SKU intelligent classification and label generation method fusing visual pre-training model

Abstract

The invention discloses an SKU intelligent classification and label generation method fused with a visual pre-training model. The method comprises the following steps: obtaining image data and text data of an SKU; extracting multi-modal features based on the image data and the text data; according to the multi-modal features, zero sample classification of the SKU is realized; generating an SKU label through an improved fractal condition variational auto-encoder based on a multi-modal feature; on the basis of newly added SKU data, continuous learning of the model is realized; carrying out compression and mobile terminal deployment on the model; through an innovative fractal attention encoder and a chaotic dynamics condition generation decoder, the classification accuracy and the label generation quality are remarkably superior to those of an existing method, the BLEU scores of 95.8% and 0.76 are achieved respectively, commodities can be classified and described more accurately, and the search and recommendation effects of a platform are improved; and due to the continuous learning ability and the dynamic category extension mechanism, the performance is excellent when emerging commodity categories are processed, and the recognition rate reaches up to 89.2%.

Description

Technical Field

[0001] The present invention relates to the technical field of SKU intelligent classification and label generation methods, and more specifically, to a SKU intelligent classification and label generation method integrating a visual pre-training model. Background Art

[0002] With the rapid development of e-commerce, the number of SKUs (Stock Keeping Units) has grown exponentially. Large e-commerce platforms need to process millions of new SKUs every day, which poses a huge challenge to SKU classification and label generation. Accurate and efficient SKU management can not only improve the platform's operational efficiency, but also significantly improve users' shopping experience.

[0003] The development status of existing technologies mainly includes the following stages:

[0004] 1. Rule-based method: This is the earliest SKU management method, which mainly relies on predefined rules and keyword matching. Although it is simple to implement and fast to calculate, its accuracy and flexibility are low. With the increase in SKU types and complexity, this method has been unable to meet the needs of modern e-commerce platforms.

[0005] 2. Traditional machine learning methods: such as support vector machine (SVM) for classification, latent Dirichlet allocation (LDA) for label generation, etc. These methods have improved the level of automation and accuracy to a certain extent, but have limited ability to process multimodal data such as images and texts, and are difficult to adapt to rapidly changing product categories.

[0006] 3. Deep learning methods: In recent years, deep learning-based methods, such as using convolutional neural networks (CNN) to process images and using pre-trained models such as BERT to process text, have been widely used in SKU management. These methods have greatly improved classification accuracy and label generation quality, but there are still some key issues.

[0007] The main problems of the prior art are as follows:

[0008] 1. Insufficient fusion of multimodal information: Existing methods often process image and text information separately and then simply concatenate the features. This approach makes it difficult to fully utilize the complementary information between different modalities.

[0009] 2. Lack of multi-scale understanding of SKU features: Product features often exist at multiple levels, from broad categories to detailed attributes. Existing methods are difficult to capture information at these different scales at the same time, resulting in limited accuracy in classification and label generation.

[0010] 3. Poor adaptability to new categories: New categories and new concepts emerge in an endless stream in the e-commerce field, but most existing models need to be retrained to adapt to new categories and cannot meet real-time requirements.

[0011] 4. Low quality of label generation: Labels generated by existing methods are often stereotyped, lacking diversity and detailed descriptions, and are difficult to accurately reflect the unique characteristics of the products.

[0012] 5. Difficulty in model deployment: Many high-performance deep learning models are large in size and difficult to deploy in resource-constrained environments such as mobile terminals, which limits their practical application scope.

[0013] Taking the deep learning method based on ResNet50 and BERT as an example, although its classification accuracy reached 92.1%, its adaptability to new categories was only 62.8%, and the model size was as high as 230MB, making it difficult to deploy on mobile devices. This method performs poorly when dealing with emerging product categories and cannot meet the needs of the rapidly changing e-commerce environment. At the same time, its large model size also limits its application in diverse scenarios.

[0014] In view of the above problems, the present invention proposes a SKU intelligent classification and label generation method integrating a visual pre-training model. The method comprises the following steps: obtaining image data and text data of SKU; extracting multimodal features based on the image data and the text data; realizing zero-sample classification of SKU according to the multimodal features; generating SKU labels through an improved fractal conditional variational autoencoder based on the multimodal features; realizing continuous learning of the model based on the newly added SKU data; compressing the model and deploying it on a mobile terminal. Summary of the invention

[0015] The present invention aims at the above problems and proposes a SKU intelligent classification and label generation method and system integrating a visual pre-training model. The method integrates cutting-edge technologies such as quantum computing, fractal theory and deep learning, and achieves high-precision emotion recognition, personalized music generation and adaptive adjustment strategy optimization.

