A product trend prediction method and system based on zero-shot learning

Abstract

The application relates to a product popular trend prediction method based on zero sample learning and belongs to the technical field of data analysis and prediction. The method comprises the following steps: acquiring image data and text data of a product and respectively converting the data into a visual feature vector and a semantic vector; generating a style label set according to the two vectors and constructing style category semantic features; acquiring time-series heat data of each style category of the product and constructing a heat evolution sequence of each style category of the product; extracting trend features of the sequence and optimizing the trend features by adopting a time guide mechanism to obtain optimized time-series trend features; constructing a weighted adjacency matrix according to clustering results of each style category of the product; performing multi-order information diffusion processing on the matrix to generate graph features; and fusing the style category semantic features, the optimized time-series trend features and the graph features to obtain comprehensive features of each style category of the product, so that a popular trend result of the product is obtained. The application improves zero sample recognition capability and product popular trend prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of data analysis and prediction technology, and in particular to a product popularity trend prediction method and system based on zero-shot learning. Background Technology

[0002] As a concentrated reflection of consumer preferences, aesthetic orientations, and core needs, fashion trends directly drive adjustments in market supply and demand and iterative product design, increasingly impacting corporate market competitiveness. Accurate trend forecasting can provide forward-looking guidance for product development and optimize production planning and supply chain management efficiency. However, fashion trends exhibit significant dynamic evolutionary characteristics, displaying phased, cyclical, and non-linear development patterns, with emerging styles constantly appearing, leading to rapid changes in consumer preferences and posing a significant challenge to trend forecasting. Traditional trend forecasting methods rely primarily on qualitative analysis based on human experience, which is limited by individual cognitive biases of analysts, resulting in highly subjective and unreliable predictions. Furthermore, manual analysis suffers from inherent bottlenecks in data processing efficiency and scale, hindering the rapid integration and analysis of multi-dimensional market data, leading to significant lag in prediction results and failing to meet current market demands for accuracy and real-time performance.

[0003] To overcome the limitations of traditional methods, trend analysis is gradually transforming towards data-driven and intelligent approaches. The application of advanced technologies such as big data and artificial intelligence has provided new possibilities for accurate prediction. Among these, visual form analysis, which can directly focus on core visual elements such as product appearance, design style, color matching, material texture, and style characteristics, has become one of the core technical paths for trend prediction. However, existing visual analysis methods are essentially static perception tools, lacking the ability to deeply model the evolutionary patterns of trends. Analysis models built solely based on image content struggle to capture the temporal context and potential causal relationships in style evolution, severely impacting the accuracy and foresight of trend prediction.

[0004] In practical applications, the rapid emergence of new styles often leads to a scarcity of samples. Traditional deep learning models heavily rely on a large number of labeled samples when dealing with such "unknown styles," exhibiting weak generalization ability and failing to effectively adapt to the needs of dynamic style updates. Zero-Shot Learning (ZSL) technology offers a potential solution to this problem. Its core idea is to leverage the semantic association between known and unknown categories, using a semantic transfer mechanism to identify and infer unseen categories, thus breaking through the dependence of traditional supervised learning on labeled samples. Applying zero-shot learning to trend prediction can not only alleviate the dilemma of insufficient new style samples but also improve the model's flexibility and scalability in recognizing novel categories, enhancing the system's practicality and generalization ability.

[0005] Large Language Models (LLMs), with their powerful data processing and multimodal fusion capabilities, are widely used in zero-shot tasks. They can perform tasks such as image classification, style recognition, and text generation through natural language prompts, demonstrating excellent knowledge transfer and semantic generalization abilities. However, existing LLMs still have significant technical shortcomings in multimodal trend analysis tasks, including: first, they mostly adopt a single-path prediction architecture, which does not fully utilize the effective information in multimodal data, resulting in limited prediction accuracy; second, they lack the ability to model the temporal dynamics of trend evolution, making it difficult to accurately depict the temporal changes in trends; and third, they have not built a modeling mechanism for the intrinsic association between style tags, which limits the semantic transfer and linkage analysis effects between styles and cannot meet the needs of complex trend prediction.

[0006] Existing zero-shot learning techniques typically construct a mapping relationship between visual features and semantic descriptions, projecting input data into a pre-defined semantic space (composed of attribute vectors or text descriptions), and then calculating the similarity between samples and semantic categories to identify unknown categories. This approach is practical in scenarios with relatively stable category structures, but it still has significant limitations, including: firstly, a strong dependence on a fixed semantic space; when new concept combinations or semantic changes occur in the analyzed object, the pre-defined semantic descriptions cannot fully cover the actual scenario, leading to a decline in model generalization ability; secondly, existing methods mostly focus on static recognition tasks, lacking the ability to model the temporal evolution of the analyzed object, and thus failing to support the practical needs of high-order trend analysis. In recent years, the development of large-scale pre-trained models and multimodal large models has provided a new technical path for zero-shot inference without explicitly constructing a semantic space, driving the application upgrade of zero-shot learning in the field of trend prediction.

