained by averaging the historical embedding vectors of all products under that category, with each category corresponding to one standard feature vector, and obtaining a set of candidate categories and corresponding similarity scores; using the set of candidate categories and similarity scores, normalizing the similarity scores of each candidate category using the softmax function, so that the sum of the probability values ​​of all candidate categories is 1, and determining the preliminary category assignment probability of each product based on the normalized probability distribution.

[0017] By aggregating multi-layered semantic information weighted by a decay coefficient and comparing and normalizing it with predefined category standard features, semantic alignment from local neighborhood to global category system is achieved. This approach not only utilizes multi-layered contextual information but also introduces prior category knowledge, making the initial category attribution probability both data-driven and domain-reasonable. This provides a high-confidence input for subsequent sensitivity adjustments, demonstrating robustness, especially when handling cross-category edge products.

[0018] As a further optimization of the above scheme, the step of analyzing the dominant proportion of substitute and complementary edges by combining the preliminary category assignment probability with the information transmission loss value, increasing the sensitivity of category determination when substitute edges dominate, and decreasing the sensitivity of category determination when complementary edges dominate, to determine the product category inference result, includes: statistically analyzing the distribution of all connection edge types in the product association network using the preliminary category assignment probability, calculating the proportion of substitute and complementary relationship edges, and obtaining an edge type dominance index; according to the edge type dominance index, if the proportion of substitute edges exceeds a preset threshold, then the probability values ​​of each category are exponentially amplified by the difference from the highest probability; if the proportion of complementary edges exceeds a preset threshold, then all probability values ​​are logarithmically compressed to obtain an adjusted category probability distribution; using the adjusted category probability distribution, the category with the highest probability is selected as the product category inference result.

[0019] By nonlinearly adjusting the initial classification probability based on the dominance of substitution and complementary edges in the network, when substitution is dominant, it indicates that the products are functionally interchangeable, thus amplifying the confidence of high-probability categories to enhance discriminative sensitivity. When complementarity is dominant, it indicates that the products are functionally compatible but belong to different categories, thus compressing the probability distribution to avoid misclassification caused by over-normalization. This nonlinear adjustment strategy enables the model to intelligently switch between "strict classification" and "relaxed inclusion" modes based on the association pattern. This improves the classification accuracy and robustness under different association patterns, namely, the accuracy of category inference and business adaptability in complex association scenarios.

[0020] As a further optimization of the above scheme, the evaluation of the degree of consistency between the inferred product category and the actual product category attribution, the identification of category inference bias, and the determination of the correction direction for adaptive identification of e-commerce product categories include: evaluating the degree of consistency between the inferred product category and the actual product category attribution, comparing the inferred category with the actual attribution of the product in the preset category system of the e-commerce platform, and obtaining the category identification accuracy and bias distribution; based on the bias distribution, statistically analyzing frequently misjudged category pairs and their frequency of occurrence, identifying systematic bias patterns, and determining the correction direction for adaptive identification of e-commerce product categories by increasing the difference in the judgment threshold between easily confused categories.

[0021] By analyzing the deviation between the inferred results and the actual categories, the system identifies frequently misclassified category pairs, providing a clear direction for correction of the classification model. Based on this, the system dynamically adjusts the judgment threshold or feature distance between easily confused categories, realizing the system's self-feedback and continuous optimization capabilities. It can also be used to guide the reconstruction of the category system or the improvement of the labeling strategy, giving the entire system the ability to learn and correct itself, and significantly improving the stability of category recognition and the efficiency of platform governance under long-term operation.

[0022] Compared with traditional processing methods, the present invention has the following advantages: (1) Dynamic adjustment of classification strategy: By analyzing the dominant proportion of substitution edges and complementary edges in the commodity association network, the sensitivity of category determination is adaptively adjusted, so that the classification decision can flexibly adapt to different association patterns, significantly improving the accuracy and robustness of the model in complex scenarios.

[0023] (2) Adaptive domain aggregation: By dynamically monitoring the similarity convergence trend of semantic embedding at each layer, the optimal information aggregation depth and weight are adaptively determined for different products, realizing intelligent optimization of the propagation range and effectively capturing the semantics of products under multi-hop complex associations. Attached Figure Description

[0024] Figure 1 is a flowchart of an adaptive identification and semantic evaluation method for e-commerce product categories according to the present invention; Figure 2 is a schematic diagram of an adaptive identification and semantic evaluation method for e-commerce product categories according to the present invention; Figure 3 is another schematic diagram of an adaptive identification and semantic evaluation method for e-commerce product categories according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example: As shown in Figures 1-3, this example provides an adaptive identification and semantic evaluation method for e-commerce product categories, including the following steps: Step S101, extract the path hop count and semantic type labels of associated edges between target products from the product relationship graph, extract all shortest path sets through graph traversal, and obtain the hop distance and edge type combination between products.

[0027] Read all product nodes and their direct connections from the product relationship graph, and obtain the semantic type label of each edge. Among them, the same purchase edge represents the product pair purchased by the user in the same order, the substitution edge represents the product pair that the user returned and bought instead, and the complementary edge represents the product pair that the user purchased in separate orders but has a functional matching relationship. The breadth-first traversal algorithm is used to expand layer by layer from the starting product node. During the expansion process, the jump value and the edge sequence traversed from the starting node to each target node are recorded to obtain the traversal result containing path information.

