A cross-border e-commerce product classification method based on cloud computing and deep learning

Abstract

The present invention proposes a cross-border e-commerce commodity classification method based on cloud computing and deep learning, including: collecting multimodal commodity data from a cross-border e-commerce platform and preprocessing the multimodal commodity data; performing multimodal feature extraction and feature fusion on the preprocessed multimodal commodity data; designing a meta-learning framework; using a graph neural network to construct a commodity relationship graph for commodities and their relationships on the cross-border e-commerce platform; constructing a classification model based on deep learning, using the final commodity feature representation as the input of the classification model, and deploying the classification model to a cloud computing platform; using the cloud computing platform to perform distributed training on the classification model; and deploying the trained classification model to a cross-border e-commerce platform. The present invention can not only efficiently process multimodal data, solve the cold start problem of new commodities, have dynamic adaptability, and make full use of the powerful resources of cloud computing, but also improve the commodity classification effect of the cross-border e-commerce platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud computing and deep learning, and in particular to a cross-border e-commerce commodity classification method based on cloud computing and deep learning. Background Art

[0002] Cross-border e-commerce platforms have extremely high requirements for accurate classification of goods due to their multilingual, multicultural and multi-category characteristics. However, current commodity classification methods are obviously insufficient in dealing with this complex scenario. The most critical problem is that existing methods are difficult to achieve effective fusion and processing of multimodal data.

[0003] Specifically, multimodal data includes product images, text descriptions, user comments, and voice information. These data are rich in sources and in various formats, but most existing classification technologies only process single-modal data. The limitation of single-modal processing methods is that they cannot fully utilize the rich information in different modal data, resulting in low classification accuracy and reliability. For example, text-based classification methods rely on the completeness and accuracy of product descriptions, but they often perform poorly when faced with multilingual or poorly translated descriptions. Similarly, although image-based classification methods can capture the visual features of products, the classification accuracy will also be affected when there is a lack of contextual information.

[0004] In addition, new products emerge in an endless stream on cross-border e-commerce platforms, but new products usually lack sufficient labeled data. Traditional classification models require a large amount of labeled data for training, and exhibit obvious cold start problems when facing new products, making it difficult to provide accurate classification results. This further exacerbates the difficulty of multimodal data processing.

[0005] The rapid changes in market and user needs also pose challenges to classification models. Existing methods usually lack the ability to dynamically adjust and cannot perform adaptive optimization based on real-time data and market changes, resulting in unstable classification results and difficulty in maintaining sustained and efficient classification performance. Large-scale commodity classification tasks also require powerful computing resources and storage capabilities. Traditional single machines or small-scale clusters are difficult to meet this demand, limiting the training efficiency and processing capabilities of classification models.

[0006] Therefore, in order to improve the product classification effect of cross-border e-commerce platforms, a product classification method is needed that can effectively integrate multimodal data, solve the cold start problem of new products, have dynamic adaptability and rely on powerful computing resources to comprehensively improve the classification performance. Summary of the invention

[0007] The purpose of the present invention is to design a cross-border e-commerce commodity classification method based on cloud computing and deep learning. Through innovative multimodal data fusion, adaptive learning, graph neural network modeling and cloud computing platform support, the commodity classification effect and user experience of the cross-border e-commerce platform are comprehensively improved. By solving the problems of multimodal data processing, new commodity cold start, dynamic adaptability and computing resource limitations, the present invention provides an efficient, accurate and adaptable solution for cross-border e-commerce commodity classification.

[0008] In order to achieve the above object, the present invention provides a cross-border e-commerce commodity classification method based on cloud computing and deep learning, and the method comprises the following steps:

[0009] S1. Collect multimodal commodity data from cross-border e-commerce platforms and pre-process the multimodal commodity data, and perform alignment and standardization on the pre-processed multimodal commodity data;

[0010] S2, performing multimodal feature extraction and feature fusion on the preprocessed multimodal commodity data to obtain a multimodal fusion feature vector;

[0011] S3. Use model-independent meta-learning to design a meta-learning framework, perform multi-task training and few-shot learning on the meta-learning framework, introduce feature enhancement and self-supervised learning in the few-shot learning of the meta-learning framework, and use the parameters of the few-shot learning of the meta-learning framework to classify and predict the fused feature vector;

[0012] S4. Use graph neural network to construct a product relationship graph for the products and their relationships on the cross-border e-commerce platform, and the product relationship graph outputs the final product feature representation;

[0013] S5. Build a classification model based on deep learning, use the final product feature representation as the input of the classification model, deploy the classification model to the cloud computing platform, use distributed computing resources for parallel training, and then evaluate and verify the classification model;

[0014] S6, dynamically allocate and manage cloud computing resources, and use the cloud computing platform to perform distributed training on the classification model;

[0015] S7. Deploy the trained classification model to the cross-border e-commerce platform;

[0016] Wherein, the step S3 specifically includes:

[0017] S31. Design a meta-learning framework:

[0018] The multimodal fusion feature vector F obtained by S2 finalInput into the meta-learning model. In the learning phase, the meta-learning model learns the initial parameter θ. Then, the meta-learning model is trained on multiple tasks to optimize the initial parameter θ, which is expressed as follows:

[0019]

[0020] Among them, θ is the initial parameter, T i is the task, p(T) is the task distribution, is the loss function, α is the learning rate, θ * represents the learned model parameters, It means to find the gradient of the initial parameter θ. Represents the loss function The initial parameter θ to be minimized, Represents task T i The loss function on ;

[0021] S32. Define different commodity classification tasks to form a task set T. In the multi-task training phase of the meta-learning model, update the initial parameter θ:

[0022] Randomly select task T from task set T i , for each task T i , the meta-learning model uses a small number of samples to adapt, calculate the gradient and update the parameters. After each task training, the updated parameters are back-propagated to optimize the global parameter θ all , which is expressed as follows:

[0023]

[0024] Among them, β represents the meta-learning rate;

[0025] S33. After the meta-learning phase is completed, few-sample learning optimization is performed for the specific new product classification task:

[0026] A small number of training samples of new products are input into the meta-learning model, and the learned model parameters θ are used * The training samples and meta-learning model are adapted, and the model is fine-tuned using a small number of samples to optimize the parameters of the new product classification task, as shown below:

[0027]

[0028] Among them, θ new represents the optimization parameters of the new task, represents the loss function of the new task, Represents the learned model parameters θ * Find the gradient;