[0007] In the research of trend reasoning based on large language models, existing technologies are gradually developing from image-text alignment to multimodal fusion. Existing technologies, such as integrating GPT-3 into image-text recommendation systems and achieving multi-turn interactive style responses by combining image retrieval with GPT, largely remain at the retrieval and recommendation level, lacking systematic modeling of trend evolution patterns. Furthermore, while existing technologies utilize ChatGPT to generate fashion content, demonstrating the expansion potential of large language models, they still have significant shortcomings in temporal dynamic modeling and structured reasoning.

[0008] To address the temporal evolution of fashion trends, related research has attempted to incorporate time factors to optimize prediction models. Early studies primarily relied on traditional statistical models such as ARIMA. For example, combining ARIMA models with basic neural networks for apparel sales data analysis and skirt fashion trend prediction, while validating the interpretability and short-term predictive advantages of traditional statistical methods in specific domains, could only model simple or cyclical data patterns, making it difficult to adapt to the highly volatile real-world scenarios of fashion trend data. With the development of deep learning, recurrent neural networks (RNNs) and their variant, Long Short-Term Memory networks (LSTMs), have demonstrated excellent performance in time series prediction. Existing technologies include the KERN model, which incorporates style semantic knowledge into the LSTM structure, and the REAR model, proposed based on the KERN model. However, these models still fail to achieve effective interactive modeling of visual features and the time dimension.

[0009] In recent years, trend prediction has entered a multimodal time modeling stage, with researchers beginning to explore the integration of image, text, and time-series data to optimize prediction performance. Existing technologies include the Visuelle model, which integrates product metadata, images, and Google Trends search indexes to solve the prediction challenge of new products without historical sales records; and the MuQAR architecture, which combines image attribute time series with a non-autoregressive mechanism to achieve a balance between prediction accuracy and delayed response in new product popularity trend modeling. However, these models lack the construction of intrinsic relationships between style semantics, making their prediction results susceptible to data timeliness and modal bias, and their stability needs improvement.

[0010] Style tags, as the core semantic units for trend prediction, possess a hierarchical structure and semantic relevance. However, existing technologies often treat them as independent classification targets, neglecting the inherent connections between tags and limiting the model's semantic transfer and trend linkage analysis capabilities. For example, DeepFashion improved the performance of clothing image classification and retrieval by constructing a large-scale style and attribute tag dataset, but it only used tags as supervisory signals and did not model the semantic relationships between tags.

[0011] In summary, existing trend prediction technologies still have many shortcomings in terms of multimodal information fusion, time dynamic modeling, style semantic association mining, and zero-shot generalization capabilities, and cannot fully adapt to the complex and dynamic market trend prediction needs in the future industrial context. Summary of the Invention

[0012] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art, such as low accuracy of trend prediction, insufficient foresight, weak generalization ability, insufficient utilization of multimodal information, lack of time dynamic modeling, and ineffective mining of style tag semantic association.

[0013] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a product popularity trend prediction method based on zero-shot learning, comprising:

[0014] Acquire product data, which includes image data and text data; convert the image data and text data into visual feature vectors and semantic vectors, respectively.

[0015] The visual feature vector and the semantic vector are jointly input into a multimodal large model to generate a style tag set; the style tag set is preprocessed to obtain a normalized tag set; the normalized tag set is clustered to construct style category semantic features;

[0016] Acquire time-series popularity data for each style category of the product and construct a popularity evolution sequence for each style category; extract trend features from the popularity evolution sequence to obtain time-series trend features; optimize the time-series trend features using a time-guided mechanism to obtain optimized time-series trend features;

[0017] The clustering results of each style category of the product are used as a set of graph nodes, and a weighted adjacency matrix is ​​constructed using the semantic similarity of the labels and the frequency of historical occurrence. Based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles.

[0018] The semantic features of the style categories, the optimized time-series trend features, and the graph features are fused to obtain the comprehensive features of each style category of the product; the comprehensive features are then used to infer the popular trend results for the future period.

[0019] In one embodiment of the present invention, the step of optimizing the time-series trend features using a time-guided mechanism to obtain optimized time-series trend features is as follows:

[0020] Obtain the popularity distribution of any style category within multiple time windows, and construct the popularity evolution sequence of the current style category; wherein, the popularity evolution sequence is a sequence formed by arranging the popularity values ​​corresponding to each time window in chronological order;

[0021] The heat evolution sequence is subjected to a time model to extract trend features, thereby obtaining time-series trend features;

[0022] By using the decay function as a time-guided mechanism, the time weights of the time-series trend features are redistributed to enhance the influence of recent time-series data on trend prediction, thereby obtaining optimized time-series trend features.

[0023] In one embodiment of the present invention, the expression for the optimized time-series trend feature is:

[0024] ;

[0025] in, Indicates the first Optimization time-series trend characteristics of each style category Indicates the first The temporal trend characteristics of each style category, Indicates the attenuation coefficient. Indicates the total number of time windows. Indicates a point in time.

[0026] In one embodiment of the present invention, the steps of converting the image data and the text data into visual feature vectors and semantic vectors, respectively, are as follows:

[0027] The image data is subjected to resolution normalization, illumination correction and noise suppression to obtain optimized image data;

[0028] Deep visual features of the optimized image data are extracted using a visual coding network to obtain a visual feature vector; wherein, the expression of the visual feature vector is:

[0029] ;

[0030] in, This represents the i-th visual feature vector. Represents the visual encoding function. This represents the i-th input image;

[0031] The text data is subjected to terminology standardization, lexical denoising, and syntax normalization to obtain optimized text data;

[0032] The semantic features of the optimized text data are extracted using a language encoder to obtain a semantic vector; wherein the expression of the semantic vector is:

[0033] ;

[0034] in, This represents the i-th semantic vector. This represents a text encoding function. This represents the text description corresponding to the i-th image data.