[0028] Based on the traversal results, the shortest path between each pair of products is extracted, and the number of co-purchase edges, substitute edges, and complementary edges on the path is counted. If the number of co-purchase edges exceeds a preset threshold, the path is marked as a strongly associated type. If the number of substitute edges exceeds a preset threshold, the path is marked as a replaceable type. If the number of complementary edges exceeds a preset threshold, the path is marked as a matching type. An edge type sequence string is generated according to the order in which the edge types appear, and the hop distance value and edge type combination identifier for each pair of products are determined.

[0029] In this embodiment, the product relationship graph can be constructed based on historical transaction data from an e-commerce platform. Each product serves as a node in the graph, and the purchase relationships between products are represented as edges. Co-purchase edges are identified by analyzing product combinations within the same order, substitute edges are extracted from return and exchange records, and complementary edges are determined based on users' continuous purchase behavior within a short time window.

[0030] The breadth-first traversal starts from a specified initial node and maintains a queue to store nodes to be visited. Each time, a node is taken from the head of the queue and all its adjacent nodes are visited, recording the hop count to reach each node. During the traversal, the algorithm simultaneously saves the complete path from the starting node to each target node, including all intermediate nodes and their corresponding edge types, forming a path tracking record.

[0031] For example, when analyzing a path from athletic shoes to a sports water bottle, if the path passes through a sportswear node and both edges are "same-purchase" edges, the path is marked as a strongly associated type, indicating that these items are often purchased by users as a combination of sports equipment. The edge type combination identifier is stored in string form, such as "same-purchase and same-purchase" or "substitute and complementary", directly reflecting the semantic characteristics of the path.

[0032] Step S102: Based on the hop distance between products and the combination of edge types, evaluate the proportion of co-purchase edges, substitute edges, and complementary edges, identify the diversity of edge types and the magnitude of propagation loss, and obtain the information transmission loss value based on the semantic association strength of edge types and the path hop count increasing attenuation rule.

[0033] Based on the distance between product hops and the edge type combination identifier, the number of same-purchase edge, substitute edge, and complementary edge on each path is counted, and the proportion of each type of edge in the total number of edges in the path is calculated. If the proportion of a single edge type exceeds the first preset threshold, it is marked as a single-type path. If the difference between the proportions of the three edge types is less than the second preset threshold, it is marked as a mixed-type path. The edge type diversity index of the path is calculated based on the distribution of the number of each edge type, and the edge type distribution characteristics of the path are obtained.

[0034] Based on the path edge type distribution characteristics, the pre-set semantic association strength value of the edge type is read, where the same edge is given a strong association weight, the substitute edge is given a medium association weight, and the complementary edge is given a weak association weight. For the edge sequence on each path, the semantic association weight corresponding to each edge is multiplied. At the same time, an attenuation factor is introduced according to the number of hops in the path. The more hops, the greater the attenuation. The path propagation strength value is then calculated.

[0035] Based on the propagation intensity value along the path, the degree of information attenuation from the starting product to the target product is calculated using a negative exponential function. The degree of attenuation is inversely proportional to the propagation intensity value. Where I(d) represents the information intensity after traveling a distance d; I0 represents the initial information intensity of the starting product; λ represents the information attenuation coefficient; d represents the path distance from the starting product to the target product; and e represents the base of the natural constant. This formula describes the law that information intensity decreases exponentially with the propagation distance. If the attenuation exceeds a preset loss threshold, the path is determined to be a high-loss path. The information transmission loss value between products is determined by multiplying the attenuation level by the number of hops in the path.

[0036] In this embodiment, when performing edge type statistics for each product association path, a sliding window is used to traverse all edges on the path, recording the occurrence frequency of co-purchase edges, substitute edges, and complementary edges, and dividing by the total number of edges on the path to obtain their respective proportion values. When the proportion of co-purchase edges exceeds 0.7, it is marked as a strongly cooperative path; when the proportion of substitute edges exceeds 0.6, it is marked as a competitive path.

[0037] In this embodiment, the edge type diversity index is calculated based on information theory principles, reflecting the uniformity of edge type distribution in the path. The diversity index is obtained by calculating the negative logarithmic weighted sum of the proportions of each edge type. When the path contains only a single edge type, the diversity index is close to zero, indicating that the path has a simple semantic meaning. When the proportions of the three edge types are similar, the diversity index reaches its peak, indicating that the path has complex semantic meaning. Specifically, in an e-commerce scenario, if the path from a laptop to a mouse pad passes through two intermediate nodes—the keyboard and the mouse—and the four edges are related as "same purchase edge," "same purchase edge," "substitute edge," and "complementary edge," then the path has a high diversity index, reflecting the complexity of product associations. This complexity directly affects the accuracy of product category determination; a high-diversity path often means increased uncertainty in product category attribution.

[0038] In this embodiment, the weighting of semantic association strength is based on the actual business rules of the e-commerce platform. "Same-purchase relationship" indicates that users actively combine products for purchase, reflecting strong category relevance, and is assigned a high weight; "substitution relationship" indicates that products have similar functions but different brands or models, and is assigned a medium weight; "complementary relationship" only indicates complementary functions with weaker category relevance, and is assigned a low weight.

