A commodity implicit association strength calculation method based on multi-scale feature fusion

Abstract

The application discloses a commodity implicit association strength calculation method based on multi-scale feature fusion, belongs to the technical field of commodity data analysis, and aims to solve the problems of single feature utilization, rough multi-scale feature fusion and insufficient capture of commodity implicit association in the prior art; the application carries out preprocessing on collected commodity multi-dimensional original data, extracts features of three scales of commodities, i.e., fine, medium and coarse, and converts the features into preset dimension feature vectors, after feature optimization, completes deep feature fusion through a hierarchical mode of same-scale fusion-cross-scale attention weighted fusion, then calculates the association strength between commodities and determines the association type, finally outputs the result and completes algorithm closed-loop optimization; the application significantly improves the calculation precision and recall rate of commodity implicit association, takes into account the result explainability and real-time performance, and is suitable for the application requirements of multiple scenes of e-commerce.

Description

Technical Field

[0001] This invention belongs to the field of commodity data analysis technology, specifically involving a method for calculating the strength of implicit associations of commodities based on multi-scale feature fusion. Background Technology

[0002] Product association strength calculation is a key technology for e-commerce platforms to realize core businesses such as product recommendation, scenario-based shopping guide, and cross-platform price comparison. Its calculation accuracy directly affects the platform's conversion efficiency and user shopping experience.

[0003] Current product association strength calculation techniques are mainly divided into three categories: (1) Computational techniques based on traditional association rules: Taking the Apriori algorithm and FP-Growth algorithm as the core, by mining the co-occurrence frequency of products in the user's shopping basket, the support and confidence between products are calculated as the association strength indicators. The focus is on capturing the explicit co-occurrence association between products. It is suitable for simple product combination recommendation scenarios and is the mainstream technical means of early e-commerce association analysis. (2) Semantic association calculation technology based on single feature: relying on natural language processing or computer vision technology, extract single-dimensional features of the product to calculate the association strength. Among them, the text dimension often uses models such as BERT and Word2Vec to convert the text information such as product title and description into feature vectors, and calculate the association strength through cosine similarity. The image dimension often uses models such as ResNet to extract the main image features of the product to achieve visual association matching. Some solutions will combine full-text search tools such as Elasticsearch to help improve the matching accuracy. (3) Fusion computing technology based on simple feature splicing: Some solutions attempt to simply splice the text features of goods and user behavior features, and then calculate the association strength through traditional machine learning models or simple deep learning models. A few solutions introduce vector databases to store feature vectors to improve retrieval efficiency, but do not perform deep fusion of features of different scales and types, and still belong to the category of shallow fusion.

[0004] In addition, existing technologies also include correlation analysis schemes that combine offline batch processing and real-time stream computing, as well as end-to-end correlation mining models based on graph neural networks. However, the former is difficult to balance the needs of historical data mining and real-time response, while the latter suffers from black-box characteristics that lead to insufficient interpretability. Neither of these solutions has solved the problem of effective fusion of multi-scale features.

[0005] Based on the existing technological implementation methods, its shortcomings are mainly reflected in the following aspects: 1. Single feature utilization and insufficient capture of implicit associations: Traditional association rule algorithms rely only on the co-occurrence frequency of products and cannot capture implicit associations in non-co-occurrence scenarios; single feature calculation schemes can only mine associations from a single dimension and ignore the synergistic effect of multi-dimensional features of products, resulting in one-sided association strength calculations that cannot cover implicit association scenarios, especially with low association recall rates for mid- and long-tail products.

[0006] 2. The feature fusion method is crude and does not consider multi-scale differences: The simple feature splicing scheme does not distinguish the scale differences of different features, nor does it perform weight allocation and deep fusion of features, resulting in feature redundancy, dilution of useful information, failure to highlight core related features, and easy occurrence of feature aliasing effect after fusion, affecting the accuracy of association calculation. At the same time, it is difficult to adapt to the characteristic differences of different product categories.