[0029] S34. Based on the optimization of few-sample learning for the specific new product classification task, feature enhancement and self-supervised learning techniques are introduced to optimize the training of few-sample learning:

[0030] Using self-supervised learning methods, we generate the feature vector F by pre-training on unlabeled data. ssl , the feature vector F generated by self-supervised learning ssl And the feature vector F generated by few-shot learning few Fusion is performed to obtain the enhanced feature vector F enhanced , which is expressed as follows:

[0031] F ssl =SelfSupervisedLearning(X)

[0032] F enhanced =F few +λF ssl

[0033] Among them, X represents unlabeled data, λ represents fusion weight, and SelfSupervisedLearning represents self-supervised learning;

[0034] S35, use the optimized parameter θ new The fusion feature vector F enhanced To make a categorical prediction:

[0035] The fused feature vector F enhanced Input into the meta-learning model optimized by few-shot learning, and use the optimized meta-learning model to optimize the parameters θ of the new task new Perform classification prediction and output the final product category, as shown below

[0036]

[0037] in, Indicates the predicted product category, and Predict indicates the classification prediction function.

[0038] Furthermore, the multimodal product data includes images, text descriptions, user comments and voice information.

[0039] Furthermore, the preprocessing includes image enhancement, text data preprocessing, and speech data preprocessing; the alignment operation includes image-text alignment and text-speech alignment; and the standardization processing includes image standardization, text standardization, and speech standardization.

[0040] Furthermore, in S2, an improved visual Transformer model is used to extract image features; an improved BERT model is used to extract text features; and an improved speech Transformer model is used to extract speech features; wherein the improved visual Transformer model is represented as follows:

[0041] F i =WeightedAverage(TransformerEncoder(P(I′+PosEmb)))

[0042] Where I′ represents the preprocessed image, P represents the image block flattening operation, PosEmb represents the position embedding, TransformerEncoder represents the Transformer encoder, and F i Represents the final image feature vector;

[0043] The improved BERT model is used to extract text features, which is expressed as follows:

[0044] F t =WeightedAverage(BERTEncoder(T″))

[0045] Where T″ represents the preprocessed text, BERTEncoder represents the BERT encoder, and F t Represents the final text feature vector;

[0046] The improved speech Transformer model is expressed as follows:

[0047] F s =WeightedAverage(TransformerEncoder(Q(FFT(S′)+PosEmb)))

[0048] Where S′ represents the preprocessed speech signal, FFT represents Fourier transform, Q represents framing operation, PosEmb represents position embedding, TransformerEncoder represents Transformer encoder, and F s represents the final speech feature vector, and WeightedAverage represents the weighted average function;

[0049] The final image feature vector F is extracted i , the final text feature vector F t And the final speech feature vector F s Use the fully connected layer to reduce the dimension of the concatenated feature vector to obtain the final multimodal feature vector F multi, which is expressed as follows:

[0050] F multi =FC([F i ; F t ; F s ])

[0051] Among them, [F i ; F t ; F s ] represents the concatenation operation of feature vectors, and F init =[F i ; F t ; F s ], F init represents the initial fusion feature vector, FC represents the fully connected layer, and F multi represents the final multimodal feature vector,

[0052] Furthermore, S2 also includes: using a self-attention mechanism to initially fuse the feature vector F init Perform weighted processing to generate a weighted fusion feature vector F attn , use the deep fusion network to weight the fusion feature vector F attn Processing is performed to generate the final fusion feature vector F final .

[0053] Furthermore, the S4 comprises:

[0054] S41. Construct commodity relationship diagram:

[0055] Define the nodes of the graph as each product v i , define edge e based on the relationship between commodities ij ; Among them, the node characteristics The multimodal feature vector F of the product final Indicates that the edge feature e ij It is represented by the commodity relationship characteristics, as follows:

[0056] G=(V,E,H (0) ,E)

[0057] Among them, V represents the node set, E represents the edge set, and H (0) represents the initial feature matrix of the node, and E represents the edge feature matrix;

[0058] S42. Design multiple graph convolutional layers to update node features layer by layer and capture the relationship information between products:

[0059] Each graph convolutional layer receives the node features H of the previous layer (l)The weighted sum of the neighbor node features of each node is calculated, and the aggregated features are processed using a nonlinear activation function to obtain the updated node feature H. (l+1) , which is expressed as follows:

[0060]

[0061] in, represents the feature vector of node i in the l+1 layer, N(i) represents the set of neighbor nodes of node i, c ij is the normalization coefficient, W (l) represents the weight matrix of the lth layer, b (l) represents the bias term, σ represents the nonlinear activation function;

[0062] S43. Introduce the attention mechanism in the graph convolution layer to weight the features of neighbor nodes:

[0063] For each node, the attention weight is calculated for its neighboring nodes. At the same time, the neighboring node features are weighted and aggregated according to the attention weight, and the node features are updated, which is expressed as follows:

[0064]

[0065]

[0066] in, represents the attention weight between node i and node j, represents the normalized attention weight, a represents the attention parameter vector, || represents the vector connection operation, LeakyReLU represents the LeakyReLU activation function, represents the attention weight between node i and node k, represents the feature vector of node i in layer l, Represents the feature vector of node j in layer l;

[0067] S44. Generate a global graph embedding vector by globally aggregating all node features to capture the global information of the entire product relationship graph:

[0068] Use the global aggregation function to aggregate all node features and generate a global graph embedding vector h G , which is expressed as follows:

[0069]

[0070] Among them, h G represents the global graph embedding vector, GlobalPooling represents the global aggregation function, Represents the feature vector of node i in the Lth layer;

[0071] S45. Perform graph embedding and multimodal feature fusion:

[0072] Embed the global graph into vector h G and the multimodal fusion feature vector F multi Concatenate and use the fully connected layer to process the concatenated feature vectors to generate the final product feature representation F final , which is expressed as follows:

[0073] F final =FC([h G ; F multi ])

[0074] Among them, [h G ; F multi ] represents the concatenation operation of feature vectors, FC represents the fully connected layer, and F final Represents the final product feature representation.

[0075] Furthermore, the loss function of the deep learning-based classification model is expressed as follows:

[0076]

[0077] in, represents the cross entropy loss, N represents the number of node samples, y true,i and i Represent the true label and predicted probability of the i-th node sample respectively;

[0078] In the loss function The L2 regularization term is added to control the complexity of the model parameters, which is expressed as follows:

[0079]

[0080] in, is the loss function after regularization, λ is the regularization coefficient, is the kth item in the model parameters;

[0081] The classification model is deployed to the cloud computing platform and parallel training is performed using distributed computing resources, as shown below:

[0082] W f ←CloudTrain(X,y,W f ,b f )

[0083] Among them, CloudTrain represents the operation of training on the cloud computing platform, W f and b fThey represent the weight and bias parameters of the updated classification model respectively, X represents the input feature matrix of the classification model, and y represents the output product category probability distribution.