[0039] In this embodiment, the path propagation strength value is calculated using a continuous multiplication method, starting from the starting node and multiplying the semantic association weights of each edge on the path sequentially. The hop count decay factor is set according to an exponential decreasing rule: the first hop retains its original value, the second hop is multiplied by a decay factor of 0.8, the third hop is multiplied by a decay factor of 0.64, and so on.

[0040] In this embodiment, a negative exponential function is used to simulate the natural decay process of information in a product network. The function uses the path propagation strength value as the coefficient of the exponential term, and the calculation result represents the degree of information retention. When the propagation strength value is low, the negative exponential function value is close to zero, indicating that the information is almost completely lost; conversely, it indicates that the information is well retained and the category correlation between products is strong.

[0041] Step S103: Using the information transmission loss value, a multi-layer graph convolutional neural network is used to propagate information across multiple layers of the hop distance between products, and the semantic embedding vector of the product after each layer of propagation is determined.

[0042] Based on the information transmission loss value, graph information propagation is performed. This involves reading the initial feature vector of the product node, including the product category code, sales attributes, and user behavior features. For each product node, its one-hop neighbor node set is obtained and grouped according to the edge type: same-purchase edge, substitute edge, and complementary edge. The sampling ratio of neighbors connected by same-purchase edges is set to the highest, followed by substitute edges, and then complementary edges, resulting in neighborhood node groups distinguished by edge type.

[0043] Based on the neighboring node groups distinguished by edge type and their initial feature vectors, a graph convolutional neural network is used to aggregate features for each group of neighboring nodes. For edge groups with the same edge type, mean pooling is used to aggregate neighbor features to retain common attributes. For edge groups with different edge types, max pooling is used to extract significantly different features. For edge groups with complementary edge types, weights are calculated using an attention mechanism and then summed. The three aggregation results are concatenated to form the node representation of the current layer. For the node representation, multi-layer information propagation is performed. During each propagation, the output of the previous layer is used as the input of the current layer, expanding the neighborhood range to two-hop and three-hop nodes. The feature aggregation process is repeated. The retention rate is calculated based on the information transmission loss value. Only for high-loss paths with a loss value exceeding a preset threshold of 0.5, the retention rate is set to 1 - the loss value, and multiplied by the aggregation result of that path to obtain the attenuated feature. ; where H (l+1) i H represents the feature representation of node i in the (l+1)th layer. (l) j Let N represent the feature representation of the l-th layer neighbor node j. k(i)Let represent the set of k-hop neighbor nodes of node i, and AGG represent the feature aggregation function. This formula describes the feature aggregation mechanism in the multi-layer information propagation process, where each layer uses the output of the previous layer as the input of the current layer. The path retention rate, with a loss value not exceeding 0.5, is set to 1 and does not decay. After propagation through a preset number of layers, the semantic embeddings of each layer of goods are output. Through the semantic embeddings of each layer of goods, the cosine similarity between the target goods and indirectly related goods in the embedding space is calculated. If the similarity exceeds the strong association threshold, a high category semantic strength is assigned; if it is below the weak association threshold, a low category semantic strength is assigned; and if it is between the two thresholds, a medium category semantic strength is assigned. This determines the category semantic strength of the indirectly related goods.

[0044] In this embodiment, the initial feature vector of the product node is constructed by integrating multi-dimensional data, including the product's three-level category code, brand identifier, price range, material attributes, functional tags, and historical sales data. These features are normalized and then concatenated to form a dense vector of fixed dimensions, which serves as the input to the graph convolutional neural network. The graph information propagation unit adopts a layered architecture design: the bottom layer is responsible for sampling neighboring nodes, the middle layer performs feature aggregation operations, and the top layer outputs the updated node representation. Each propagation unit maintains three parallel processing channels, corresponding to the information flows of same-purchase edges, substitute edges, and complementary edges, respectively. Each channel is processed independently and then fused at the output layer. The propagation unit retains the original feature information through residual connections, avoiding the gradient vanishing problem during deep propagation.

[0045] In this embodiment, the neighbor node sampling process is differentiated based on the actual business characteristics of the e-commerce platform. For neighbor nodes connected by the same-purchase edge, the sampling ratio is set to 0.7, because simultaneously purchased goods often belong to the same or similar usage scenarios and have strong category correlation. The sampling ratio for alternative edges is 0.2, reflecting the competitive relationship between products; although their functions are similar, their brand positioning may differ. The sampling ratio for complementary edges is 0.1, as these products, although complementary, usually belong to different categories. During actual sampling, if the number of neighbor nodes for a certain type of edge is insufficient for the preset sampling number, a repeated sampling strategy is used to supplement it; if it exceeds the preset number, representative nodes are selected through random sampling. The sampling process also introduces a degree normalization mechanism to suppress the influence of nodes with high elevation, preventing popular products from excessively dominating the dissemination process.

[0046] In this embodiment, the selection of the three aggregation functions is based on the semantic characteristics of different edge types. Mean pooling is used for edge groups of similar products. By calculating the arithmetic mean of the features of all neighbors, it extracts the common features of the product group, which is suitable for capturing the common attributes of similar products. Max pooling is used for edge groups of alternative products. It selects the maximum value on each feature dimension to highlight the significant differences between different brands or models, which helps to identify the unique selling points of products. Attention mechanism is used for edge groups of complementary products. It dynamically allocates weights based on the relevance between the query node and its neighboring nodes. The weights are calculated through dot product attention, reflecting the differences in the strength of association between complementary products. This differentiated aggregation strategy enables the network to extract product association information from different perspectives.