[0016] The present invention provides a SKU intelligent classification and label generation method integrating a visual pre-training model, comprising the following steps:

[0017] Obtain image data and text data of the SKU; extract multimodal features based on the image data and the text data; implement zero-sample classification of the SKU based on the multimodal features; generate SKU labels through an improved fractal conditional variational autoencoder based on the multimodal features; implement continuous learning of the model based on newly added SKU data; and compress the model and deploy it on a mobile terminal.

[0018] Specifically, the step of extracting multimodal features includes:

[0019] Using the pre-trained CLIP visual encoder, extract the image feature vector V from the image data img′ ; Using the pre-trained CLIP text encoder, extract the text feature vector V from the text data txt ; Through the adaptive weighted fusion strategy, the fusion feature vector V is obtained fused ,in:

[0020] V fused =α·V img +(1-α)·V txt

[0021] α=σ(W α ·[V img ; V txt ]+b α )

[0022] Among them, σ is the sigmoid function, W α and b α are learnable parameters.

[0023] Specifically, the steps of zero-sample classification include: constructing a category prototype

[0024] Among them, w i is the sample weight, which is calculated by the sample quality assessment module Q: Calculate the similarity between the features of the SKU to be classified and the prototypes of each category:

[0025]

[0026] Among them, ε is the smoothing factor, β is the learnable temperature parameter, and the SKU is classified into the category with the highest similarity: c * = argmax c sin(V fused , P c ).

[0027] Specifically, the improved fractal conditional variational autoencoder includes a fractal attention encoder, a chaotic dynamics conditional generation decoder and a quantum annealing adversarial training module.

[0028] Specifically, the calculation process of the fractal attention encoder is as follows:

[0029] h=x

[0030] For l in the range L:

[0031] h sub =[h]

[0032] For s in range S:

[0033] h s =sub u nit(h,s)

[0034] h sub .append(h s )

[0035]

[0036] Return h

[0037] h fractal =fractal attention (x, L = 3, S = 4)

[0038]

[0039] μlogσ=MLP(h fused )

[0040] Among them, L is the number of fractal layers, S is the number of subunits in each layer, α i is the attention weight for adaptive learning, x is the input feature, h is the intermediate feature, Wq, W k , W v is the learnable weight matrix, d k is the scaling factor of the attention mechanism, μ and logσ are the posterior distribution parameters of the latent variable z, and h sub It takes the natural logarithm of each element of the vector sub.

[0041] Specifically, the calculation process of the chaotic dynamics condition generation decoder is as follows:

[0042] Z tilde =Reparameterize(μ,σ)

[0043] Z chaos =logistic map (z tilde , r = 3.9)

[0044] h coupled =concat[z chaos , c]

[0045] h lorenz =odeint(lorenz rnn,h coupled , t, args = (10, 28, 8 / 3))

[0046] y pred =MLP(hlorenz [-1])

[0047] Among them, z tilde is the latent variable obtained by reparameterization sampling, c is the conditional vector, r is the parameter σ of the Logistic mapping, ρ, β are the parameters of the Lorenz system, t is the time step sequence, y pred is the generated label probability distribution.

[0048] Specifically, the loss function of the quantum annealing adversarial training module is as follows:

[0049] L G =BCE(D(G(z,c)),1)+λ1KL(q(z|x)||p(z))+λ2CE(y pred ,y true )+λ3L fractal_contrast

[0050] L D =BCE(D(y true ), 1) + BCE(D(G(z,c)), 0)

[0051] Among them, BCE is the binary cross entropy loss, KL is the KL divergence, and CE is the classification cross entropy loss L fractal_contrast is the fractal contrast learning loss, λ1, λ2, λ3 are balance factors, D is the discriminator, G is the generator, z is the hidden variable, c is the conditional vector, y true is the true label, y pred The generated label.

[0052] Specifically, the steps of continuous learning include:

[0053] Implementing knowledge distillation: L distill =KL(softmax(z old / T)||softmax(z new / T))

[0054]

[0055] Among them, z old and z new are the logits of the old model and the new model respectively, T is the temperature parameter, T0 is the initial temperature, τ r is the temperature annealing time constant;

[0056] Maintenance priority experience buffer pool: p i =|δ i |+ε,

[0057] Among them, δ iis the TD error of sample i, ε is a small positive number, and α is the priority index;

[0058] Implementing dynamic category extension:

[0059] If confidence(x) <threshold:

[0060] create_new_class()

[0061] otherwise:

[0062] update_existing_class()

[0063] confidence(x)=max(softmax(f(x)))

[0064] Among them, f(x) is the logits output by the classifier, and threshold is the confidence threshold for creating a new category.