[0084] Furthermore, the distributed training of the classification model using the cloud computing platform is expressed as follows:

[0085]

[0086] Among them, θ t represents the model parameters at time t, η represents the learning rate, N represents the number of node samples, Represents the loss function of the i-th node sample.

[0087] Furthermore, in the S7, it also includes encapsulating the classification model using containerization technology and performing A / B testing on the classification model.

[0088] Furthermore, the classification model uses the Adam optimizer to calculate the gradient and update the classification model parameters to minimize the cross entropy loss, which is expressed as follows:

[0089]

[0090] Among them, W f and b f They represent the classification model parameters, η represents the learning rate, and denote the first-order and second-order moment estimates respectively, and ∈ denotes a stable term.

[0091] The beneficial technical effects of the present invention are at least as follows:

[0092] (1) The present invention adopts advanced multimodal data fusion technology, using the visual Transformer (ViT) to extract product image features, the BERT model to process product description and user comment text data, and the speech recognition model to process speech data. The self-attention mechanism is used to fuse the features of data of different modalities to form a unified product feature vector, effectively improving the accuracy of classification. This solves the problem that existing methods are difficult to fully utilize the rich information of multimodal data.

[0093] (2) The present invention introduces a meta-learning framework, which enables the model to quickly adapt to new product categories and market changes. At the same time, by using the few-sample learning technology, the model can still maintain efficient classification performance even when data is scarce, solving the cold start problem of new products. This ensures that the classification model can still provide accurate classification results when new products continue to emerge on the cross-border e-commerce platform.

[0094] (3) The adaptive learning model of the present invention can be dynamically adjusted according to the characteristics of commodities and user behavior data in different markets. Through the adaptive learning algorithm, the classification standard can be self-optimized to improve the accuracy and adaptability of classification, thereby better meeting the dynamic needs of cross-border e-commerce platforms. This innovation solves the problem that the existing methods lack dynamic adjustment capabilities and cannot be optimized according to market changes.

[0095] (4) The present invention uses graph neural networks to model products and their relationships (such as brands, categories, user reviews, etc.) and capture the complex relationship information between products. By dynamically updating the product relationship graph and adjusting the graph structure in real time according to user behavior and market changes, the dynamic adaptability of the classification model is further improved. This enables the classification model to accurately classify in a complex product relationship network.

[0096] (5) The present invention relies on cloud computing platforms (such as AWS and Google Cloud) to conduct large-scale distributed training to improve the training efficiency and processing power of the model. The cloud platform also supports model sharing and collaborative learning. Different cross-border e-commerce platforms can share classification models to improve overall classification performance. Real-time data processing and model updates are achieved through the cloud computing platform to ensure the response speed and accuracy of the system. This innovation solves the limitations of traditional methods in computing resources and storage capacity.

[0097] (6) The present invention can not only efficiently process multimodal data and solve the cold start problem of new products, but also has dynamic adaptability and makes full use of the powerful resources of cloud computing, significantly improving the product classification effect and user experience of the cross-border e-commerce platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0099] Figure 1 This is a flow chart of a cross-border e-commerce commodity classification method based on cloud computing and deep learning in the present invention. DETAILED DESCRIPTION

[0100] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0101] In one or more embodiments, Figure 1As shown, a cross-border e-commerce commodity classification method based on cloud computing and deep learning is disclosed, including the following steps S1-S7:

[0102] S1. Collect multimodal commodity data from cross-border e-commerce platforms and pre-process the multimodal commodity data, and perform alignment and standardization on the pre-processed multimodal commodity data.

[0103] Specifically, by collecting and preprocessing product images, text descriptions, user comments and voice information on cross-border e-commerce platforms, the diversity and consistency of the data are ensured. This step provides high-quality basic data for subsequent steps. Through image enhancement, text cleaning and voice conversion, the noise and inconsistency in the data are eliminated, laying a solid foundation for subsequent feature extraction.

[0104] Furthermore, multimodal product data including images, text descriptions, user comments, and voice information are obtained from cross-border e-commerce platforms to ensure data diversity and high quality.

[0105] Image data collection: Use the API provided by the cross-border e-commerce platform to obtain product images. Ensure that each product contains at least one main image and several auxiliary images (such as detailed images, panoramic images, etc.).

[0106] Text data collection: Use the API interface of the cross-border e-commerce platform to collect product titles, descriptions, specifications and user reviews in multiple languages ​​to ensure the completeness and accuracy of multilingual texts.

[0107] Voice data collection: Use the voice comment interface of the cross-border e-commerce platform to extract voice information from user comments. Ensure the clarity and recognizability of voice data.

[0108] The pre-processing comprises:

[0109] Image data preprocessing: Enhance the product images to increase data diversity. Perform random cropping, rotation, flipping, scaling, color change and other operations. Convert the images to a size of 224x224 pixels for subsequent processing. Let the original image be I and the enhanced image be I′. The enhancement process can be expressed as:

[0110] I′=T(I,θ) (1)

[0111] Wherein, T represents the enhancement transformation, and θ is the transformation parameter (such as angle, scaling factor, etc.).

[0112] Text data preprocessing: noise removal, word segmentation, stop word removal, etc. Use regular expressions to clean text and remove irrelevant symbols and characters. Use BERT for subword segmentation and process multilingual text. Let the original text be T and the cleaned text be T′. The cleaning process can be expressed as:

[0113] T′=clean(T) (2)

[0114] Among them, clean() represents the cleaning function.

[0115] Multilingual processing: Convert texts in different languages ​​into a unified language (such as English) for processing, using a multilingual translation model:

[0116] T″=translate(T′) (3)

[0117] Among them, translate() represents the translation function.

[0118] Speech data preprocessing: Use a speech recognition model to convert speech data into text. Use an end-to-end speech recognition model (such as DeepSpeech) to convert speech waveforms into text. Let the original speech data be S and the converted text be T s The conversion process can be expressed as:

[0119] T s =ASR(S) (4)

[0120] Among them, ASR() represents the speech recognition function.