[0047] In this embodiment, multi-layer information propagation expands the receptive field layer by layer in a breadth-first manner. The first layer aggregates information from one-hop neighbors, the second layer expands to two-hop neighbors, and so on. After each layer of propagation, the node representation is updated, and the new representation incorporates information from a wider range of neighborhoods. The propagation depth is typically set to 3 to 5 layers; too deep a propagation can lead to oversmoothing, causing the representations of different products to become similar.

[0048] In this embodiment, the retention rate calculation transforms the loss value from a negative indicator to a positive one. If the information transmission loss value is 0.3, the retention rate is 0.7, indicating that 70% of the original information strength is retained after path propagation. The retention rate is applied as a scaling factor to the output features of each layer. Furthermore, the cosine similarity is calculated in a normalized embedding space by dividing the inner product of two product semantic embedding vectors by their modulus product. The similarity range is [-1, 1], and in practical applications, the positive correlation in the [0, 1] interval is primarily considered. When the similarity exceeds 0.8, it indicates that the products have a strong category correlation.

[0049] In this embodiment, the three-level classification of category semantic strength corresponds to different business processing strategies. High strength indicates that products should be classified into the same sub-category for accurate recommendations; medium strength indicates that products belong to adjacent categories and are suitable for associated recommendations; low strength indicates that product categories are far apart and are usually not displayed together. This hierarchical mechanism allows e-commerce platforms to flexibly adjust product association strategies according to different scenarios.

[0050] For example, if the propagation path from athletic shoes to professional running socks goes through multiple aggregations and the final similarity reaches 0.85, it is judged as having high category semantic strength, and the platform will classify the two into the same subcategory of sports equipment for joint display.

[0051] Step S104: Based on the semantic embedding vectors of the products after propagation at each layer, extract the vector distance between the semantic embeddings of the products, evaluate the similarity, and obtain the optimized number of information propagation aggregation layers based on the similarity convergence trend.

[0052] Obtain the semantic embedding vectors of the products after propagation at each layer. Starting from the first layer, extract the multidimensional embedding / 128-dimensional embedding representation of each product node layer by layer. Calculate two metrics, Euclidean distance and cosine similarity, for a randomly sampled subset of products. If the Euclidean distance is less than the preset distance threshold of 1.0 and the cosine similarity is greater than the preset similarity threshold of 0.8, then mark the product pair as having a strong semantic association at the current level, and obtain the product pair similarity matrix at each level.

[0053] Based on the similarity matrix, the rate of change in product similarity is calculated with each additional layer of propagation depth. When the average rate of change in similarity between two consecutive layers is lower than a preset convergence threshold of 0.01, the information propagation is considered to have entered a convergence state, and the layer number at this point is recorded as a candidate aggregation depth, thus obtaining the convergence layer distribution for different product categories. Through the convergence layer distribution and the candidate aggregation depth, differentiated convergence patterns for different product categories at different levels are identified. The aggregation weight for each layer is calculated based on the initial product category label and the distribution of neighboring node categories. The aggregation weight decreases exponentially with increasing propagation depth, and the weight formula is: wl = alpha l Where wl is the weight of the l-th layer, alpha is the decay factor (its value is 0.9), and l is the layer number. The optimized information propagation aggregation layer for each product category is determined by weighted summation. Using the optimized information propagation aggregation layer, the propagation path of products that have not reached the convergence layer is forcibly truncated. For products that have converged, their embedding state at the time of convergence is retained. The truncated product embedding vectors are classified using a K-nearest neighbor classifier, k=5, with an embedding dimension d=128. The consistency between the classification result and the actual category label in the training set is calculated. If the consistency is lower than the threshold of 0.9, the aggregation layer is iteratively adjusted to obtain the final adaptive aggregation depth configuration.

[0054] In this embodiment, the extraction process of product semantic embedding vectors begins with the product association network of the e-commerce platform. Each product node initially has an original embedding representation generated based on the product title, description text, and category tag. The multidimensional embedding representation is propagated between layers through the message passing mechanism of the graph neural network. The first layer aggregates the feature information of directly adjacent products, the second layer fuses the product semantics within a two-hop neighborhood, and so on up to the fifth layer or deeper.

[0055] In this embodiment, the Euclidean distance is calculated using a standard vector space metric. For the embedding vectors of two items, the square root of the sum of the squares of the differences in each dimension is calculated. Cosine similarity is obtained by dividing the inner product of the two vectors by the product of their respective moduli, reflecting the similarity of the vectors in direction. When a pair of items simultaneously meets both the distance threshold and the similarity threshold, the system marks them as a strongly semantically related pair. This dual verification mechanism avoids misjudgments that may occur with a single metric.

[0056] In this embodiment, when processing clothing items, although a white shirt and a pair of dark trousers differ significantly in color attributes, their semantic embeddings become significantly closer after the second layer of propagation because they are often sold together. The distance matrix records the distance values ​​of all item pairs in each layer, forming a symmetric matrix structure. The diagonal elements are zero, and the off-diagonal elements reflect the semantic distance of the corresponding item pair in the current layer. By comparing the distance matrices of adjacent layers, the evolution trend of the semantic relationship of items with the depth of propagation can be observed.