[0035] In one embodiment of the present invention, the formula for calculating each element in the weighted adjacency matrix is ​​as follows:

[0036] ;

[0037] in, Indicates style category With style category The connection weights between them This represents the i-th style category. This represents the j-th style category. and This represents an adjustable weight parameter. Indicates the semantic similarity of tags. Indicates style category With style category The conditional probability of simultaneous occurrence.

[0038] In one embodiment of the present invention, the expression for the spectral features is:

[0039] ;

[0040] in, Representation of spectral features, Represents the adjacency matrix. Degree matrix, Represents the initial feature matrix of the node. Represents the learnable weight matrix. This indicates the total number of time windows.

[0041] In one embodiment of the present invention, the expression of the style category semantic feature is:

[0042] ;

[0043] in, Represents the semantic features of style categories. This represents the embedding of the label semantic vector. Indicates the first Each style category center This represents the i-th normalized label set.

[0044] In one embodiment of the present invention, the expression for the trend result is:

[0045] ;

[0046] in, Indicates the first The results of the popular trends for each style category, Indicates the first The combined characteristics of each style category This represents the model's prediction function.

[0047] In one embodiment of the present invention, after reasoning about the comprehensive features to obtain the popular trend results in the future period, the method further includes: constructing multi-carrier fusion trend presentation content based on the optimized time-series trend features, the map features, and the popular trend results; organizing the trend presentation content through visualization and natural language description to generate a trend report.

[0048] Secondly, to solve the above-mentioned technical problems, the present invention provides a product popularity trend prediction system based on zero-shot learning, used to implement the above-mentioned product popularity trend prediction method based on zero-shot learning, comprising:

[0049] The acquisition and conversion module is used to acquire product data, which includes image data and text data; and to convert the image data and text data into visual feature vectors and semantic vectors, respectively.

[0050] The label graph space module is used to jointly input the visual feature vector and the semantic vector into a multimodal large model to generate a style label set; preprocess the style label set to obtain a normalized label set; and perform clustering processing on the normalized label set to construct style category semantic features.

[0051] The time-guided inference module is used to acquire time-series popularity data of each style category of the product and construct the popularity evolution sequence of each style category; extract trend features from the popularity evolution sequence to obtain time-series trend features; and optimize the time-series trend features using a time-guided mechanism to obtain optimized time-series trend features.

[0052] The graph feature acquisition module is used to take the clustering results of each style category of the product as a set of graph nodes, and construct a weighted adjacency matrix using the semantic similarity of the labels and the frequency of historical occurrence; based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles.

[0053] The trend reasoning module is used to fuse the semantic features of the style categories, the optimized time-series trend features, and the graph features to obtain the comprehensive features of each style category of the product; and to reason about the comprehensive features to obtain the popular trend results in the future period.

[0054] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0055] (1) The product popularity trend prediction method and system based on zero-shot learning described in this invention integrates the visual semantic features of style with historical time-series trend features and introduces a time-guided trend evolution modeling mechanism, so that the trend prediction process takes into account both current design features and historical evolution rules, thereby improving the stability and foresight of trend analysis. By constructing a semantic association network between style tags, the originally isolated style tags are transformed into association graph nodes, and the relationship and evolution path between styles are integrated into the trend reasoning process, which effectively improves the accuracy and interpretability of trend prediction results.

[0056] (2) This invention employs a simplified graph convolutional network to perform information diffusion and feature propagation on the style tag map. Under the premise of reducing computational complexity, it achieves efficient modeling and reasoning of style association relationships, enhancing the system's ability to identify and predict new styles and style evolution trends. Finally, the semantic features of style categories, optimized temporal trend features, and style map features are fused and input into a large language model to complete trend prediction. The output includes analysis results containing trend conclusions and their basis, realizing intelligent prediction and application output of clothing fashion trends under zero-sample conditions. Attached Figure Description

[0057] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0058] Figure 1 This is a flowchart of a product popularity trend prediction method based on zero-shot learning in a preferred embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of the large model trend inference process in a preferred embodiment of the present invention;

[0060] Figure 3 This is a structural diagram of a product popularity trend prediction system based on zero-shot learning in a preferred embodiment of the present invention. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0062] Example 1:

[0063] Reference Figures 1 to 2 As shown, this embodiment of the invention provides a product popularity trend prediction method based on zero-shot learning, including but not limited to the following steps:

[0064] S1. Obtain product data, which includes image data and text data; convert the image data and text data into visual feature vectors and semantic vectors, respectively.

[0065] S2. Input the visual feature vector and semantic vector into the multimodal large model to generate a style label set; preprocess the style label set to obtain a normalized label set; perform clustering processing on the normalized label set to construct style category semantic features;

[0066] S3. Obtain the time-series popularity data of each style category of the product and construct the popularity evolution sequence of each style category; extract trend features from the popularity evolution sequence to obtain time-series trend features; optimize the time-series trend features using a time-guided mechanism to obtain optimized time-series trend features.