[0121] Furthermore, align the data of different modes to ensure that the data of each product can be effectively associated:

[0122] Image and text alignment: Ensure that each product image corresponds to the corresponding text description and specification parameters. Associate the image ID with the text ID to build a multimodal data alignment table. Let the image ID be ID i , the text ID is ID t , the alignment process can be expressed as:

[0123] Align(ID i ,ID t )(5)

[0124] Among them, Align() represents the alignment operation.

[0125] Text and voice alignment: Ensure that the text and voice data of each user comment correspond to each other. Build an alignment table of comment ID and voice ID. Set the text ID as ID t , voice ID is ID s , the alignment process can be expressed as:

[0126] Align(ID t ,ID s )(6)

[0127] Furthermore, the standardization includes:

[0128] Image normalization: Normalize the pixel values ​​of all images to the interval [0,1]. Let the image pixel value be P, and the normalized pixel value be P′:

[0129]

[0130] Among them, P min and P max Represents the minimum and maximum pixel values ​​respectively.

[0131] Text normalization: convert all text to lowercase and remove stop words. Use word vectors to represent text, such as GloVe or Word2Vec. Let the original text be T and the normalized text be T′:

[0132] T′=normalize(T) (8)

[0133] Among them, normaliz() represents the text normalization function.

[0134] Speech standardization: Normalize the speech signal to eliminate the influence of different recording conditions. Suppose the original speech signal is S and the standardized speech signal is S′:

[0135]

[0136] Among them, S mean and S std Represent the mean and standard deviation of the speech signal respectively.

[0137] S2. Perform multimodal feature extraction and feature fusion on the preprocessed multimodal commodity data to obtain a multimodal fusion feature vector.

[0138] Specifically, after preprocessing, the visual Transformer (ViT) is used to extract image features, the BERT model is used to extract text features, and the pre-trained speech recognition model is used to extract speech features. Through advanced feature extraction models, it is ensured that the representative features of each modality of data can be fully mined. The core of this step is to ensure that data of different modalities can be represented efficiently and accurately, providing rich information for subsequent feature fusion.

[0139] Based on the extracted features, the image, text and voice features are fused using the self-attention mechanism to form a unified product feature vector. This step uses the self-attention mechanism to achieve effective weighting and integration of different modal features, ensuring that the fused feature vector can integrate information from various types of data and improve the accuracy of the classification model.

[0140] Among them, in S2, an improved visual Transformer model is used to extract image features; an improved BERT model is used to extract text features; and an improved speech Transformer model is used to extract speech features.

[0141] Specifically, an improved visual Transformer (ViT) model is used to extract image features:

[0142] Input the preprocessed image I′ into the visual Transformer model;

[0143] The image is divided into fixed-size image blocks, and each image block is flattened into a one-dimensional vector;

[0144] Add a position embedding vector to each image block to preserve the spatial information of the image;

[0145] Use a multi-layer Transformer encoder to process the embedding vector;

[0146] Extract the output features of each layer of encoder and perform weighted average on the outputs of all layers to obtain the final image feature vector F i ; means as follows:

[0147] F i =WeightedAverage(TransformerEncoder(P(I′+PosEmb)))(10)

[0148] Where I′ represents the preprocessed image, P represents the image block flattening operation, PosEmb represents the position embedding, TransformerEncoder represents the Transformer encoder, and F i Represents the final image feature vector.

[0149] Specifically, the improved BERT model is used to extract text features:

[0150] Input the preprocessed text T″ into the BERT model;

[0151] Use a multi-layer Transformer encoder to process text;

[0152] Extract the output features of each layer of encoder and perform weighted average on the outputs of all layers to obtain the final text feature vector F t ,formula:

[0153] F t =WeightedAverage(BERTEncoder(T″))(11)

[0154] Where T″ represents the preprocessed text, BERTEncoder represents the BERT encoder, and F t Represents the final text feature vector.

[0155] Specifically, an improved speech Transformer model is used to extract speech features:

[0156] Input the preprocessed speech signal S′ into the speech Transformer model;

[0157] Divide the speech signal into frames and perform Fourier transform on each frame to obtain a spectrogram;

[0158] Add a position embedding vector to each frame of the spectrogram to preserve the time information;

[0159] Use a multi-layer Transformer encoder to process the embedding vector;

[0160] Extract the output features of each layer of encoder and perform weighted average on the outputs of all layers to obtain the final speech feature vector F s ,formula:

[0161] F s =WeightedAverage(TransformerEncoder(Q(FFT(S′)+PosEmb)))(12)

[0162] Where S′ represents the preprocessed speech signal, FFT represents Fourier transform, Q represents framing operation, PosEmb represents position embedding, TransformerEncoder represents Transformer encoder, and F s Represents the final speech feature vector, and WeightedAverage represents the weighted average function.

[0163] The final image feature vector F is extracted i , the final text feature vector F t And the final speech feature vector F s Use the fully connected layer to reduce the dimension of the concatenated feature vector to obtain the final multimodal feature vector F multi , which is expressed as follows:

[0164] F multi =FC([F i ; F t ; F s ])(13)

[0165] Among them, [F i ; F t ; F s] represents the concatenation operation of feature vectors, FC represents the fully connected layer, and F milti Represents the final multimodal feature vector.

[0166] Furthermore, it also includes: using a self-attention mechanism to initially fuse the feature vector F init Perform weighted processing to generate a weighted fusion feature vector F attn , use the deep fusion network to weight the fusion feature vector F attn Processing is performed to generate the final fusion feature vector F final .

[0167] Specifically, the image feature vector F i , text feature vector F t and speech feature vector F s Splice in a specific order and use the splicing operation to form a preliminary fusion feature vector F init . formula:

[0168] F init =[F i ; F t ; F s ] (14)

[0169] Among them, [F i ; F t ; F s ] represents the concatenation operation of feature vectors, F init Represents the preliminary fused feature vector.

[0170] The initial fusion feature vector F init Perform linear transformation to obtain the query matrix Q, key matrix K and value matrix V. Calculate the dot product of the query matrix Q and the key matrix K, and perform scaling and softmax normalization to obtain the attention weight matrix A. Use the attention weight matrix A and the value matrix V for weighted summation to obtain the weighted fusion feature vector F attn . formula:

[0171] Q=W Q F init , K=W K F init , V=W V F init (15)

[0172]

[0173] F attn =AV (17)

[0174] Among them, W Q ,W K ,W Vis the linear transformation matrix, A is the attention weight matrix, F attn is the weighted fusion feature vector, d k is the dimension of the key matrix.