[0065] Specifically, the model compression step includes:

[0066] Implementing multi-teacher knowledge distillation:

[0067] Implementing multi-teacher knowledge distillation:

[0068] L multi_distill =∑ i w i KL(softmax(z i / T i )||softmax(z s / T s ))

[0069] Among them, z i is the logits of the i-th teacher model, z s is the logits of the student model, w i is the teacher model weight, T i and T s are the temperature parameters of the teacher and student models, respectively;

[0070] Use a mixed precision quantization strategy: quantize the weights of the convolutional and fully connected layers to 8 bits, quantize the weights of the attention and normalization layers to 16 bits, and keep 32-bit floating points for key intermediate activations: Use a sensitivity-based iterative pruning algorithm: Calculate parameter sensitivity Sort by sensitivity and remove the bottom p% of parameters, then fine-tune the pruned model, repeating until the target compression ratio is achieved.

[0071] Specifically, the steps of deploying the mobile terminal include:

[0072] Using TensorFlowLit e Transformation model;

[0073] Apply the XNNPACK acceleration library to optimize inference performance;

[0074] Achieve 8-bit fixed-point conversion to reduce memory usage and computational complexity.

[0075] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0076] 1. High accuracy: Through the innovative fractal attention encoder and chaotic dynamics conditional generative decoder, the present invention significantly outperforms existing methods in classification accuracy and label generation quality, reaching 95.8% and 0.76 BLEU scores respectively. This means that products can be classified and described more accurately, which is conducive to improving the search and recommendation effects of the platform.

[0077] 2. Excellent adaptability to new categories: The continuous learning ability and dynamic category expansion mechanism of the present invention enable it to excel in handling emerging product categories, with a recognition rate of up to 89.2%. This enables the system to respond quickly to market changes without frequent manual intervention and model retraining.

[0078] 3. Efficient computing performance: Despite providing more complex functions, the method of the present invention still maintains a high computing efficiency, processing 1,000 SKUs in just 5.2 seconds. This ensures that the system can process a large number of SKUs in real time and meet the needs of large e-commerce platforms.

[0079] 4. Easy to deploy: Through advanced model compression technology, the present invention controls the model size to 42MB, which is much smaller than traditional deep learning methods, while maintaining excellent performance. This compact model size makes it easy to deploy on mobile devices, providing merchants with a convenient SKU management tool.

[0080] 5. Full fusion of multimodal information: The present invention realizes the deep fusion of image and text information through the innovative fractal attention mechanism, fully utilizes the complementary information between different modalities, and improves the richness and accuracy of feature representation.

[0081] 6. Generate diverse and high-quality labels: Through the quantum annealing adversarial training module, the present invention can generate more realistic, diverse and detailed product labels, effectively improving the discoverability of products and user experience.

[0082] In summary, the present invention not only solves many key problems existing in the prior art, but also achieves significant improvements in performance, efficiency, and practicality. This comprehensive improvement makes the present invention particularly suitable for the SKU management needs of modern large-scale e-commerce platforms, and is expected to significantly improve the platform's operational efficiency and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is the overall system logic block diagram of the present invention.

[0084] Figure 2 It is a logic block diagram of the data preprocessing step of the present invention.

[0085] Figure 3 This is a logic block diagram of the multimodal feature extraction of the present invention.

[0086] Figure 4 This is a logic block diagram of the zero-sample classification of the present invention.

[0087] Figure 5 Generate a logic block diagram for the tag of the present invention.

[0088] Figure 6 This is a logic block diagram of the continuous learning module of the present invention.

[0089] Figure 7 This is a logic block diagram of model compression and deployment of the present invention.

[0090] Figure 8 This is a schematic diagram of the framework of the improved fractal conditional variational autoencoder of the present invention.

[0091] Fig. 9 It is a multi-dimensional performance radar chart of the present invention and comparative documents 1-3.

[0092] Fig.10 This is a heat map of the new category adaptability of the present invention and comparative documents 1-3. DETAILED DESCRIPTION

[0093] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0094] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0095] Please refer to Figure 1-10The present invention provides a SKU intelligent classification and label generation method integrating a visual pre-training model, which is particularly suitable for the product management system of a large e-commerce platform. Taking a world-renowned e-commerce platform as an example, the platform needs to process the SKU information of millions of new products every day. Traditional manual classification and label generation methods can no longer meet such large-scale needs, so there is an urgent need for an efficient, accurate and scalable automated solution.