[0067] S4. The clustering results of each style category of the product are used as a set of graph nodes, and a weighted adjacency matrix is ​​constructed using the semantic similarity of the labels and the frequency of historical occurrence. Based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles.

[0068] S5. The semantic features of style categories, optimized time-series trend features, and graph features are fused together to obtain the comprehensive features of each style category of the product; the comprehensive features are used to infer the popular trend results in the future period.

[0069] This invention discloses a product trend prediction method based on zero-shot learning. Without requiring manually labeled samples or a pre-defined fixed style category system, it utilizes a multimodal large-scale model to jointly understand images and their corresponding text data, automatically generating a style tag set to achieve zero-shot identification and analysis of unknown or emerging styles. By combining the visual semantic features of styles with historical time-series trend features, a time-guided trend evolution modeling mechanism is introduced, allowing the trend prediction process to consider both current design features and historical evolution patterns, thereby improving the stability and foresight of trend analysis. By constructing a semantic relationship network between style tags, style tags are transformed from isolated analysis objects into graph nodes with associated structures, and the relationships and evolutionary paths between styles are incorporated into the trend inference process, improving the logical consistency and interpretability of the trend prediction results. A simplified graph convolutional network is used to diffuse information and propagate features through the style tag graph, achieving efficient modeling and inference of style relationships while reducing computational complexity, enhancing the method's ability to infer new styles and style evolution trends. Furthermore, this embodiment integrates style category semantic features, optimized temporal trend features, and style map features, and inputs them into a large language model to complete trend prediction, generating analysis results that include trend conclusions and their basis, thus realizing intelligent prediction and application output of popular trends under zero-sample conditions.

[0070] This embodiment uses the apparel industry as a verification scenario. Due to its high correlation with fashion trends and the diverse data sources available in the apparel industry (such as runway visual data, historical sales timelines, and social media text), these data provide rich training materials for studying fashion trend prediction methods. Therefore, this embodiment uses apparel products as an example for detailed explanation.

[0071] Specifically, in step S1, fashion data (i.e., product data) from different time and space channels is acquired and organized, including multimodal data such as runway images, e-commerce new product display images, brand press releases, trend media texts, KOL discussion content, and historical trend popularity curves. Among them, runway images and e-commerce new product display images are image data; brand press releases, trend media texts, KOL discussion content, and historical trend popularity curves are text data.

[0072] Furthermore, the specific steps for converting image data and text data into visual feature vectors and semantic vectors, respectively, are as follows:

[0073] S110. Perform resolution normalization, illumination correction, and noise suppression on the image data to obtain optimized image data.

[0074] S120. Utilize a visual coding network to extract deep visual features from the optimized image data, obtaining a visual feature vector; where the expression for the visual feature vector is:

[0075] ;

[0076] in, This represents the i-th visual feature vector; This represents a visual encoding function used to extract visual features such as contours, textures, and colors. Let i represent the i-th input image.

[0077] S130. Perform terminology standardization, lexical denoising, and syntax normalization on the text data to obtain optimized text data.

[0078] S140. Extract and optimize the semantic features of the text data using a language encoder to obtain a semantic vector; wherein the expression of the semantic vector is:

[0079] ;

[0080] in, This represents the i-th extracted semantic vector. This represents a text encoding function. This represents the text description corresponding to the i-th image data.

[0081] For example, for image content in fashion data, resolution normalization, illumination correction, and noise suppression are performed, and deep visual feature vectors such as structural contours, silhouette segmentation, material textures, and color matching are extracted using a visual encoding network. For show reviews, keyword texts, and design descriptions corresponding to clothing content, the original text is transformed into a computable expression through terminology standardization, lexical denoising, and syntax regularization, and then semantic vectors are obtained through a language encoder.

[0082] After processing in step S1, the original image and text information in this embodiment are transformed into structured numerical representations and have the ability to be processed and inferred by large models. This solves the problems of mixed fashion corpora, inconsistent formats, and complex noise distribution, and lays the input foundation for subsequent style tag generation and trend modeling.

[0083] Specifically, in step S2, relying on the semantic understanding and cross-modal reasoning capabilities of multimodal large models such as GPT-4V, clothing style tags are automatically generated from runway images and corresponding text descriptions, without the need for manual annotation or pre-constructing a training category system. Visual feature vectors With semantic vectors After inputting a multimodal large model in the form of a combined prompt, the multimodal large model will output a set of style tags based on visual details, material combinations, shape structure, and descriptive semantics. The expression for this set is:

[0084] ;

[0085] in, The large model is the first The set of style tags generated from the input group is not fixed and may include terms related to pattern, material, structural design, or newly created naming styles. This represents the GPT-4V multimodal large model.

[0086] Furthermore, because the multimodal large model possesses the ability to combine concepts that have never appeared before, when the input clothing reflects a new trend, even if that trend is not in the historical style experience base, it can still generate new style names, achieving zero-sample style tag recognition. To ensure consistency in naming conventions, this embodiment also performs lemmatization and synonym merging on the generated tags, thereby forming a normalized tag set. The expression for the normalized tag set is:

[0087] ;

[0088] in, This indicates the label normalization process; This represents the deduplication function, which ultimately yields... To standardize the tag set, i.e. the first Group normalized label set.