[0175] The weighted fusion feature vector F attn Input into a multi-layer fully connected network (MLP). Add a nonlinear activation function (such as ReLU) after each layer of the fully connected network to increase the nonlinear representation ability of the feature. Add batch normalization after each layer of the fully connected network to accelerate training and stabilize the model. After multiple layers of processing, the final fused feature vector F is obtained. final . formula:

[0176] F final =MLP(F attn ) (18)

[0177] Among them, MLP represents a multi-layer fully connected network, F final Represents the final fused feature vector.

[0178] S3. Use model-agnostic meta-learning (MAML) to design a meta-learning framework, perform multi-task training and few-shot learning on the meta-learning framework, introduce feature enhancement and self-supervised learning in the few-shot learning of the meta-learning framework, and use the parameters of the few-shot learning of the meta-learning framework to classify and predict the fused feature vector.

[0179] Specifically, the fused feature vectors are used to construct a meta-learning framework, which enables the classification model to quickly adapt to new product categories and market changes. At the same time, through the few-sample learning technology, efficient classification performance can still be maintained when data is scarce. This step solves the cold start problem of new products through the adaptive learning framework, enabling the model to quickly respond to the emergence of new products and changes in market demand.

[0180] The specific steps include:

[0181] S31. Design a meta-learning framework:

[0182] The multimodal fusion feature vector F obtained by S2 final Input into the meta-learning model. In the learning phase, the meta-learning model learns the initial parameter θ. Then, the meta-learning model is trained on multiple tasks to optimize the initial parameter θ, which is expressed as follows:

[0183]

[0184] Among them, θ is the initial parameter, T i is the task, p(T) is the task distribution, is the loss function, α is the learning rate, θ * represents the learned model parameters, It means to find the gradient of parameter θ. Represents the loss function The parameter θ to be minimized, Represents task T i The loss function on ;

[0185] S32. Define different commodity classification tasks to form a task set T. In the multi-task training phase of the meta-learning model, update the initial parameter θ:

[0186] Randomly select task T from task set T i , for each task T i , the meta-learning model uses a small number of samples to adapt, calculate the gradient and update the parameters. After each task training, the updated parameters are back-propagated to optimize the global parameter θ all , which is expressed as follows:

[0187]

[0188] Among them, β represents the meta-learning rate;

[0189] S33. After the meta-learning phase is completed, few-sample learning optimization is performed for the specific new product classification task:

[0190] A small number of training samples of new products are input into the meta-learning model, and the learned model parameters θ are used * The training samples and meta-learning model are adapted, and the model is fine-tuned using a small number of samples to optimize the parameters of the new product classification task, as shown below:

[0191]

[0192] Among them, θ new represents the optimization parameters of the new task, represents the loss function of the new task, Represents the learned model parameters θ * Find the gradient;

[0193] S34. Based on the optimization of few-sample learning for the specific new product classification task, feature enhancement and self-supervised learning techniques are introduced to optimize the training of few-sample learning:

[0194] Using self-supervised learning methods, we generate the feature vector F by pre-training on unlabeled data. ssl , the feature vector F generated by self-supervised learning ssl And the feature vector F generated by few-shot learningfew Fusion is performed to obtain the enhanced feature vector F enhanced , which is expressed as follows:

[0195] F ssl =SelfSupervisedLearning(X) (22)

[0196] F enhanced =F few +λF ssl (twenty three)

[0197] Among them, X represents unlabeled data, λ represents fusion weight, and SelfSupervisedLearning represents self-supervised learning;

[0198] S35, use the optimized parameter θ new The fusion feature vector F enhanced To make a categorical prediction:

[0199] The fused feature vector F enhanced Input into the meta-learning model optimized by few-shot learning, and use the optimized meta-learning model to optimize the parameters θ of the new task new Perform classification prediction and output the final product category, as shown below

[0200]

[0201] in, Indicates the predicted product category, and Predict indicates the classification prediction function.

[0202] S4. Use graph neural network to construct a product relationship graph for the products and their relationships on the cross-border e-commerce platform, and the product relationship graph outputs the final product feature representation.

[0203] Specifically, based on the features extracted and fused in the previous steps, the product relationship graph is constructed using information such as product brand, category, and user reviews, and the product relationship graph is trained through a graph neural network (GNN) to extract the relationship features between products. This step further improves the accuracy and adaptability of the classification model by capturing the complex relationship information between products, enabling the model to accurately classify in more complex product relationship networks.

[0204] The specific steps include:

[0205] S41. Construct commodity relationship diagram:

[0206] Define the nodes of the graph as each product v i , define edge e based on the relationship between commodities ij , such as brand similarity, category similarity, regional relevance, etc.; among them, node features The multimodal feature vector F of the product final Indicates that the edge feature e ij It is represented by the commodity relationship characteristics, as follows:

[0207] G=(V,E,H (0) ,E) (25)

[0208] Among them, V represents the node set, E represents the edge set, and H (0) represents the initial feature matrix of the node, and E represents the edge feature matrix;

[0209] S42. Design multiple graph convolutional layers to update node features layer by layer and capture the relationship information between products:

[0210] Each graph convolutional layer receives the node features H of the previous layer (l) The weighted sum of the neighbor node features of each node is calculated, and the aggregated features are processed using a nonlinear activation function to obtain the updated node feature H. (l+1) , which is expressed as follows:

[0211]

[0212] in, represents the feature vector of node i in the l+1 layer, N(i) represents the set of neighbor nodes of node i, c ij is the normalization coefficient, W (l) represents the weight matrix of the lth layer, b (l) represents the bias term, σ represents the nonlinear activation function;

[0213] S43. Introduce the attention mechanism in the graph convolution layer to weight the features of neighbor nodes:

[0214] For each node, the attention weight is calculated for its neighboring nodes. At the same time, the neighboring node features are weighted and aggregated according to the attention weight, and the node features are updated, which is expressed as follows:

[0215]

[0216] in, is the attention weight between node i and node j, is the normalized attention weight, a is the attention parameter vector, and || represents the vector connection operation.

[0217] S44. Generate a global graph embedding vector by globally aggregating all node features to capture the global information of the entire product relationship graph:

[0218] Use the global aggregation function to aggregate all node features and generate a global graph embedding vector h G , which is expressed as follows:

[0219]

[0220] Among them, h G represents the global graph embedding vector, GlobalPooling represents the global aggregation function, Represents the feature vector of node i in the Lth layer;

[0221] S45. Perform graph embedding and multimodal feature fusion:

[0222] Embed the global graph into vector h G and the multimodal fusion feature vector F multi Concatenate and use the fully connected layer to process the concatenated feature vectors to generate the final product feature representation F final , which is expressed as follows:

[0223] F final =FC([h G ; F multi ]) (31)

[0224] Among them, [h G ; F multi ] represents the concatenation operation of feature vectors, FC represents the fully connected layer, and F fimal Represents the final product feature representation.