[0096] This method includes the following six main steps:

[0097] Step 1: Get the image data and text data of the SKU.

[0098] In this embodiment, when a merchant uploads a new product on the e-commerce platform, the system will automatically obtain the product's image and text description. Specifically:

[0099] 1.1 Image data processing,First, all product images are uniformly scaled to 224x224 resolution using the bicubic interpolation algorithm. This size is the best balance point obtained through a large number of experiments, which can retain sufficient image details without causing excessive computational burden.

[0100] Next, the system applies contrast-limited adaptive histogram equalization (CLAHE) to enhance image details. The parameters of the CLAHE algorithm are set as: tile_grid_size = (8,8), clip_limit = 2.0. These parameter values ​​are obtained through the analysis of a large number of e-commerce product images, and can effectively improve the detail performance of various types of product images.

[0101] To further enhance the generalization ability of the model, the system also randomly applies the following data augmentation techniques:

[0102] Random cropping: The cropping ratio range is [0.8, 1.0], which can simulate different shooting angles and distances while retaining the main product information.

[0103] Random horizontal flip: The probability is 0.5, simulating different placement directions of the product.

[0104] Random rotation: The angle range is [-10°, 10°], simulating slight changes in shooting angle.

[0105] Color Jitter: Randomly adjusts brightness, contrast, saturation, and hue within a range of ±0.2 to simulate different lighting conditions.

[0106] 1.2 Text data processing,For the text description of the product, BERT tokenizer is used for word segmentation, and the maximum sequence length is set to 128. This length is based on the statistical analysis of a large number of product descriptions and can cover the key information of most product descriptions.

[0107] The system also removes stop words and punctuation marks to reduce the interference of irrelevant information. At the same time, in order to enhance the model's ability to understand different expressions, the system applies the following text enhancement techniques:

[0108] Synonym replacement: Randomly select 30% of non-stop words and replace them with synonyms using WordNet. This ratio is the best value obtained through experiments, which can increase the diversity of text while maintaining the semantics of the original text.

[0109] Back-translation enhancement: Using Chinese as the intermediate language, we can perform English-Chinese-English back-translation through the Google Translate API. This method can produce descriptions with slightly different grammatical structures but similar semantics, effectively improving the robustness of the model.

[0110] 1.3 Data Formatting,The processed image data is converted into a PyTorch tensor with a shape of (3,224,224) and a data type of float32. This format is designed to match the input requirements of the subsequent CLIP model.

[0111] The processed text data is converted into a word index sequence and padded to a uniform length of 128. This is done to maintain data consistency during batch processing and improve computing efficiency.

[0112] Step 2: Extract multimodal features based on the image data and the text data

[0113] 2.1 Image feature extraction, use the pre-trained CLIPViT-B / 32 model as the visual encoder to extract the 512-dimensional image feature vector Vimg. The CLIP (Contrastive Language-Image Pre-training) model was developed by OpenAI and uses about 400 million pairs of image-text data for pre-training. These data come from publicly available resources on the Internet, including images of various topics and fields and their related descriptions. The pre-training process used hundreds of GPUs and took several weeks. This large-scale pre-training gives CLIP powerful cross-modal understanding capabilities, which is particularly suitable for our SKU classification and label generation tasks. ViT-B / 32 is a variant of CLIP, which is based on the VisionTransformer architecture and can capture global and local information of the image.

[0114] 2.2 Text feature extraction, also use CLIP's text encoder to extract the 512-dimensional text feature vector Vtxt. CLIP's text encoder is based on the Transformer architecture and can effectively understand complex language structures and semantics.

[0115] 2.3 Feature Fusion,In order to make full use of the complementary information of images and texts, the system adopts an adaptive weighted fusion strategy:

[0116] V fused =α·V img +(1-α)·V txt

[0117] α=σ(W α ·[V img ; V txt ]+b α )

[0118] Where σ is the sigmoid function, W α and b α is a learnable parameter.

[0119] Step 3: Implement zero-shot classification of SKUs based on the multimodal features

[0120] 3.1 Constructing category prototypes. For each product category c on the e-commerce platform, the system calculates the weighted average of all sample features of the category as the category prototype:

[0121]

[0122] Among them, w i is the sample weight, which is calculated by the sample quality assessment module

[0123] 3.2 Sample Quality Assessment Module,A lightweight neural network Q is designed to evaluate the sample quality.,The lightweight neural network Q is used to evaluate the quality or importance of each sample.