[0089] Furthermore, this embodiment clusters all labels based on semantic embedding vectors to form style category semantic features (i.e., style category center matrix). A style structure boundary, i.e., the style category center matrix, is then established for subsequent relationship inference. The expression is:

[0090] ;

[0091] in, This represents the center matrix of all style categories. This represents the embedding of the label semantic vector. Indicates the first Each style category center Indicates the first A normalized label set, This represents the clustering function. This process forms a style semantic system, which is the basic expression space for trend inference.

[0092] To make the prediction future-oriented, this embodiment constructs a trend evolution trajectory based on the popularity sequence of clothing styles over time, and infers future trends through time-series modeling structure.

[0093] Specifically, in step S3, the time-guided mechanism is used to optimize the time-series trend features, and the specific steps to obtain the optimized time-series trend features are as follows:

[0094] S310, Obtain any style category The popularity distribution across multiple time windows is used to construct the popularity evolution sequence of the current style category; where the popularity evolution sequence is a sequence formed by arranging the popularity values ​​corresponding to each time window in chronological order, and the specific expression is:

[0095] ;

[0096] in, Representing style At the point of time Popularity value This represents the complete popularity evolution sequence of this style over time. This indicates the total number of time windows.

[0097] S320, The heat evolution sequence The data is input into a time-based network to obtain a trend representation. Specifically, a time-based model is used to analyze the evolution sequence of popularity. Perform trend feature extraction to obtain time series trend features. The specific expression is:

[0098] ;

[0099] in, This represents a function used by time-series trend models (such as LSTM) to extract trend features. It can also be called a time trend vector.

[0100] S330. To make the time model more focused on the latest market signals, this embodiment also designs a decay function to redistribute time weights, so that data closer to the current period contributes more. Specifically, the decay function is used as a time-guiding mechanism to influence time-series trend characteristics. Time weights are reallocated to enhance the impact of recent time series data on trend prediction, thereby obtaining optimized time series trend features. Among them, optimizing time-series trend features The expression is:

[0101] ;

[0102] in, Indicates the first The optimized time-series trend characteristics of each style category after weight adjustment make the influence greater the closer to the current cycle. Indicates the first The temporal trend characteristics of each style category Indicates the attenuation coefficient. Indicates a point in time. Indicates the time interval from the present.

[0103] Step S3 effectively solves the problems of lagging and insensitivity in traditional trend prediction, enabling the method described in this embodiment to capture the precursor signals of an impending style breakout or decline, making it an important driving variable for future trend prediction.

[0104] This embodiment does not base its style understanding on isolated label classification, but rather establishes a structural relationship network of semantic, cultural, profile, and combination paths among various styles by constructing a style knowledge graph.

[0105] Specifically, the steps for generating the map features reflecting the relationships between styles in step S4 are as follows:

[0106] S410. The clustering results of each style category of the product are used as a set of graph nodes. The expression for this set is:

[0107] ;

[0108] in, Indicates the first Style categories, This represents the total number of style categories.

[0109] S420. Construct a weighted adjacency matrix using label semantic similarity and historical frequency of occurrence. The formula for calculating each element in the weighted adjacency matrix is ​​as follows:

[0110] ;

[0111] in, Indicates style category With style category The connection weights between them This represents the i-th style category. This represents the j-th style category. and This represents an adjustable weight parameter. Indicates the semantic similarity of tags. Indicates style category With style category The conditional probability of simultaneous occurrence.

[0112] The development of Graph Neural Networks (GNNs) has provided an effective tool for modeling label relationships. For example, Graph Convolutional Networks (GCNs) propagate node features through adjacency matrices to model high-order relationships between labels, and have been widely applied to trend computing tasks such as image attribute prediction and style compatibility analysis. However, the complex neighborhood aggregation computation of GCNs leads to high computational overhead, making it difficult to adapt to the construction and inference requirements of large-scale style label knowledge graphs. To address this, Simplified Graph Convolutional Networks (SGCs) have emerged. By removing intermediate nonlinear mappings and retaining core convolutional steps, SGCs significantly reduce training and inference time while maintaining the ability to model structural relationships between labels, providing technical support for the efficient construction of large-scale style label knowledge graphs. Therefore, in step S430, a Simplified Graph Convolutional Network is used to obtain graph features.

[0113] S430. This embodiment is based on a weighted adjacency matrix and uses S... 2 GC (Simple Spectral Graph Convolution) performs multi-order information diffusion, enabling style nodes to receive neighborhood information and form multi-layered semantic structure representations, thereby obtaining graph features. The expression for the graph features is:

[0114] ;

[0115] in, The spectral features representing diffusion generation are represented. Represents the adjacency matrix. Degree matrix, Represents the initial feature matrix of the node. Represents the learnable weight matrix. This indicates the total number of time windows.

[0116] In this embodiment, the diffused spectral features It can not only represent known style relationships, but also infer the location, adjacent style groups and potential evolution direction when a completely new style appears, thereby enhancing the ability of the method described in this embodiment to understand trend structure changes.