[0057] In this embodiment, the calculation of the similarity change rate involves the concept of moving average in statistics. The average change in similarity between all product pairs from layer k to layer k+1 is calculated. When this change rate falls below a preset convergence threshold twice consecutively, it indicates that information propagation has stabilized, and further increasing the number of propagation layers will yield limited semantic gain. The candidate aggregation depth is the layer at which the convergence condition is first met. Different categories of products will converge at different depths due to differences in their association network structures.

[0058] In this embodiment, the identification process of differentiated convergence patterns is based on statistical analysis of a large amount of historical transaction data. Electronic products, due to their clear accessory relationships and limited substitutes, typically allow for sufficient expression of semantic relationships at a shallower level. However, clothing and accessories involve complex combinations and seasonal factors, often requiring deeper propagation to capture complete semantic information. The aggregation strategy for each product is dynamically adjusted by analyzing its initial category label distribution and the proportion of categories within its k-hop neighborhood. The calculation of the aggregation weight follows an exponential decay law, specifically, the weight of the nth layer equals the initial weight multiplied by the nth power of the decay factor. The value of the decay factor is dynamically determined based on the product category and neighborhood purity. Neighborhood purity refers to the proportion of products of the same category within the neighborhood. When neighborhood purity is high, the decay factor is set larger, allowing information to propagate at deeper levels; when the neighborhood contains a large number of dissimilar products, the decay factor is smaller, limiting the influence of long-distance information and preventing semantic pollution.

[0059] In this embodiment, the propagation path truncation mechanism implements differentiated processing based on the optimized aggregation layer number. For products that have reached their optimal aggregation depth, the system retains the embedding state at that depth as the final representation and stops propagating; for products that have not yet converged, the system forcibly stops when the preset maximum propagation depth is reached to avoid over-smoothing.

[0060] For example, a newly listed smartwatch is linked to mobile phone accessories and sports equipment in the product network. Through the first two layers of propagation, it is found that the semantic distance between it and mobile phone accessories rapidly decreases and tends to stabilize, while the association with sports equipment requires indirect paths such as fitness apps and heart rate monitoring to be fully reflected in the third layer. Based on this differentiated convergence characteristic, the aggregation depth for this product is determined to be three layers, which retains the attributes of electronic products while incorporating the semantic features of sports and health.

[0061] In this embodiment, the application of the K-nearest neighbor classifier in the category determination process enhances the robustness of the system. The classifier, based on the truncated product embedding vector, searches for the k nearest neighbor products in the embedding space and determines the category of the target product through a voting mechanism. The calculation of consistency considers not only whether the determination result matches the actual label but also evaluates the distribution of confidence scores; correct determinations with high confidence and incorrect determinations with low confidence are both considered signals of good model performance.

[0062] For example, it exhibits unique adaptability when handling seasonal goods. During the summer off-season, the network for winter down jackets is sparse, so the system automatically reduces the aggregation depth, relying primarily on the product's own characteristics for category determination. However, during the winter peak season, rich purchase associations allow the system to acquire matching information through deeper propagation, improving the accuracy of category recognition. Through this adaptive aggregation depth configuration mechanism, e-commerce platforms can dynamically adjust their processing strategies based on product characteristics and network structure when handling massive product classifications, ensuring classification accuracy while avoiding unnecessary computational overhead.

[0063] Step S105: Based on the optimized information propagation aggregation layer, perform aggregation operation on the category semantic strength, evaluate the similarity between the neighborhood aggregation result and the standard features of each commodity category, and obtain the preliminary category belonging probability.

[0064] Based on the optimized information propagation aggregation layers, the semantic embedding vector of each product at the corresponding layer is extracted. A weighted average method is used to aggregate the semantic strength of each layer, where the weight of the first layer is set to 1, and the weight of each subsequent layer decreases by a preset decay coefficient. The aggregated comprehensive semantic representation is obtained by summing the products of the embedding vectors of each layer and their corresponding weights, and then dividing by the total weights. Using this comprehensive semantic representation, the cosine similarity between it and the standard feature vectors of each product category is calculated. The standard feature vectors are obtained by averaging the historical embedding vectors of all products in that category. Each category corresponds to one standard feature vector. If the similarity exceeds a preset threshold, the category is recorded as a candidate category, resulting in a set of candidate categories and their corresponding similarity scores. Using the set of candidate categories and the similarity scores, the softmax function is used to normalize the similarity scores of each candidate category. The softmax function calculates the exponent value of each score and divides it by the sum of all score exponent values, making the sum of the probability values ​​of all candidate categories equal to 1. Based on the normalized probability distribution, the preliminary probability of each product belonging to each category is determined. The specific calculation formula is as follows: Among them, P i s represents the normalized probability of a product belonging to the i-th candidate category. i s represents the similarity score between the product and the i-th candidate category. j Let represent the similarity score between the product and the j-th candidate category, C represent the total number of candidate categories, exp represent the exponential function, and the denominator is the sum of the exponential values ​​of the similarity scores of all candidate categories; Among them, S ik α represents the initial probability of the k-th item being assigned to the i-th category. ik α represents the similarity score between the k-th item and the i-th category. im Let m represent the similarity score between the m-th item and the i-th category, M represent the total number of categories, exp represent the exponential function operation, and the denominator ensures that the sum of the probabilities of all categories is 1.