[0225] S5. Build a classification model based on deep learning, use the final product feature representation as the input of the classification model, deploy the classification model to the cloud computing platform, use distributed computing resources for parallel training, and then evaluate and verify the classification model.

[0226] Specifically, the final product feature representation F generated in the previous steps is used final As the input of the classification model, the feature vector F that combines multimodal features and product relationship information is final As input data, prepare to train the classification model. Formula:

[0227] X=F final (32)

[0228] Where X represents the input feature matrix of the classification model.

[0229] Furthermore, a deep learning-based classification model (e.g., a Transformer-based classifier) ​​is constructed for the classification task of cross-border e-commerce products:

[0230] Design a deep learning model with a multi-layer structure, where each layer contains a fully connected layer, an activation function, and a regularization layer. The last layer uses a softmax activation function to output the probability distribution of each product category. Formula:

[0231] y=softmax(W f X+b f ) (33)

[0232] Among them, W f and b f are the weight and bias parameters of the classification model, and y is the output product category probability distribution.

[0233] Preferably, a cross entropy loss function is used to evaluate the prediction accuracy of the classification model. Calculate the predicted category probability distribution y and the true category label y true The cross entropy loss between .

[0234] formula:

[0235]

[0236] in, is the cross entropy loss, N is the number of node samples, y true,i and i are the true label and predicted probability of the i-th node sample respectively.

[0237] Preferably, an adaptive optimization algorithm (such as the Adam optimizer) is used to train and optimize the classification model:

[0238] Initialize model parameters W f and b f . Use the Adam optimizer to calculate the gradient and update the model parameters to minimize the cross entropy loss. Formula:

[0239]

[0240] Where η is the learning rate, and are the first-order and second-order moment estimates respectively, and ∈ is a stable term.

[0241] Furthermore, regularization techniques (such as Dropout and L2 regularization) are introduced to prevent overfitting and enhance the generalization ability of the model. Dropout layers are added between fully connected layers to randomly discard some neurons and reduce the risk of overfitting of the model. L2 regularization terms are added to the loss function to control the complexity of the model parameters. Formula:

[0242]

[0243] in, is the loss function after regularization, λ is the regularization coefficient, is the kth item in the model parameters.

[0244] Furthermore, we use cloud computing platforms (such as AWS and Google Cloud) to conduct large-scale distributed training to improve model training efficiency and processing capabilities. We deploy the classification model to the cloud computing platform, use distributed computing resources for parallel training, and use the automatic expansion and fault tolerance mechanisms provided by the cloud computing platform to ensure the stability and efficiency of model training. Formula:

[0245] W f ←CloudTrain(X,y,W f ,b f )(38)

[0246] Among them, CloudTrain represents the operation of training on the cloud computing platform.

[0247] Preferably, use a validation set to evaluate the classification model, tune hyperparameters, and ensure the model's generalization ability on different datasets:

[0248] The training data is divided into a training set and a validation set. After each training cycle, the validation set is used to evaluate the model performance and the accuracy and loss value on the validation set are recorded. Based on the validation results, the model hyperparameters (such as learning rate, regularization coefficient, etc.) are adjusted to improve the generalization performance of the model. Formula:

[0249]

[0250] Among them, N val is the number of samples in the validation set, 1 is the indicator function, argmax(y i ) is the model prediction category, y true,i is the real category.

[0251] S6. Dynamically allocate and manage cloud computing resources, and use the cloud computing platform to perform distributed training on the classification model.

[0252] Specifically, the cloud computing platform is used to improve the computing power and efficiency of the cross-border e-commerce product classification system, ensuring the efficiency and stability of model training, deployment and reasoning.

[0253] Among them, cloud computing resource allocation includes dynamic allocation and management of cloud computing resources to ensure efficient execution of model training and reasoning tasks.

[0254] Dynamically allocate computing resources (such as CPU, GPU, memory) according to the needs of model training tasks. Use resource scheduling algorithms to optimize resource utilization, reduce resource waste and task waiting time. Formula:

[0255] R t =Allocate(C t ,M t ,D t ) (40)

[0256] Among them, R t represents the resources allocated at time t, C t represents the computing resource requirement, M t Indicates the memory requirement, D t Indicates data storage requirements.

[0257] Furthermore, we can use the distributed computing capabilities of the cloud computing platform to conduct large-scale model training and improve training efficiency:

[0258] Distribute training data and model parameters to multiple computing nodes for parallel training. Use parameter server architecture for parameter synchronization and update. Formula:

[0259]

[0260] Among them, θ t is the model parameter at time t, η is the learning rate, N is the number of node samples, is the loss function of the i-th node sample.

[0261] Preferably, realize the elastic expansion and fault tolerance mechanism of the system to ensure the high availability and stability of the system:

[0262] Dynamically expand or reduce computing resources according to task load to ensure that the system has sufficient computing power during peak hours. Through health checks and failover mechanisms, node failures can be quickly detected and handled to ensure continuous operation of the system. Formula:

[0263]

[0264] Among them, R t+1 is the resource allocation at time t+1, L t is the task load at time t, and ΔR is the resource adjustment amount.

[0265] Preferably, cloud storage services are used to efficiently manage and store large amounts of training data, model parameters, and inference results.

[0266] The training data and model parameters are stored in a distributed cloud storage system to ensure high availability and security of the data. Data version control and access control policies are used to manage data updates and access rights. Formula:

[0267] D t+1 =Store(D t ,P t) (43)

[0268] Among them, D t+1 is the data storage state at time t+1, D t is the stored data at time t, P t is the storage strategy at time t.

[0269] Preferably, the real-time reasoning and online learning of the classification model are realized to improve the response speed and classification accuracy of the system:

[0270] Deploy the classification model to the inference service of the cloud computing platform to process real-time commodity classification requests. Use online learning algorithms to dynamically update model parameters based on new data to keep the model up to date and accurate. Formula:

[0271]

[0272] Among them, θ t is the model parameter at time t, η is the learning rate, is the loss function, x t is the new input data, y t For new label data.

[0273] S7. Deploy the trained classification model to the cross-border e-commerce platform.