[0007] 3. Low calculation accuracy and weak anti-interference ability: Existing solutions do not perform effective preprocessing and noise filtering on the original features. Redundant information and abnormal data in the product features will seriously affect the association strength calculation results. At the same time, the dynamic changes of features are not considered, resulting in the association strength calculation results being lagging behind. They cannot meet the needs of real-time scenarios such as e-commerce promotions. Some solutions even have a query delay of more than 500ms, which cannot support high-concurrency requests. Summary of the Invention

[0008] To address the problems mentioned in the background art, this invention provides a method for calculating the strength of implicit associations of commodities based on multi-scale feature fusion, which solves the problems of existing technologies such as single feature utilization, insufficient capture of implicit associations, coarse feature fusion methods, failure to consider multi-scale differences, low calculation accuracy, and weak anti-interference ability.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for calculating the strength of implicit associations among commodities based on multi-scale feature fusion includes the following steps: S1: Data Acquisition and Preprocessing; Collect multi-dimensional raw data of commodities, and perform missing value imputation, outlier removal, and numerical standardization on the raw data in sequence. Then, store the different types of raw data into the corresponding databases to obtain standardized datasets. S2: Multi-scale feature extraction; Based on a standardized dataset, features at three scales—fine-grained, medium-grained, and coarse-grained—are extracted from the product, and each scale feature is converted into a feature vector of a preset dimension. S3: Feature optimization; The feature vectors of the three scales obtained in S2 are subjected to secondary normalization, noise filtering, redundant feature removal and category-adaptive feature selection in sequence to obtain the optimized multi-scale feature vectors. S4: Multi-scale feature fusion; First, the optimized feature vectors at each scale are fused within the same scale to obtain the fused feature vectors at the corresponding scale. Then, the fused feature vectors at the three scales are deeply fused through cross-scale attention weighted fusion to obtain the global fused feature vector of the product. S5: Calculation of latent association strength; Based on the global fusion feature vector of the product to be calculated, the basic association strength between products is calculated, and then a dynamic adjustment factor is introduced to optimize the basic association strength to obtain the final association strength. The association type between products is determined according to the final association strength and the contribution of features at each scale. S6: Results Output and Feedback; Based on the final association strength, association type, and feature contribution at each scale obtained in S5, generate and output the association strength calculation results; Manually review the calculation results where the association strength reaches the preset review threshold; Based on the review results and association effect evaluation data, perform incremental training and closed-loop optimization on the parameters of feature extraction, feature fusion, and association strength calculation corresponding to this method.

[0010] Compared with the prior art, the beneficial effects of the present invention are: 1. This application achieves comprehensive extraction of multi-dimensional and multi-scale features, solving the problem of insufficient capture of single and implicit associations by existing technologies. Through feature extraction at three scales—fine, medium, and coarse—this application comprehensively covers multi-dimensional information such as product attributes, text, images, user behavior, and scene categories. It can capture functional associations at the detailed level of products, as well as potential associations at the user behavior level and generalized associations at the scene category level, significantly improving the association recall rate of mid-to-long-tail products.

[0011] 2. This application adopts a hierarchical fusion architecture of "same-scale fusion - cross-scale attention-weighted fusion," which solves the problems of coarse feature fusion and failure to consider multi-scale differences in existing technologies. This application uses an adaptive fusion method for the characteristics of features at different scales, dynamically allocating the weights of features at different scales through an attention mechanism. This avoids feature aliasing and information dilution, effectively highlighting core related features, adapting to the characteristic differences of different product categories, and significantly improving the accuracy of association calculation.

[0012] 3. The entire process of data and feature optimization, combined with an improved association strength calculation algorithm, solves the problems of low calculation accuracy and weak anti-interference ability in existing technologies. This application effectively removes interfering data through data preprocessing, feature secondary normalization, noise filtering, and redundant feature deletion; the improved cosine similarity algorithm compensates for the shortcomings of traditional algorithms in ignoring vector amplitude differences, and the dynamic adjustment factor ensures the timeliness of the calculation results, making it suitable for high-concurrency real-time scenarios such as e-commerce promotions. Attached Figure Description

[0013] Figure 1This is a flowchart illustrating the process of this application. Detailed Implementation

[0014] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0015] A method for calculating the strength of implicit associations among commodities based on multi-scale feature fusion, such as... Figure 1 As shown, it includes the following steps: Step 1: Data Acquisition and Preprocessing; The core of this step is to acquire comprehensive product-related data and perform standardization processing to provide high-quality data support for subsequent feature extraction. This process is divided into four stages: (1) Data collection: Collect multi-dimensional raw data of products to ensure comprehensive data sources, including basic product data, user behavior data, scene tag data, and cross-platform data.

[0016] (2) Missing value filling: Various missing data are filled in a targeted manner. Missing values ​​of product attributes are filled with the mean / mode of product category, missing values ​​of user behavior data are filled with 0, and missing text data are filled with product title expansion.