[0065] In this embodiment, the semantic strength aggregation operation is achieved through multi-level weighted fusion. The weighted averaging method is designed based on the principle of distance attenuation in information propagation. The first layer retains the original semantic features of the product itself and assigns a weight of 1. The second layer incorporates the semantic information of directly neighboring products, with the weight decreasing by an attenuation coefficient of 0.8. The weights of the third layer and deeper layers decrease sequentially by powers of this attenuation coefficient. The embedding vector of each product in each layer is multiplied by its corresponding weight, summed, and then divided by the sum of the weights of all layers to obtain the comprehensive semantic representation. This decreasing weight design reflects the characteristic that semantic information gradually dilutes with increasing propagation distance, and that nearby products have a greater impact on the determination of the target product category than distant products.

[0066] In this embodiment, the construction of standard feature vectors is based on statistical learning from historical transaction data. Historical embedding vectors of all products under each given category are extracted from the e-commerce platform database. These vectors are trained using product descriptions, image features, and user behavior data when products are listed. The arithmetic mean of all vectors within the same category is calculated dimension-wise to form the standard feature vector for that category. The calculation of cosine similarity measures the directional consistency between the aggregated semantic representation and the standard features of each category. The closer the cosine value of two vectors is to 1, the more similar their orientations are in the semantic space, and the greater the likelihood that the product belongs to that category.

[0067] In this embodiment, the application of the softmax function realizes the transformation from similarity score to probability distribution. The function calculates the natural index of the similarity score of each candidate category, and then divides it by the sum of the index values ​​of all candidate categories, so that the probability values ​​of all categories are strictly between 0 and 1 and the sum is 1, forming a distribution that meets the requirements of probability theory.

[0068] For example, after aggregation, a newly listed Bluetooth headset has a similarity of 0.85 with the standard features of the "digital accessories" category, 0.65 with the "sports equipment" category, and 0.45 with the "office supplies" category. After softmax normalization, these similarities are converted into belonging probabilities of 0.58, 0.27, and 0.15, respectively.

[0069] In this embodiment, a probability threshold filtering mechanism is also set up. When the probability of a certain category is lower than 0.05, it is directly excluded from the candidate category set, which reduces the computational burden of subsequent processing and improves the accuracy and efficiency of category determination.

[0070] Step S106: By analyzing the dominance ratio of substitute edges and complementary edges through the preliminary category assignment probability and the information transmission loss value, the sensitivity of category determination is increased when substitute edges dominate and decreased when complementary edges dominate, thus determining the product category inference result.

[0071] By using preliminary category classification probabilities, the distribution of all connection edges in the network is statistically analyzed. Substitute and complementary relationship edges are identified based on the purchase relationship attributes marked on the edges. The proportion of each type of edge to the total number of edges is calculated. At the same time, the cumulative loss value of each edge in the information transmission process is calculated to obtain the edge type dominance index.

[0072] Based on the edge type dominance index, if the proportion of substitute edges exceeds a preset threshold, the probability values ​​of each category are P-valued. 1.5The data is then normalized to enhance category differentiation, where P represents the probability value. If the proportion of complementary edges exceeds a preset threshold, all probability values ​​are logarithmically compressed to reduce inter-category differences, resulting in an adjusted category probability distribution. Using this adjusted distribution, the category with the highest probability is selected as the final category inference for the product, while the second-highest probability category is recorded as a candidate category. When the difference between the highest and second-highest probabilities is less than 0.1, a dual-category label is output to determine the product category inference result.

[0073] In this embodiment, edge type identification is based on the purchase relationship attribute labeling of edges in the product association network. Substitutional relationship edges indicate that a user chooses one product and abandons another after comparison; typical scenarios include competing purchases of similar products from different brands. Complementary relationship edges indicate that products are frequently purchased together, such as the combined sale of suits and ties. The type attribute of each edge is automatically labeled by analyzing the co-occurrence and mutual exclusion patterns of products in historical order data.

[0074] In this embodiment, the calculation of information transmission loss reflects the degree of attenuation of semantic information on different types of edges. Alternative edges, connecting competing goods, have an attenuation coefficient of 0.3 during information transmission, indicating a rapid decrease in semantic similarity; complementary edges, connecting cooperative goods, have an attenuation coefficient of 0.7, retaining more semantic information. The cumulative loss value is equal to the product of the attenuation coefficients of all edges on the path.

[0075] In this embodiment, the edge type dominance index comprehensively considers two dimensions: quantity proportion and loss degree. When the proportion of substitute edges exceeds 60% and the average loss value is higher than a preset threshold, the system determines that the product is in a highly competitive category environment, and it is necessary to improve the classification sensitivity to accurately distinguish subtle differences.

[0076] In this embodiment, the probability distribution is adjusted using a nonlinear transformation method. When substitution edges dominate, the probability value P of each category is converted to P raised to the power of 1.5 and then normalized. This exponential amplification makes high-probability categories more prominent and low-probability categories further suppressed, enhancing the distinguishability between categories. Conversely, when complementary edges dominate, a logarithmic compression formula log(1+P) is used to reduce the probability differences between categories, reflecting the characteristic that a product may simultaneously possess multiple category attributes.