[0117] After completing the above-mentioned standardized labels (which yields style category semantic features), time trend modeling (which yields optimized time-series trend features), and style relationship topology construction (which yields graph features), this embodiment will jointly fuse the three features and input them into a large model, which will then complete the final trend inference and result generation.

[0118] Specifically, in step S5, the style category semantic features, optimization time-series trend features, and graph features are fused to obtain the comprehensive features of each style category of the product. The expression of the comprehensive features is as follows:

[0119] ;

[0120] in, Indicates the first A center matrix for each style category is used to represent the semantic class center; Indicates the first Optimized temporal trend features for each style category are used to characterize trend dynamics; Indicates the first The graph features of each style category are used to characterize the diffusion feature representation of nodes of that category in the knowledge graph. This represents the vector concatenation operation, while This is a comprehensive characteristic of this style.

[0121] Furthermore, in this embodiment, LLM is preferably used to infer the comprehensive features, thereby generating trend results and interpretable inferences for future periods. The expression for the trend results is as follows:

[0122] ;

[0123] in, Indicates the first Results of future trends for each style category Indicates the first The combined characteristics of each style category This represents the model's prediction function. In this embodiment, the prediction results may include the probability of rise and fall, potential breakout-level new styles, the replacement path of old and new styles, and the judgment of future trend strength. The model not only outputs conclusions but also explains the reasons for the formation of trend directions, achieving transparency and interpretability in trend inference.

[0124] Furthermore, after trend prediction is completed in step S5, the process also includes: constructing multi-carrier integrated trend presentation content based on optimized time-series trend characteristics, graph characteristics, and popular trend results; organizing the trend presentation content through visualization and natural language description to generate a trend report. Specifically, the trend report will describe future popular trends, the structural evolution basis between related styles, and the logical source of trend changes in long-form natural language text. It will also automatically draw popularity change trend curves, style evolution topology diagrams, and style intensity probability distribution visualizations, and further select representative runway images for visual presentation, resulting in a comprehensive output structure that combines images, charts, and text. The final trend report is readable, analyzable, and directly applicable to brand strategy formulation and product planning decisions, achieving a closed-loop process from data calculation to commercial implementation.

[0125] The product trend prediction method described in this embodiment extracts visual semantic features of clothing design elements through LLM and combines them with a time-series encoder to capture market cycle patterns, ultimately achieving generalized prediction of clothing styles under zero-sample conditions. This method not only enables zero-sample identification of styles not covered by historical data but also predicts trend evolution based on fashion evolution patterns, thus providing more accurate and efficient trend decision support.

[0126] Furthermore, to verify the feasibility and effectiveness of the method described in this embodiment of the invention in real-world apparel trend prediction scenarios, a complete and reproducible implementation demonstration is provided. This demonstration uses quarterly runway data as input, and through the method described in this embodiment, completes multimodal information parsing, zero-sample style tag generation, trend time-series inference, and knowledge graph diffusion reasoning, ultimately generating trend prediction and applied output. This implementation demonstration aims to illustrate how the method described in this embodiment can achieve trend identification and prediction without sample labeling and independent training when users face unknown styles or new season design releases, and further demonstrates the correspondence between the model output results and real market trends, thus proving its research and industrial promotion value. The specific content is as follows:

[0127] Users first prepare the input data for analysis, namely the latest season's runway images and brand press releases. Users simply need to upload several look images (e.g., 200 images) along with corresponding text to the system used to implement this method. After uploading, the system automatically enters the first stage of processing: the images are analyzed by a visual model to extract fabric, cut, color scheme, and silhouette, while the text is standardized into phrase tags to aid understanding. In this stage, users do not need to perform any manual categorization or labeling; the system will automatically complete the processing in the background.

[0128] After preprocessing, the system enters the large-scale model inference stage. GPT-4V automatically generates design style tags based on the images and text provided by the user. For example, it may identify new trend terms such as glossy leather, minimalist silhouettes, soft mist tulle, and metallic ballet, even if these names have never appeared in historical data. This process does not rely on training samples, thus capturing entirely new styles and demonstrating the system's zero-shot capability. Subsequently, the system further categorizes these tags and forms style category centers, facilitating subsequent analysis of how trends evolve.

[0129] After style tagging is completed, the system automatically retrieves similar style data from the past two years (24 months), including sales, search volume, and social media buzz, and plots a line curve showing the style's evolution over time. The time-trend model analyzes the speed, cycle, and decay of this line's changes, allowing the system to not only know what's currently popular but also whether the trend is rising, stable, or about to fade. For example, in a particular run, the system might find that the "metal ballet" style has seen a continuous increase in discussion over the past six months, with color schemes shifting towards silver and soft pink, indicating its potential to become a breakout trend.

[0130] To make the conclusions more reliable, the system further incorporates these styles into a style association graph for diffusion inference. The system examines which styles "Metallic Ballet" is most closely associated with and analyzes whether it has merged with past styles such as Gothic Romance, Industrial Chic, or Sheer Tulle. For example, if the graph calculations find a high connection weight with the "Industrial Misty Romance" node, the system infers a potential strong rise in the "Industrial × Ballet Hybrid Style" in the next quarter and provides design and fabric suggestions; for instance, combining rigid metal buttons with soft tulle, and light silver with mother-of-pearl pink tones are safe options.