[0124]

[0125] The specific structure of the lightweight neural network Q is as follows:

[0126] Input layer: receives the fused feature vector V fused ;

[0127] Hidden layer: one or more fully connected layers, using ReLU activation function;

[0128] Output layer: one neuron that uses a sigmoid activation function to output a weight between 0 and 1;

[0129] For example, a simple implementation would be:

[0130] Q(V fused )=sigmoid(W2·ReLU(W1·V fused +b1)+b2).

[0131] 3.3 Similarity calculation: Use the improved cosine similarity to calculate the similarity between the features of the SKU to be classified and the prototypes of each category:

[0132]

[0133] Among them, ε is the smoothing factor (value is 1e-5), and β is the learnable temperature parameter.

[0134] 3.4 Category determination, finally, classify the SKU into the category with the highest similarity:

[0135]

[0136] Step 4: Based on the multimodal features, the SKU label is generated by the improved fractal conditional variational autoencoder. This step uses the improved fractal conditional variational autoencoder (IFCVAE) to achieve the generation of SKU labels. The IFCVAE model contains the following innovative modules:

[0137] 4.1 Fractal Attention Encoder 1

[0138] The calculation process is as follows:

[0139] h fractal =fractal attention (x, L = 3, S = 4)

[0140]

[0141] μlogσ=MLP(h fused )

[0142] Where L is the number of fractal layers, S is the number of subunits in each layer, and α i is the attention weight of adaptive learning, and MLP refers to Multi-Layer Perceptron. It is a commonly used neural network structure for nonlinear transformation. In order to further enhance the representation ability of the encoder, we introduce fractal contrast learning loss:

[0143]

[0144] 4.2 Chaotic dynamics condition generation decoder 2, the calculation process is as follows:

[0145] Z tilde =Reparameterize(μ,σ)

[0146] Z chaos =logistic map (z tilde , r = 3.9)

[0147] h coupled =concat[z chaos , c]

[0148] h lorenz =odeint(lorenz rnn,h coupled , t, args = (10, 28, 8 / 3))

[0149] y pred =MLP(h lorenz [-1])

[0150] Among them, z tilde is the latent variable obtained by reparameterization sampling, c is the conditional vector, r is the parameter σ of the Logistic mapping, ρ, β are the parameters of the Lorenz system, t is the time step sequence, y pred is the generated label probability distribution.

[0151] 4.3 The loss functions of the generator and discriminator of quantum annealing adversarial training module 3 are as follows:

[0152] L G =BCE(D(G(z,c)),1)+λ1KL(q(z|x)||p(z))+λ2CE(y pred ,y true )+λ3L fractal_contrast

[0153] L D =BCE(D(y true ), 1) + BCE(D(G(z,c)), 0)

[0154] Among them, BCE is the binary cross entropy loss, KL is the KL divergence, and CE is the classification cross entropy loss L fractal_contrast is the fractal contrast learning loss, λ1, λ2, λ3 are balance factors, D is the discriminator, G is the generator, z is the hidden variable, c is the conditional vector, y true is the true label, y pred The generated label.

[0155] The implementation of the quantum annealing strategy is as follows:

[0156]

[0157] Where noise0 is the initial noise intensity (usually set between 0.1 and 0.5), t is the current training step number, and τ is the annealing time constant (usually set to 1 / 2 of the total training steps). arrive ).

[0158] This annealing strategy allows the model to have a larger exploration space in the early stage of training, and gradually converges to the classical state as the training progresses. In practical applications, this quantum annealing adversarial training mechanism can help generate more diverse and realistic SKU labels. For example, for a newly launched smartwatch, through quantum annealing adversarial training, the system can generate more detailed and diverse labels.

[0159] 5.1 Knowledge distillation, realizing the knowledge distillation strategy of dynamic temperature:

[0160]

[0161] Among them, z old and z new are the logits of the old model and the new model respectively, T is the temperature parameter, T0 is the initial temperature (usually set between 2 and 5), τ T is the temperature annealing time constant (usually set to the total number of training steps arrive ).

[0162] 5.2 Experience replay, maintain a priority experience buffer pool, use TD error as priority:

[0163]

[0164] Among them, δ i is the TD error of sample i, ε is a small positive number (usually set to 0.01), and α is the priority exponent (usually set to 0.6).

[0165] 5.3 Dynamic category expansion, using the AdaBest algorithm to dynamically adjust the classifier structure:

[0166] Ifconfidence(x) <threshold:

[0167] create_new_class()

[0168] Else:

[0169] update_existing_class()

[0170] confidence(x)=max(softmax(f(x)))

[0171] Where f(x) is the logits output by the classifier, and threshold is the confidence threshold for creating new categories (usually set between 0.3 and 0.5).