[0017] (3) Outlier Removal and Data Standardization: Outlier data is removed, including clicks / browsing durations exceeding 3 times the category average in user behavior data, product prices far below 10 times the category average, and blurry images with a resolution below 300 dpi; a min-max normalization algorithm is used to normalize all numerical data to the [0,1] interval, using the following formula: x norm =(xx min ) / (x max -x min ); Among them, X norm The result is the normalized value, falling within the interval [0,1]. X is the original value to be normalized. min X is the minimum value among all values ​​in this field. max This is the maximum value among all values ​​in this field.

[0018] (4) Data storage: MySQL is used to store structured data, Milvus vector database is used to store feature vectors, and Elasticsearch is used to store text information, thus completing the data preprocessing loop and obtaining a standardized dataset.

[0019] Step 2: Multi-scale feature extraction; Based on the standardized dataset preprocessed in step 1, features at three scales (fine, medium, and coarse) are extracted from the products. All features are ultimately converted into 128-dimensional feature vectors, which is divided into three steps: (1) Fine-grained feature extraction: First, the key attributes of the product are extracted. One-hot encoding is used to convert discrete attributes into vectors, and the normalized values ​​of continuous attributes are directly used. The vectors are concatenated to obtain a 128-dimensional attribute feature vector. Then, the product title and description are segmented by jieba to remove stop words and extract the top 50 high-frequency keywords. The keywords are converted into vectors by the Word2Vec model and mean pooled. The Word2Vec model has a window size of 5 and an iteration count of 100 to obtain a 128-dimensional text keyword feature vector. At the same time, the ResNet50 model is used to extract local features of the product main image. After local pooling, a 128-dimensional image local feature vector is obtained. The three types of fine-grained feature vectors are concatenated and L2 normalized to obtain a 384-dimensional fine-grained feature vector.

[0020] (2) Medium-granularity feature extraction: Calculate the statistical features of four dimensions: product click conversion rate, add-to-cart rate, purchase rate, and co-occurrence frequency. After normalization, a 4-dimensional statistical feature vector is obtained. Calculate the jump probability between products based on user behavior data, extract the jump probability of the Top 30 related products, and obtain a 30-dimensional preference association feature vector. Use the global pooling layer of the ResNet50 model to extract the global features of the product main image, and obtain a 128-dimensional image global feature vector. Concatenate the three types of medium-granularity feature vectors and normalize them with L2, and add zero vectors to make it 128-dimensional.

[0021] (3) Coarse-grained feature extraction: One-hot encoding is used to convert scene, season and festival tags into 64-dimensional scene feature vectors and third-level categories into 64-dimensional category feature vectors; the two types of feature vectors are concatenated and L2 normalized to obtain 128-dimensional coarse-grained feature vectors.

[0022] During feature extraction, Kafka is used to buffer feature data and write it in batches to the Milvus vector library. The batch write latency is ≤50ms and the write TPS is ≥2000 to ensure real-time performance.

[0023] Step 3: Feature optimization; The feature vectors extracted in step 2 at the fine, medium, and coarse scales are optimized to remove redundancy and noise and improve feature quality. This process consists of four steps: (1) Secondary normalization: Perform L2 normalization on all feature vectors again to ensure that the magnitude is 1, so as to avoid the influence of numerical differences on subsequent fusion. The formula is: v norm =v / ||v||; Where v is the original feature vector, and ||v|| is the L2 norm of vector v.

[0024] (2) Noise filtering: Gaussian filtering algorithm is used to filter out abnormal feature points, where the standard deviation σ = 0.5, the window size = 3 × 3, and the calculation formula is: G(x,y)=(1 / (2πσ²))×e^(-(x²+y²) / (2σ²)); Where G(x,y) is the Gaussian filter weight value at coordinates (x,y), σ is the standard deviation of the Gaussian filter, which controls the smoothing intensity of the filter, and x,y are the pixel coordinate offsets within the filter window.

[0025] (3) Redundant feature deletion: The correlation between features within the same scale is calculated using the Pearson correlation coefficient. If the correlation coefficient between two features is ≥0.85, they are considered redundant features, and features with smaller variance and simpler information are deleted. Cross-scale features are not subject to redundancy deletion to ensure the integrity of multi-scale information.