[0131] After all calculations are completed, the system generates a trend report, which users can read directly or use in planning meetings without additional editing. The report consists of three parts: a trend description summarized by the large language model, a popularity evolution curve, and a style evolution path derived from graph inference, along with representative images of the trend, similar to an automatically generated trend research briefing.

[0132] Actual simulation tests show that when 200 images and corresponding text data are input, the system completes trend tag generation and prediction in less than 3 minutes. After verification by comparing with real market performance, the search popularity of the "metallic ballet tulle" style increased by about 42% in the following quarter, which is consistent with the trend predicted by this system, verifying the reliability of the method described in this embodiment of the invention in real-world clothing trend analysis scenarios.

[0133] In summary, this demonstration showcases a complete trend prediction process that allows users to complete without manual annotation or model training. The method described in this embodiment can not only understand images and text but also combine historical trends to make future inferences and generate directly usable trend reports, demonstrating practical industrial feasibility.

[0134] The large model-based inference framework and S described in this embodiment of the invention 2 The GC-enhanced style relationship-based intelligent analysis method for clothing trends has significant advantages over traditional trend analysis methods that rely on labeled training data or expert experience. These advantages are verifiable and possess clear value for industry application. Specific advantages are as follows:

[0135] First, traditional visual models rely on a large number of labeled samples and fixed category labels, making it difficult to identify newly emerging shapes or aesthetic styles in the design industry. This invention uses a large model as the core inference subject for trend analysis. Style labels are automatically generated by inputting runway images and text, eliminating the need for manual training or expanding the label set. This fundamentally eliminates the dependence on labeling, enabling the method to understand new styles and predict future trends.

[0136] Secondly, by introducing time trend modeling, this embodiment of the invention moves beyond static classification and judgment, enabling it to predict trend direction, strength, and reversal probability. The time decay and sequence encoding mechanisms effectively address the shortcomings of traditional trend research, such as lag and insensitivity, allowing this method to capture early signals of impending style rise, surge, or decline, significantly improving the accuracy and foresight of trend prediction.

[0137] Third, in the embodiments of the present invention, S 2 GC style knowledge graphs serve as a trend inference structure, enabling the modeling of the relationships and evolutionary connections between different styles. This method not only identifies existing styles but also infers the possible evolutionary directions, fusion paths, and replacement trends of new styles based on topological propagation structures. This mechanism overcomes the traditional limitations of isolated label recognition, making trend prediction structured, interpretable, and extensible.

[0138] Fourth, the method described in this embodiment of the invention can directly output trend analysis reports, generating complete trend decision-making materials in multimodal forms such as natural language, chart visualization, style relationship networks, and trend evidence images. These materials can be directly used for brand planning, buyer procurement, fabric supply chain scheduling, and e-commerce advertising strategy decisions, without requiring additional manual analysis or secondary processing. Therefore, this method can significantly reduce trend research costs, improve new product development efficiency and market response speed, and has clear industrial value.

[0139] Therefore, the method described in this embodiment of the invention has achieved substantial improvements in zero-sample recognition capability, trend prediction accuracy, style structure interpretability, and industrial application implementation capability.

[0140] Example 2:

[0141] Based on the same inventive concept, this embodiment provides a product popularity trend prediction system based on zero-shot learning. The principle of solving the problem is similar to the product popularity trend prediction method based on zero-shot learning provided in Embodiment 1, and the repeated parts will not be described again.

[0142] Reference Figure 3 As shown, this embodiment provides a product popularity trend prediction system based on zero-shot learning, used to implement the product popularity trend prediction method based on zero-shot learning provided in Embodiment 1, including:

[0143] The acquisition and conversion module is used to acquire product data, which includes image data and text data; and to convert the image data and text data into visual feature vectors and semantic vectors, respectively.

[0144] The label graph space module is used to jointly input visual feature vectors and semantic vectors into a multimodal large model to generate a style label set; the style label set is preprocessed to obtain a normalized label set; the normalized label set is clustered to construct style category semantic features;

[0145] The time-guided inference module is used to acquire time-series popularity data for each style category of the product and construct the popularity evolution sequence for each style category; extract trend features from the popularity evolution sequence to obtain time-series trend features; and optimize the time-series trend features using a time-guided mechanism to obtain optimized time-series trend features.

[0146] The graph feature acquisition module is used to take the clustering results of each style category of the product as a set of graph nodes, and construct a weighted adjacency matrix using the semantic similarity of the labels and the frequency of historical occurrence. Based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles.

[0147] The trend reasoning module is used to fuse style category semantic features, optimized time-series trend features, and graph features to obtain comprehensive features of each style category of the product; and to reason about the comprehensive features to obtain the popular trend results in the future period.

[0148] Specifically, the product popularity trend prediction system described in this embodiment also includes a multimodal report generation module, which is used to generate trend analysis results for enterprise applications.

[0149] The product trend prediction system described in this embodiment is based on LLM, integrates visual guidance and time guidance dual-path mechanism, and introduces style tag map for structured reasoning. It can not only achieve zero-sample identification of styles not covered by historical data, but also predict trend evolution based on fashion evolution law, thus providing more accurate and efficient trend decision support.