[0274] Specifically, use containerization technology (such as Docker) to encapsulate the classification model to ensure the portability and ease of deployment of the model. Specific steps:

[0275] Create a Dockerfile file to define the model's dependency environment and startup script.

[0276] Use Docker to build the model container image, ensuring that all dependencies and model codes are fully packaged in the image.

[0277] Dockerfile example:

[0278] FROM python:3.8-slim

[0279] COPY. / app

[0280] WORKDIR / app

[0281] RUN pip install-r requirements.txt

[0282] CMD["python","app.py"]

[0283] Among them, the Dockerfile file defines the building process of the model container, including the base image, working directory, dependency installation, and startup commands.

[0284] Furthermore, we use container orchestration tools such as Kubernetes to deploy and manage model services to ensure high availability and scalability of services. Specific steps:

[0285] Write Kubernetes deployment files to define the deployment strategy and resource quota of the model service.

[0286] Use Kubernetes to deploy and manage model services to achieve automatic expansion, load balancing, and fault recovery.

[0287] Deployment file example:

[0288]

[0289] The deployment file defines the number of replicas, image name, port, and other configurations of the model service.

[0290] Preferably, the access entry of the model service is managed through an API gateway (such as Kong, Istio) to achieve unified API management and security control. Specific steps:

[0291] Configure the routing rules of the API gateway to forward external requests to the corresponding model service. Set access control policies (such as authentication, rate limiting, etc.) to ensure the security and stability of the service. Formula:

[0292] R api =Route(Request,Service) (45)

[0293] Among them, R api It is the API routing rule, Request is the external request, and Service is the model service.

[0294] Preferably, deploy a real-time monitoring and log management system to ensure visualization of the running status of the model service and rapid location of problems. Specific steps:

[0295] Use Prometheus to monitor the performance indicators of the model service (such as number of requests, response time, resource usage, etc.).

[0296] Use the ELK (Elasticsearch, Logstash, Kibana) stack to collect and analyze log data, and implement centralized management and visual display of logs. Monitoring indicator formula:

[0297]

[0298] Among them, M req is the total number of requests, and Requests(t) is the number of requests at time t.

[0299] Furthermore, we adopt model version control and rolling update strategies to ensure smooth upgrade and rollback of models and reduce the impact of updates on services. We label each model version to ensure that different versions of models can be deployed and managed independently. We use rolling update strategies to gradually replace old versions of models to ensure service continuity and stability. Formula:

[0300] V t+1 =Update(V t ) (47)

[0301] Among them, V t+1 is the model version at time t+1, V t is the model version at time t.

[0302] Preferably, perform A / B testing on the model, compare the effects of different versions of the model, and select the best model version for deployment. Randomly distribute user traffic to different versions of the model, and collect performance data of each version. Analyze and compare the effect indicators of each version of the model, and select the model version with the best performance for comprehensive deployment. Formula:

[0303]

[0304] Among them, E best The model version with the best effect, Metric(V i ) is the performance indicator of the i-th model version.

[0305] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A cross-border e-commerce commodity classification method based on cloud computing and deep learning, It is characterized in that The method comprises: S1. Collect multimodal commodity data from cross-border e-commerce platforms and pre-process the multimodal commodity data, and perform alignment and standardization on the pre-processed multimodal commodity data; S2, performing multimodal feature extraction and feature fusion on the preprocessed multimodal commodity data to obtain a multimodal fusion feature vector; S3. Use model-independent meta-learning to design a meta-learning framework, perform multi-task training and few-shot learning on the meta-learning framework, introduce feature enhancement and self-supervised learning in the few-shot learning of the meta-learning framework, and use the parameters of the few-shot learning of the meta-learning framework to classify and predict the fused feature vector; S4. Use graph neural network to construct a product relationship graph for the products and their relationships on the cross-border e-commerce platform, and the product relationship graph outputs the final product feature representation; S5. Build a classification model based on deep learning, use the final product feature representation as the input of the classification model, deploy the classification model to the cloud computing platform, use distributed computing resources for parallel training, and then evaluate and verify the classification model; S6, dynamically allocate and manage cloud computing resources, and use the cloud computing platform to perform distributed training on the classification model; S7. Deploy the trained classification model to the cross-border e-commerce platform; Wherein, the step S3 specifically includes: S31. Design a meta-learning framework: The multimodal fusion feature vector F obtained by S2 final Input into the meta-learning model. In the learning phase, the meta-learning model learns the initial parameter θ. Then, the meta-learning model is trained on multiple tasks to optimize the initial parameter θ, which is expressed as follows: Among them, θ is the initial parameter, T i is the task, p(T) is the task distribution, is the loss function, α is the learning rate, θ * represents the learned model parameters, It means to find the gradient of the initial parameter θ. Represents the loss function The initial parameter θ to be minimized, Represents task T i The loss function on ; S32. Define different commodity classification tasks to form a task set T. In the multi-task training phase of the meta-learning model, update the initial parameter θ: Randomly select task T from the task set T i , for each task T i , the meta-learning model uses a small number of samples to adapt, calculate the gradient and update the parameters. After each task training, the updated parameters are back-propagated to optimize the global parameter θ all , which is expressed as follows: Among them, β represents the meta-learning rate; S33. After the meta-learning phase is completed, few-sample learning optimization is performed for the specific new product classification task: A small number of training samples of new products are input into the meta-learning model, and the learned model parameters θ are used * The training samples and meta-learning model are adapted, and the model is fine-tuned using a small number of samples to optimize the parameters of the new product classification task, as shown below: Among them, θ new represents the optimization parameters of the new task, represents the loss function of the new task, Represents the learned model parameters θ * Find the gradient; S34. Based on the optimization of few-shot learning for specific new product classification tasks, feature enhancement and self-supervised learning techniques are introduced to optimize the training of few-shot learning: Using self-supervised learning methods, we generate the feature vector F by pre-training on unlabeled data. ssl , the feature vector F generated by self-supervised learning ssl And the feature vector F generated by few-shot learning few Fusion is performed to obtain the enhanced feature vector F enhanced , which is expressed as follows: F ssl =SelfSupervisedLearning(X) F enhanced =F few +λF ssl Among them, X represents unlabeled data, λ represents fusion weight, and SelfSupervisedLearning represents self-supervised learning; S35, use the optimized parameter θ new The fusion feature vector F enhanced To make a categorical prediction: The fused feature vector F enhanced Input into the meta-learning model optimized by few-shot learning, and use the optimized meta-learning model to optimize the parameters θ of the new task new Perform classification prediction and output the final product category, as shown below in, Indicates the predicted product category, and Predict indicates the classification prediction function.

2. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 1, It is characterized in that The multimodal product data includes images, text descriptions, user comments, and voice information.

3. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 2, It is characterized in that The preprocessing includes image enhancement, text data preprocessing, and speech data preprocessing; the alignment operation includes image-text alignment and text-speech alignment; the standardization processing includes image standardization, text standardization, and speech standardization.

4. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 1, It is characterized in that In S2, an improved visual Transformer model is used to extract image features; an improved BERT model is used to extract text features; and an improved speech Transformer model is used to extract speech features; wherein the improved visual Transformer model is represented as follows: F i =WeightedAverage(TransformerEncoder(P(I′+PosEmb))) Where I′ represents the preprocessed image, P represents the image block flattening operation, PosEmb represents the position embedding, TransformerEncoder represents the Transformer encoder, and F i Represents the final image feature vector; The improved BERT model is used to extract text features, which is expressed as follows: F t =WeightedAverage(BERTEncoder(T″)) Where T″ represents the preprocessed text, BERTEncoder represents the BERT encoder, and F t Represents the final text feature vector; The improved speech Transformer model is expressed as follows: F s =WeightedAverage(TransformerEncoder(Q(FFT(S′)+PosEmb))) Where S′ represents the preprocessed speech signal, FFT represents Fourier transform, Q represents framing operation, PosEmb represents position embedding, TransformerEncoder represents Transformer encoder, and F s represents the final speech feature vector, and WeightedAverage represents the weighted average function; The final image feature vector F is extracted i , the final text feature vector F t And the final speech feature vector F s Use the fully connected layer to reduce the dimension of the concatenated feature vector to obtain the final multimodal feature vector F multi , which is expressed as follows: F multi =FC([F i ;F t ;F s ]) Among them, [F i ; F t ; F s ] represents the concatenation operation of feature vectors, and F init =[F i ; F t ; F s ], F init represents the initial fusion feature vector, FC represents the fully connected layer, and F multi Represents the final multimodal feature vector.

5. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 4, It is characterized in that The S2 also includes: using the self-attention mechanism to initially fuse the feature vector F init Perform weighted processing to generate a weighted fusion feature vector F attn , use the deep fusion network to weight the fusion feature vector F attn Processing is performed to generate the final fusion feature vector F final .

6. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 1, It is characterized in that The S4 comprises: S41. Construct commodity relationship diagram: Define the nodes of the graph as each product v i , define edge e based on the relationship between commodities ij ; Among them, the node characteristics The multimodal feature vector F of the product final Indicates that the edge feature e ij It is represented by the commodity relationship characteristics, as follows: G=(V,E,H (0) ,Q) Among them, V represents the node set, E represents the edge set, and H (0) represents the initial feature matrix of the node, Q represents the edge feature matrix; S42. Design multiple graph convolutional layers to update node features layer by layer and capture the relationship information between products: Each graph convolutional layer receives the node features H( l ) and edge features E as input, then calculate the weighted sum of the neighbor node features of each node, perform aggregation operation, and use nonlinear activation function to process the aggregated features to obtain the updated node feature h (l+1) , which is expressed as follows: in, represents the feature vector of node i at layer l+1, N(i) represents the set of neighbor nodes of node i, c ij is the normalization coefficient, W (l) represents the weight matrix of the lth layer, b (l) represents the bias term, σ represents the nonlinear activation function, and j represents node j; S43. Introduce the attention mechanism in the graph convolution layer to weight the features of neighbor nodes: For each node, the attention weight is calculated for its neighboring nodes. At the same time, the neighboring node features are weighted and aggregated according to the attention weight, and the node features are updated, which is expressed as follows: in, represents the attention weight between node i and node j, represents the normalized attention weight, a represents the attention parameter vector, || represents the vector connection operation, LeakyReLU represents the LeakyReLU activation function, represents the attention weight between node i and node k, represents the feature vector of node i in layer l, Represents the feature vector of node j in layer l; S44. Generate a global graph embedding vector by globally aggregating all node features to capture the global information of the entire product relationship graph: Use the global aggregation function to aggregate all node features and generate a global graph embedding vector h G , which is expressed as follows: Among them, h G represents the global graph embedding vector, GlobalPooling represents the global aggregation function, Represents the feature vector of node i in the Lth layer; S45. Perform graph embedding and multimodal feature fusion: Embed the global graph into vector h G and the multimodal fusion feature vector F multi Concatenate and use the fully connected layer to process the concatenated feature vectors to generate the final product feature representation F final , which is expressed as follows: F final =FC([h G ;F multi ]) Among them, [h G ; F multi ] represents the concatenation operation of feature vectors, FC represents the fully connected layer, and F final Represents the final product feature representation.

7. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 1, It is characterized in that The loss function of the deep learning-based classification model is expressed as follows: in, represents the cross entropy loss, N represents the number of computing node samples, y true,i and i Represent the true label and predicted probability of the i-th node sample respectively; In the loss function The L2 regularization term is added to control the complexity of the model parameters, which is expressed as follows: in, is the loss function after regularization, λ is the regularization coefficient, is the kth item in the model parameters; The classification model is deployed to the cloud computing platform and parallel training is performed using distributed computing resources, as shown below: W f ←CloudTrain(X,y,W f ,b f ) Among them, CloudTrain represents the operation of training on the cloud computing platform, W f and b f They represent the weight and bias parameters of the updated classification model respectively, X represents the input feature matrix of the classification model, and y represents the output product category probability distribution.

8. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning according to claim 7, It is characterized in that The distributed training of the classification model using the cloud computing platform is expressed as follows: Among them, θ t represents the model parameters at time t, η represents the learning rate, N represents the number of computing node samples, represents the loss function of the i-th node sample, Represents the loss function for the i-th node Find the gradient.

9. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning as claimed in claim 7, It is characterized in that In the S7, it also includes encapsulating the classification model using containerization technology and performing A / B testing on the classification model.

10. According to the cross-border e-commerce commodity classification method based on cloud computing and deep learning as claimed in claim 7, It is characterized in that The classification model uses the Adam optimizer to calculate the gradient and update the classification model parameters to minimize the cross entropy loss, which is expressed as follows: Among them, W f and b f They represent the classification model parameters, η represents the learning rate, and denote the first-order and second-order moment estimates respectively, and ∈ denotes a stable term.

Images