[0172] Step 6: Compress the model and deploy it on mobile devices. In order to enable the model to run efficiently on resource-constrained mobile devices, this step implements a series of model compression and optimization techniques:

[0173] 6.1 Knowledge distillation, using multi-teacher distillation strategy:

[0174]

[0175] Among them, z i is the logits of the i-th teacher model, z s is the logits of the student model, w i is the teacher model weight (determined by validation set performance), T i and T s are the temperature parameters of the teacher and student models respectively (usually set between 1 and 5).

[0176] 6.2 Model quantization, using mixed precision quantization strategy:

[0177] 8-bit quantization: used for weights of convolutional and fully connected layers;

[0178] 16-bit quantization: used for weights of attention and normalization layers;

[0179] 32-bit floating point: used for key intermediate activations;

[0180] 6.3 Model pruning, using sensitivity-based iterative pruning algorithm

[0181] (1) Calculate the sensitivity of each parameter:

[0182] (2) Sort by sensitivity and remove the bottom p% of parameters (p usually starts at 5% and gradually increases to around 30%)

[0183] (3) Fine-tune the pruned model

[0184] (4) Repeat steps 1-3 until the target compression ratio is reached.

[0185] 6.4 Mobile Optimization

[0186] Convert models using TensorFlowLite, a framework optimized for mobile and embedded devices;

[0187] Apply the XNNPACK acceleration library to optimize inference performance. XNNPACK is a highly optimized neural network inference library, especially suitable for mobile devices;

[0188] Achieve 8-bit fixed-point, reduce memory usage and computation Through these optimization strategies, the originally complex IFCVAE model can be compressed to a size of tens of MB, and can achieve millisecond-level inference speed on ordinary smartphones. This allows merchants to complete SKU classification and label generation directly on mobile devices, greatly improving the convenience and efficiency of operations.

[0189] Through these optimization strategies, the originally complex IFCVAE model can be compressed to a size of tens of MB and can achieve millisecond-level inference speed on ordinary smartphones. This allows merchants to complete SKU classification and label generation directly on mobile devices, greatly improving the convenience and efficiency of operations.

[0190] The SKU intelligent classification and label generation method proposed in this paper, which integrates the visual pre-training model, effectively solves the SKU management challenges faced by large-scale e-commerce platforms through innovative algorithm design and engineering implementation. This method can not only accurately and efficiently classify SKUs, but also generate rich and diverse labels, and has the ability of continuous learning and mobile terminal deployment.

[0191] In practical applications, this method can significantly improve the efficiency and quality of product listings. For example, for a newly listed smartwatch, the system can automatically classify it into the category of smart wearable devices and generate labels such as heart rate monitoring, GPS positioning, waterproof, and long battery life. This accurate and rich information not only helps buyers quickly find the products they need, but also improves the platform's search and recommendation effects.

[0192] In addition, the method’s continuous learning capability enables it to adapt to changing market trends. When new product categories or attributes emerge, the system can automatically adjust and expand without the need for frequent manual intervention and model retraining.

[0193] Finally, through model compression and mobile optimization, this method can be directly deployed on merchants’ mobile devices, making SKU management more convenient and efficient. Merchants can upload new products anytime and anywhere, and the system will classify and generate labels in real time, greatly shortening the time it takes to put products on the shelves.

[0194] In general, the present invention provides a comprehensive, efficient and forward-looking solution for SKU management of e-commerce platforms, and is expected to play an important role in improving user experience and optimizing platform operation efficiency.

[0195] In order to further verify the effectiveness of the present invention, the applicant provides the embodiments of the present invention and comparative examples of different conditions, and conducts detailed comparative analysis.

[0196] Embodiment 1 (method of the present invention): A SKU intelligent classification and label generation method integrating a visual pre-training model is adopted, including a fractal attention encoder, a chaotic dynamics conditional generation decoder and a quantum annealing adversarial training module.

[0197] Comparative Example 1: Traditional rule-based SKU classification and label generation method, using predefined rules and keyword matching.

[0198] Comparative Example 2: SKU classification and label generation method based on traditional machine learning, using SVM for classification and LDA for label generation.

[0199] Comparative Example 3: SKU classification and label generation method based on deep learning, using ResNet50 for image feature extraction, BERT for text feature extraction, and fully connected layers for classification and label generation.

[0200] Test indicators and detection methods:

[0201] 1. Classification accuracy: Use the test set to calculate the accuracy of the model classification results.