[0026] (4) Feature selection: A category adaptation mechanism is introduced. The importance of features is calculated through mutual information. The top 80% of the core features with mutual information values ​​under each product category are retained to further improve the feature targeting and obtain the optimized multi-scale feature vector.

[0027] Step 4: Multi-scale feature fusion; A hierarchical approach of "same-scale fusion - cross-scale attention-weighted fusion" is adopted to achieve deep fusion of multi-scale features, avoiding feature overlap and redundancy. The specific operation is as follows: (1) Same-scale feature fusion: Fine-grained feature fusion: The attributes, text keywords, and local image feature vectors are concatenated into a 384-dimensional vector, which is then compressed into a 128-dimensional fine-grained fusion feature vector through a 1×1 convolutional layer with 128 convolutional channels. Medium-granularity feature fusion: A weighted fusion method is adopted, with user behavior statistical features weighted at 0.4, user preference association features weighted at 0.3, and global image features weighted at 0.3. The weighted sum is used to obtain a 128-dimensional medium-granularity fusion feature vector. Coarse-grained feature fusion: The scene label and category feature vector are concatenated to obtain a 128-dimensional coarse-grained fused feature vector.

[0028] (2) Cross-scale attention-weighted fusion: Vector concatenation: Concatenate the 128-dimensional fused feature vectors from three scales to obtain a 384-dimensional temporary feature vector; Pooling and concatenation: Perform global average pooling and max pooling operations on the temporary feature vector to obtain a 1×128 feature vector, and then concatenate the two vectors to obtain a 1×256 feature vector. Weight Calculation: By transforming two fully connected layers with non-shared weights and substituting them into the sigmoid activation function, three attention weights w1, w2, and w3 are obtained. The sum of these weights is 1. The calculation formula is as follows: w=σ(F2(F1([avg pool (v),max pool (v)]))); Where w is the final output feature weight vector, v is the input original feature vector, and avg pool (v) represents the average pooling (average value) operation on the feature vector v, max pool (v) represents max pooling (taking the maximum value) of the feature vector v, [avg pool (v),max pool [v] concatenates the average pooling features with the max pooling features. F1 / F2 are two fully connected layers / linear transformation layers (used for feature learning). σ is the activation function (such as Sigmoid or ReLU, used for non-linear transformation). F2(...) is the output after the second transformation layer, and σ(...) is the output after activation.

[0029] Weighted fusion: Based on the attention weights, the fusion feature vectors of the three scales are weighted and summed to obtain a 128-dimensional global fusion feature vector. The formula is as follows: V global =w1×V fine +w2×V mid +w3×V coarse ; Vglobal: Global fusion feature vector of the product (ultimately used to calculate the association strength); w1: Attention weights for fine-grained features; w2: Attention weights for medium-granularity features; w3: Attention weights for coarse-grained features; V fine : Optimized fine-grained fused feature vector; V mid : Optimized medium-granularity fused feature vector; V coarse : Optimized coarse-grained fused feature vector; Fusion optimization: Subpixel convolutional upsampling optimization is adopted, and a subpixel context enhancement unit is introduced. Context features are extracted and aggregated through three branches: local, large receptive field, and global, to obtain the final global fusion feature vector.

[0030] The aforementioned attention weights can be dynamically adjusted based on product categories and application scenarios to improve integration and adaptability.

[0031] Step 5: Calculate the strength of the latent association; Based on the global fusion feature vector obtained in step 4, the strength of the implicit association between products is calculated. Combined with the results of dynamic factor optimization, the association type is clarified, which is divided into 3 steps: (1) Basic association strength calculation: An improved cosine similarity algorithm is adopted to overcome the deficiency of traditional algorithms in ignoring vector magnitude differences. The formula is as follows: Sim(A,B)=(V A ·V B ) / (||V A ||×||V B ||)×α+(min(||V A ||,||V B ||) / max(||V A ||,||V B ||))×(1-α); Where Sim(A,B) represents the basic association strength between goods A and B, and V A V B V is the globally fused feature vector of the two. A ·V B For the vector dot product, ||V A ||、||V B || represents the L2 norm of the vector, α = 0.7, and the latter part is the normalized ratio of the vector magnitude.

[0032] (2) Dynamic optimization and adjustment: Two dynamic adjustment factors are introduced to optimize the basic correlation strength and obtain the final correlation strength: Score(A,B)=Sim(A,B)×(β×F hot +(1-β)×F prefer ); Where β = 0.4, F hot The product popularity factor is calculated as follows: Current product's 7-day click-through rate + purchases / category's average 7-day click-through rate + purchases; F prefer The user preference factor is calculated as follows: 1 - | Difference in current user preference for products A and B|; Popularity factors are updated every 24 hours, while preference factors are updated in real time to ensure timeliness.