[0150] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A product popularity trend prediction method based on zero-shot learning, characterized in that, include: Acquire product data, which includes image data and text data; convert the image data and text data into visual feature vectors and semantic vectors, respectively. The visual feature vector and the semantic vector are jointly input into a multimodal large model to generate a style tag set; the style tag set is preprocessed to obtain a normalized tag set; the normalized tag set is clustered to construct style category semantic features; Obtain time-series popularity data for each style category of the product, and construct a popularity evolution sequence for each style category; extract trend features from the popularity evolution sequence to obtain time-series trend features; The time-series trend features are optimized using a time-guided mechanism to obtain optimized time-series trend features. Specifically, a decay function is used as the time-guided mechanism to redistribute the time weights of the time-series trend features, thereby strengthening the influence of recent time-series data on trend prediction, resulting in optimized time-series trend features. The expression for the optimized time-series trend features is as follows: ； in, Indicates the first Optimization time-series trend characteristics of each style category Indicates the first The temporal trend characteristics of each style category, Indicates the attenuation coefficient. Indicates the total number of time windows. Indicates a point in time; The clustering results of each style category of the product are used as a set of graph nodes, and a weighted adjacency matrix is ​​constructed using the semantic similarity of the labels and the frequency of historical occurrence. Based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles. The semantic features of the style categories, the optimized time-series trend features, and the graph features are fused to obtain the comprehensive features of each style category of the product; the comprehensive features are then used to infer the popular trend results for the future period.

2. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The steps for extracting trend features from the heat evolution sequence to obtain time-series trend features are as follows: Obtain the popularity distribution of any style category within multiple time windows, and construct the popularity evolution sequence of the current style category; wherein, the popularity evolution sequence is a sequence formed by arranging the popularity values ​​corresponding to each time window in chronological order; The time-series trend features are obtained by extracting trend features from the heat evolution sequence using a time model.

3. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The steps for converting the image data and the text data into visual feature vectors and semantic vectors, respectively, are as follows: The image data is subjected to resolution normalization, illumination correction and noise suppression to obtain optimized image data; Deep visual features of the optimized image data are extracted using a visual coding network to obtain a visual feature vector; wherein, the expression of the visual feature vector is: ； in, This represents the i-th visual feature vector. Represents the visual encoding function. This represents the i-th input image; The text data is subjected to terminology standardization, lexical denoising, and syntax normalization to obtain optimized text data; The semantic features of the optimized text data are extracted using a language encoder to obtain a semantic vector; wherein the expression of the semantic vector is: ； in, This represents the i-th semantic vector. This represents a text encoding function. This represents the text description corresponding to the i-th image data.

4. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The formula for calculating each element in the weighted adjacency matrix is ​​as follows: ； in, Indicates style category With style category The connection weights between them This represents the i-th style category. This represents the j-th style category. and This represents an adjustable weight parameter. Indicates the semantic similarity of tags. Indicates style category With style category The conditional probability of simultaneous occurrence.

5. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The expression for the spectral features is: ； in, Representation of spectral features, Represents the adjacency matrix. Degree matrix, Represents the initial feature matrix of the node. Represents the learnable weight matrix. This indicates the total number of time windows.

6. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The expression for the semantic features of the style category is: ； in, Represents the semantic features of style categories. This represents the embedding of the label semantic vector. Indicates the first Each style category center This represents the i-th normalized label set.

7. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, The expression for the trend result is: ； in, Indicates the first The results of the popular trends for each style category, Indicates the first The combined characteristics of each style category This represents the model's prediction function.

8. The product popularity trend prediction method based on zero-shot learning according to claim 1, characterized in that, After reasoning about the comprehensive features to obtain the popular trend results in the future period, the method further includes: constructing multi-carrier integrated trend presentation content based on the optimized time-series trend features, the map features, and the popular trend results; organizing the trend presentation content through visualization and natural language description to generate a trend report.

9. A product popularity trend prediction system based on zero-shot learning, used to implement the product popularity trend prediction method based on zero-shot learning as described in any one of claims 1 to 8, characterized in that, include: The acquisition and conversion module is used to acquire product data, which includes image data and text data; and to convert the image data and text data into visual feature vectors and semantic vectors, respectively. The label graph space module is used to jointly input the visual feature vector and the semantic vector into a multimodal large model to generate a style label set; preprocess the style label set to obtain a normalized label set; and perform clustering processing on the normalized label set to construct style category semantic features. The time-guided reasoning module is used to acquire time-series popularity data for each style category of the product and construct a popularity evolution sequence for each style category; the module also extracts trend features from the popularity evolution sequence to obtain time-series trend features. The time-series trend features are optimized using a time-guided mechanism to obtain optimized time-series trend features; The graph feature acquisition module is used to take the clustering results of each style category of the product as a set of graph nodes, and construct a weighted adjacency matrix using the semantic similarity of the labels and the frequency of historical occurrence; based on the weighted adjacency matrix, multi-level information diffusion processing is performed to generate graph features that reflect the relationship between styles. The trend reasoning module is used to fuse the semantic features of the style category, the optimized time-series trend features, and the graph features to obtain the comprehensive features of each style category of the product. By reasoning about the comprehensive characteristics, the popular trend results for the future period are obtained.

Images