[0202] Detection method: (number of correctly classified SKUs / total number of SKUs)*100%.

[0203] 2. Label generation quality: Use BLEU score to evaluate the quality of generated labels.

[0204] Detection method: Calculate the BLEU score between the generated labels and the manually annotated labels.

[0205] 3. New category adaptability: Evaluate the model’s ability to identify and handle newly emerging SKU categories.

[0206] Detection method: After introducing a new category, calculate the rate at which the model correctly identifies the new category.

[0207] 4. Computational efficiency: Evaluate the inference time of the model on standard hardware.

[0208] Testing method: Record the average time (in seconds) required to process 1,000 SKUs.

[0209] 5. Model size: Evaluate the size of the compressed model to accommodate mobile deployment.

[0210] Detection method: Record the size of the model file (MB).

[0211] The test results are shown in the following table:

[0212] index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Classification accuracy (%) 95.8 78.3 86.5 92.1 Label Generation Quality (BLEU) 0.76 0.45 0.58 0.69 New category adaptability (%) 89.2 25.7 41.3 62.8 Calculation efficiency (seconds) 5.2 1.8 3.5 7.9 Model size (MB) 42 0.5 85 230

[0213] |

[0214] Embodiment 1 (method of the present invention) is the best embodiment, and the test result analysis and explanation are as follows:

[0215] 1. Classification accuracy: The method of the present invention achieved a classification accuracy of 95.8%, which is significantly higher than other methods. This is mainly due to the multi-scale feature extraction capability of the fractal attention encoder and the modeling of complex nonlinear relationships by the chaotic dynamics conditional generation decoder. High accuracy means that products can be correctly classified, which is conducive to improving the search and recommendation effects of the platform.

[0216] 2. Label generation quality: The BLEU score of the proposed method reaches 0.76, which is significantly better than other methods. This shows that the generated labels are closer to the quality of manual annotation. The high-quality label generation is due to the quantum annealing adversarial training module, which can generate more realistic and diverse labels. This is crucial to improving the discoverability of products and user experience.

[0217] 3. New category adaptability: The proposed method excels in new category adaptability, achieving an 89.2% recognition rate. This is due to the model's continuous learning ability and dynamic category expansion mechanism. High new category adaptability means that the system can quickly respond to market changes and automatically handle emerging product categories without frequent manual intervention.

[0218] 4. Computational efficiency: The computational efficiency of the method of the present invention (5.2 seconds / 1000 SKU) is not as good as that of the rule-based method, but it is a good trade-off considering its significantly improved performance. Compared with the deep learning method of Comparative Example 3, the method of the present invention shows better efficiency, which is due to model compression and optimization technology. The high efficiency ensures that the system can process a large number of SKUs in real time to meet the needs of large e-commerce platforms.

[0219] 5. Model size: The compressed model size of the method of the present invention is 42MB, which is much smaller than the traditional deep learning method (Comparative Example 3), while maintaining excellent performance. This compact model size makes it easy to deploy on mobile devices, providing merchants with a convenient SKU management tool.

[0220] The method of the present invention has shown significant superiority in key indicators. It not only leads by a large margin in classification accuracy and label generation quality, but also shows excellent adaptability to new categories. Although it is slightly inferior to simple rule-based methods in computational efficiency, this is a completely acceptable trade-off considering the advanced features and performance improvements it provides. In addition, the relatively small model size makes it very suitable for mobile deployment, which is a unique advantage among competitors.

[0221] These results fully demonstrate the excellent ability of the present invention in handling large-scale, dynamically changing SKU management tasks. It not only provides high-quality classification and label generation services, but also adapts to the rapidly changing e-commerce environment while maintaining deployment flexibility. This comprehensive performance makes the present invention particularly suitable for the needs of modern e-commerce platforms, and has the potential to significantly improve the platform's operational efficiency and user experience.

[0222] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The SKU intelligent classification and label generation method integrating the visual pre-training model is characterized by: The following steps are involved: Obtain image data and text data of the SKU; extract multimodal features based on the image data and the text data; According to the multimodal features, zero-shot classification of SKU is realized; based on the multimodal features, SKU labels are generated by an improved fractal conditional variational autoencoder; Based on the newly added SKU data, the model can be continuously learned; The model is compressed and deployed on mobile terminals.

2. The method according to claim 1, characterized in that: The step of extracting multimodal features comprises: Using the pre-trained CLIP visual encoder, extract the image feature vector V from the image data img′ ; Using the pre-trained CLIP text encoder, extract the text feature vector V from the text data txt ; Through the adaptive weighted fusion strategy, the fusion feature vector V is obtained fused ,in: V fused =α·V img +(1-α)·V txt α=σ(W α ·[V img ;V txt ]+b α ) Among them, σ is the sigmoid function, W α and b α are learnable parameters.