[0033] (3) Association type determination: The association type is determined based on the association strength and feature contribution. The specific criteria are as follows: Score(A,B)≥0.7 and fine-grained feature contribution≥50% indicates functional complementarity association; Score(A,B)≥0.6 and coarse-grained feature contribution≥40% indicates scene association; Score(A,B)≥0.5 and medium-grained feature contribution≥40% indicates potential substitution association; Score(A,B)<0.5 indicates no obvious association. The feature contribution is calculated as the attention weight of that scale × the contribution value of the feature to the similarity at that scale.

[0034] Step 6: Results Output and Feedback; The algorithm outputs the results of the correlation strength calculation between products, and continuously optimizes its performance through a feedback mechanism. This process is divided into four stages: (1) Output of results: The output includes “product AID-product BID-latent association strength value-association type-feature contribution of each scale”, and outputs the confidence level and difference points of the association results to assist manual review; the output results are adapted to different application scenarios. The cross-platform price comparison scenario focuses on outputting substitution association and price association, the scenario-based shopping guide scenario focuses on outputting scenario and complementary association, and the new product recommendation scenario prioritizes category and scenario features, shortening the cold start delay to within 24 hours.

[0035] (2) Manual review: Results with a correlation strength of ≥0.6 are pushed to the manual review queue. The review SLA is set to ≤5 minutes and the daily review volume is ≤150,000. Results that pass the review are directly applied to the actual scenario, while results that fail the review are retained for subsequent algorithm optimization.

[0036] (3) Closed-loop optimization: Every 7 days, the accuracy and recall rate are used to evaluate the association calculation results. If the accuracy is lower than 90%, the feature extraction weight, attention weight and dynamic adjustment factor are readjusted. At the same time, manually corrected samples are stored in the feedback library. The feature extraction model is incrementally trained daily using the PyTorchLightning framework with batch_size=1024 and max_epochs=1 to continuously improve the calculation accuracy.

[0037] (4) Scenario adaptation: The output results are adapted to different application scenarios. The cross-platform price comparison scenario focuses on outputting substitution and price associations. The scenario-based shopping guide scenario focuses on outputting scenario and complementary associations. The new product recommendation scenario prioritizes category and scenario features, shortening the cold start delay to within 24 hours.

[0038] Those skilled in the art should understand that the above embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Any reasonable modifications, equivalent substitutions, or adaptive improvements made based on the technical concepts disclosed in this invention without departing from the essential spirit of the invention should be considered to fall within the scope of protection defined by the claims of this invention.

Claims

1. A method for calculating the strength of implicit associations among commodities based on multi-scale feature fusion, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing; Collect multi-dimensional raw data of commodities, and perform missing value imputation, outlier removal, and numerical standardization on the raw data in sequence. Then, store the different types of raw data into the corresponding databases to obtain standardized datasets. S2: Multi-scale feature extraction; Based on a standardized dataset, features at three scales—fine-grained, medium-grained, and coarse-grained—are extracted from the product, and each scale feature is converted into a feature vector of a preset dimension. S3: Feature optimization; The feature vectors of the three scales obtained in S2 are subjected to secondary normalization, noise filtering, redundant feature removal and category-adaptive feature selection in sequence to obtain the optimized multi-scale feature vectors. S4: Multi-scale feature fusion; First, the optimized feature vectors at each scale are fused within the same scale to obtain the fused feature vectors at the corresponding scale. Then, the fused feature vectors at the three scales are deeply fused through cross-scale attention weighted fusion to obtain the global fused feature vector of the product. S5: Calculation of latent association strength; Based on the global fusion feature vector of the product to be calculated, the basic association strength between products is calculated, and then a dynamic adjustment factor is introduced to optimize the basic association strength to obtain the final association strength. The association type between products is determined according to the final association strength and the contribution of features at each scale. S6: Results Output and Feedback; Based on the final association strength, association type, and feature contribution at each scale obtained in S5, generate and output the association strength calculation results; Manually review the calculation results where the association strength reaches the preset review threshold; Based on the review results and association effect evaluation data, perform incremental training and closed-loop optimization on the parameters of feature extraction, feature fusion, and association strength calculation corresponding to this method.

2. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 1, characterized in that, In S1, the collected raw data of products in multiple dimensions includes basic product data, user behavior data, scene tag data, and cross-platform data.

3. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 2, characterized in that, In S2, fine-grained features include product attribute features, text keyword features, and image local features; Medium-granularity features include user behavior statistical features, user preference correlation features, and global image features; Coarse-grained features include scene tag features and product category features; During feature extraction, a message queue is used to buffer feature data and write it to the vector database in batches.

4. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 3, characterized in that, S3 specifically includes: Perform L2 quadratic normalization on all feature vectors to ensure that the feature vector magnitude is 1; A Gaussian filtering algorithm is used to filter out outlier feature points in the feature vector. The correlation between feature vectors within the same scale is calculated using the Pearson correlation coefficient. If the correlation coefficient between two feature vectors reaches a preset correlation coefficient threshold, the redundant feature vector with smaller variance is deleted. No redundancy deletion is performed on cross-scale feature vectors. The importance of each feature vector is calculated using mutual information, and the core feature vectors with the highest mutual information values ​​under each product category are retained according to a predetermined ratio.

5. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 4, characterized in that, Same-scale feature fusion in S4 specifically includes: Fine-grained feature fusion: attributes, text keywords, and local image feature vectors are concatenated and compressed into a fine-grained fused feature vector of a preset dimension through a convolutional layer; Medium-granularity feature fusion: A weighted fusion method with preset weight allocation is used to sum the user behavior statistical features, user preference association features, and global image features in a weighted manner to obtain a medium-granularity fusion feature vector of preset dimensions; Coarse-grained feature fusion: The scene label and the category feature vector are concatenated to obtain a coarse-grained fused feature vector of a preset dimension.

6. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 5, characterized in that, Cross-scale attention-weighted fusion in S4 specifically includes: By concatenating the fused feature vectors of the three scales and preset dimensions, a temporary feature vector is obtained; Perform global average pooling and max pooling operations on the temporary feature vectors respectively, and concatenate the two pooled vectors to obtain the concatenated feature vector. After transformation through a fully connected layer, three attention weights with a sum of 1 are obtained through an activation function. The fusion feature vectors of the three scales are then weighted and summed according to the attention weights to obtain a global fusion feature vector of a preset dimension. Subpixel convolutional upsampling is used to optimize the global fusion feature vector. The final global fusion feature vector is obtained through multi-branch context feature extraction and aggregation.

7. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 6, characterized in that, In S5, an improved cosine similarity algorithm is used to calculate the basic association strength between products. The calculation formula is as follows: Sim(A,B)=(V A ·V B ) / (||V A ||×||V B ||)×α+(min(||V A ||,||V B ||) / max(||V A ||,||V B ||))×(1-α); Where Sim(A,B) represents the basic association strength between product A and product B, and V A V B V represents the global fusion feature vectors of product A and product B, respectively. A ·V B For the vector dot product, ||V A ||、||V B || represents the L2 norm of the vector, and α is the first preset weight coefficient.

8. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 7, characterized in that, In S5, the formula for calculating the final association strength is: Score(A,B)=Sim(A,B)×(β×F hot +(1-β)×F prefer ); Where Score(A,B) represents the final association strength between product A and product B, β is the second preset weighting coefficient, and F hot F is the product popularity factor. prefer The association is determined based on user preference factors and the preset association strength threshold and feature contribution threshold at each scale. The association types include functional complementary association, scenario association, potential substitution association, and no obvious association.

9. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 8, characterized in that, In S6, the output of the product association strength calculation results includes the unique identifier of the product to be associated, the final association strength value, the association type, and the contribution of features at each scale. It also outputs the confidence level of the association results and the feature difference points. The calculation results of the association strength reaching the preset review threshold are pushed to the manual review queue, and the corresponding review time limit and daily review volume limit are set.

10. The method for calculating the implicit association strength of commodities based on multi-scale feature fusion according to claim 9, characterized in that, In S6, the accuracy and recall metrics are used to evaluate the results of the product association strength calculation according to the preset evaluation cycle. If the accuracy is lower than the preset precision threshold, the feature extraction weight, feature fusion attention weight and the dynamic adjustment factor of association strength calculation are readjusted. The manually reviewed and corrected samples are stored in the feedback library, and the model used for feature extraction is incrementally trained according to the preset training cycle.

Images