[0077] For example, when classifying a sports smartwatch, if 70% of its associated network consists of alternative edges to other smartwatch brands, the system will increase its sensitivity and clearly classify it as "smart wearables" rather than the vague "sports equipment"; if 60% consists of complementary edges to sportswear and fitness equipment, the system can reduce its sensitivity and label it as both "smart wearables" and "sports equipment".

[0078] In this embodiment, the dual-category label output mechanism provides a foundation for cross-recommendation on e-commerce platforms. When a product is labeled with two categories, it can appear simultaneously in recommendation scenarios of both categories, thus expanding the product's exposure opportunities.

[0079] Step S107: Evaluate the degree of consistency between the product category inference result and the actual product category attribution, identify category inference bias, and obtain the correction direction for adaptive identification of e-commerce product categories.

[0080] The accuracy of the inferred product category is evaluated by comparing it with the actual product category classification in the e-commerce platform's preset category system. If they match, it is recorded as correct; otherwise, it is recorded as a misjudgment with specific deviation information. This yields the category recognition accuracy and deviation distribution. Based on the deviation distribution, frequently misjudged category pairs and their frequency are statistically analyzed to identify systematic deviation patterns. By increasing the difference in judgment thresholds between easily confused categories, the correction direction for adaptive e-commerce product category recognition is determined.

[0081] In this embodiment, the e-commerce platform's preset category system adopts a tree structure, subdividing from the root category to the leaf categories. When each product is listed, the operations staff marks its exact position in the category tree as its actual affiliation. The system then compares the inferred result with this actual affiliation to determine whether the classification is correct.

[0082] In this embodiment, the deviation distribution statistics cover two dimensions: misclassification direction and frequency. Misclassification direction records which category a product was incorrectly classified into another category, such as misclassifying a "power bank" as a "portable hard drive"; frequency statistics reflect the number of times this misclassification pattern occurs in the test set.

[0083] In this embodiment, the identification of systematic bias patterns is based on statistical significance testing. When the misclassification frequency of a certain class pair exceeds 5% of the total misclassifications, the system marks it as an object that needs to be optimized in a key manner, reducing confusion by increasing the distance between the two classes in the feature space.

[0084] For example, if the system detects frequent confusion between "headphones" and "speakers", it increases the weight of the sound output mode characteristics, strengthens the difference between the two in terms of wearing method, and thus improves the accuracy of differentiation.

[0085] In this embodiment, the correction also includes adjusting the feature extraction method and adding more discriminative feature dimensions for easily confused categories.

[0086] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A method for adaptive identification and semantic evaluation of e-commerce product categories, characterized in that, The method includes: extracting the path hop count and semantic type labels of associated edges between target products from a product relationship graph; extracting all shortest path sets through graph traversal to obtain the hop distance and edge type combination between products; obtaining the information transmission loss value based on the hop distance and edge type combination, according to the semantic association strength of edge types combined with the increasing attenuation rule of path hop count; using the information transmission loss value, performing multi-layer graph convolutional neural network to propagate the hop distance between products through multiple layers, determining the product semantic embedding vector after each layer of propagation; extracting the vector distance between product semantic embeddings based on the product semantic embedding vectors after each layer of propagation, evaluating the similarity level, and based on similarity... The convergence trend is optimized by increasing the number of information propagation aggregation layers. Based on the optimized number of information propagation aggregation layers, the semantic strength of the categories is aggregated, and the similarity between the neighborhood aggregation results and the standard features of each product category is evaluated to obtain the preliminary category assignment probability. Through the preliminary category assignment probability, the dominant ratio of substitution edges and complementary edges is analyzed in combination with the information transmission loss value. When substitution edges dominate, the sensitivity of category determination is increased, and when complementary edges dominate, the sensitivity of category determination is decreased to determine the product category inference result. The degree of consistency between the product category inference result and the actual product category assignment is evaluated to identify category inference bias and obtain the correction direction for adaptive recognition of e-commerce product categories.

2. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process of extracting the path hop count and semantic type labels of associated edges between target products from the product relationship graph, and extracting all shortest path sets through graph traversal to obtain the hop distance and edge type combination between products, includes: reading all product nodes and their direct connections from the product relationship graph, obtaining the semantic type label of each edge: a "same purchase" edge indicates that a user purchases a pair of products in the same order, a "substitute" edge indicates that a user returns and buys a pair of products instead, and a "complementary" edge indicates that a user purchases a pair of complementary products in separate transactions; using a breadth-first traversal algorithm to expand layer by layer from the starting product node, recording the hop count and edge sequence from the starting node to each target node; extracting the shortest path between each pair of products based on the traversal results, counting the number of "same purchase" edges, "substitute" edges, and "complementary" edges on the path, generating an edge type sequence string according to the order of edge type appearance, and determining the hop distance value and edge type combination identifier for each pair of products.

3. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process of obtaining the information transmission loss value based on the hop distance between products and the edge type combination, combined with the semantic association strength of the edge type and the path hop count increment decay rule, includes: counting the number of co-purchase edges, substitute edges, and complementary edges on each path based on the hop distance between products and the edge type combination identifier, calculating the proportion of each type of edge, and calculating the path edge type diversity index based on the distribution of the number of each edge type; reading the pre-set semantic association strength value of the edge type based on the path edge type distribution characteristics, assigning strong association weight to co-purchase edges, medium association weight to substitute edges, and weak association weight to complementary edges, multiplying the corresponding semantic association weight of each edge for each path edge sequence, and introducing a decay factor based on the path hop count to calculate the path propagation strength value; using the path propagation strength value, calculating the degree of information decay from the starting product to the target product using a negative exponential function, and determining the information transmission loss value between products based on the product of the decay degree and the path hop count.

4. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process involves using a multi-layer graph convolutional neural network to propagate information across multiple hop distances between products based on the information transmission loss value, and determining the semantic embedding vector of the product after each layer of propagation. This includes: constructing a graph information propagation unit based on the information transmission loss value; reading the initial feature vector of the product node, including product category code, sales attributes, and user behavior features; obtaining a set of one-hop neighbor nodes for each product node; grouping these nodes according to same-purchase edges, substitute edges, and complementary edges; prioritizing the sampling ratio of neighbors connected by same-purchase edges, followed by substitute edges, and then complementary edges, to obtain neighbor node groups distinguished by edge type; and then further analyzing the neighbor node groups distinguished by edge type and their initial features. The feature vector is generated by using a graph convolutional neural network to aggregate features of each group of neighboring nodes. The same-side group aggregates neighbor features through mean pooling, the alternative side group extracts significantly different features through max pooling, and the complementary side group calculates weights through an attention mechanism and then sums them by weight. The three aggregation results are concatenated to form the node representation of the current layer. For the node representation, multi-layer information propagation is performed. Each layer propagation uses the output of the previous layer as the input of the current layer to expand the neighborhood range and repeat the feature aggregation process. The retention rate is calculated based on the information transmission loss value. The retention rate is multiplied by the aggregation result to obtain the decayed features. After propagation through a preset number of layers, the semantic embedding vectors of each layer of products are output.

5. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process of extracting the vector distance between product semantic embeddings based on the product semantic embedding vectors after propagation at each layer, evaluating the similarity level, and obtaining the optimized information propagation aggregation layer based on the similarity convergence trend includes: obtaining the product semantic embedding vectors after propagation at each layer, extracting the multi-dimensional embedding representation of each product node layer by layer, calculating two metrics, Euclidean distance and cosine similarity, for each pair of products to obtain the distance matrix of product pairs at each level; based on the distance matrix, calculating the rate of change of product pair similarity with each additional layer of propagation depth, and recording the layer number as a candidate aggregation depth when the average rate of change of similarity between two consecutive layers is lower than a preset convergence threshold; identifying the convergence differentiation pattern of different product category levels through the convergence layer distribution and the candidate aggregation depth, calculating the aggregation weight of each layer based on the initial product category label and the distribution of neighboring node categories, and determining the optimized information propagation aggregation layer for each product category.

6. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The step of aggregating category semantic strengths based on the optimized information propagation aggregation layer, evaluating the similarity between the neighborhood aggregation results and the standard features of each product category, and obtaining preliminary category assignment probabilities includes: extracting the semantic embedding vector of each product at the corresponding level based on the optimized information propagation aggregation layer; aggregating the semantic strength of each layer using a weighted average method, with the first layer having a weight of 1, and the weight of each subsequent layer decreasing by a preset decay coefficient; obtaining the aggregated comprehensive semantic representation by summing the products of each layer's embedding vector and its corresponding weight, and then dividing by the total weight; calculating the cosine similarity between the comprehensive semantic representation and the standard feature vector of each product category, where the standard feature vector is obtained by averaging the historical embedding vectors of all products under that category, with each category corresponding to one standard feature vector; obtaining a candidate category set and corresponding similarity scores; using the candidate category set and similarity scores, normalizing the similarity scores of each candidate category using the softmax function, so that the sum of the probability values ​​of all candidate categories is 1; and determining the preliminary category assignment probability of each product based on the normalized probability distribution.

7. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process involves analyzing the dominance of substitute and complementary edges based on preliminary category assignment probabilities and information transmission loss values. When substitute edges dominate, the sensitivity for category determination is increased; when complementary edges dominate, the sensitivity is decreased, thus determining the product category inference result. This includes: statistically analyzing the distribution of all connection edge types in the product association network using the preliminary category assignment probabilities; calculating the proportion of substitute and complementary edges to obtain an edge type dominance index; based on the edge type dominance index, if the proportion of substitute edges exceeds a preset threshold, exponentially amplifying the probability values ​​of each category by the difference from the highest probability; if the proportion of complementary edges exceeds a preset threshold, logarithmically compressing all probability values ​​to obtain an adjusted category probability distribution; and using the adjusted category probability distribution, selecting the category with the highest probability as the product category inference result.

8. The method for adaptive identification and semantic evaluation of e-commerce product categories according to claim 1, characterized in that, The process of evaluating the degree of consistency between the inferred product category and the actual product category attribution, identifying category inference bias, and obtaining the correction direction for adaptive e-commerce product category recognition includes: evaluating the degree of consistency between the inferred product category and the actual product category attribution, comparing the inferred category with the actual attribution of the product in the preset category system of the e-commerce platform, and obtaining the category recognition accuracy and bias distribution; based on the bias distribution, statistically analyzing frequently misjudged category pairs and their frequency of occurrence, identifying systematic bias patterns, and determining the correction direction for adaptive e-commerce product category recognition by increasing the difference in judgment thresholds between easily confused categories.

Images