3. The method according to claim 1, characterized in that The steps of zero-sample classification include: constructing a category prototype Among them, w i is the sample weight, which is calculated by the sample quality assessment module Q: Calculate the similarity between the features of the SKU to be classified and the prototypes of each category: Among them, ε is the smoothing factor, β is the learnable temperature parameter, and the SKU is classified into the category with the highest similarity: c * = argmax c sin(V fused ,P c ).

4. The method according to claim 1, characterized in that: The improved fractal conditional variational autoencoder includes a fractal attention encoder, a chaotic dynamics conditional generation decoder and a quantum annealing adversarial training module.

5. The method according to claim 4, characterized in that The calculation process of the fractal attention encoder is as follows: h=x For l in the range L: h sub =[h] For s in range S: h s =sub u nit(h,s) h sub .append(h s ) Return h h fractal =fractal attention (x,L=3,S=4) μlogσ=MLP(h fused ) Among them, L is the number of fractal layers, S is the number of subunits in each layer, α i is the attention weight for adaptive learning, x is the input feature, h is the intermediate feature, and W q ,W k ,W v is the learnable weight matrix, d k is the scaling factor of the attention mechanism, μ and logσ are the posterior distribution parameters of the latent variable z, and h sub It takes the natural logarithm of each element of the vector hsub.

6. The method according to claim 4, characterized in that The calculation process of the chaotic dynamics condition generation decoder is as follows: z tilde =Reparameterize(μ,σ) z chaos =logistic map (z tilde ,r=3.9) h coupled =concat[z chaos ,c] h lorenz =odeint(lorenzrnn,h coupl ed,t,args=(10,28,8 / 3)) y pred =MLP(h lorenz [-1]) Among them, z tilde is the latent variable obtained by reparameterization sampling, c is the conditional vector, r is the parameter of the Logistic mapping, σ, ρ, β are the parameters of the Lorenz system, t is the time step sequence, y pred is the generated label probability distribution.

7. The method according to claim 4, characterized in that The loss function of the quantum annealing adversarial training module is as follows: L G =BCE(D(G(z,c)),1)+λ1KL(q(z|x)||p(z))+λ2CE(y pred ,y true )+λ3L fractal_contrast L D =BCE(D(y true ),1)+BCE(D(G(z,c)),0) Among them, BCE is the binary cross entropy loss, KL is the KL divergence, and CE is the classification cross entropy loss L fractal_contrast is the fractal contrast learning loss, λ1,λ2,λ3 are balance factors, D is the discriminator, G is the generator, z is the hidden variable, c is the conditional vector, y true is the true label, y pred The generated label.

8. The method according to claim 1, characterized in that: The steps of continuous learning include: Implementing knowledge distillation: L distill =KL(softmax(z old / T)||softmax(z new / T)) Among them, z old and z new are the logits of the old model and the new model respectively, T is the temperature parameter, T0 is the initial temperature, τ T is the temperature annealing time constant; Maintaining priority experience buffer pool: Among them, δ i is the TD error of sample i, ε is a small positive number, and α is the priority index; Implementing dynamic category extension: If confidence(x) <threshold: create_new_class() otherwise: update_existing_class() confidence(x)=max(softmax(f(x))) Among them, f(x) is the logits output by the classifier, and threshold is the confidence threshold for creating a new category.

9. The method according to claim 1, characterized in that: The steps of model compression include: Implementing multi-teacher knowledge distillation: L multi_distill ∑ i w i ·KL(softmax(z i / T i )||softmax(z s / T s )) Among them, z i is the logits of the i-th teacher model, z s is the logits of the student model, w i is the teacher model weight, T i and T s are the temperature parameters of the teacher and student models, respectively; Use a mixed precision quantization strategy: quantize the weights of the convolutional and fully connected layers to 8 bits, quantize the weights of the attention and normalization layers to 16 bits, and keep 32-bit floating points for key intermediate activations: Use a sensitivity-based iterative pruning algorithm: Calculate parameter sensitivity Sort by sensitivity and remove the bottom p% of parameters, then fine-tune the pruned model, repeating until the target compression ratio is achieved.

10. The method according to claim 1, characterized in that The steps of mobile terminal deployment include: Convert the model using TensorFlowLite; Apply the XNNPACK acceleration library to optimize inference performance; Achieve 8-bit fixed-point conversion to reduce memory usage and computational complexity.

Images