Garment display method and device based on artificial intelligence

Abstract

The invention discloses a clothing display method and device based on artificial intelligence, and relates to the technical field of artificial intelligence. Automatically generating a plurality of candidate clothing display maps by using a map generator under constraint conditions of store space layout and clothing combination; training a graph neural network discriminator model by taking the historical clothing display graph and the total sales of the corresponding time period as training samples; constructing a collaborative optimization framework based on a graph generator and a graph neural network discriminator model, continuously updating parameters of the graph generator, and gradually generating a clothing display graph with higher sales potential; and after the optimization process is converged, outputting an optimal clothing display map, and mapping the optimal clothing display map into a binding relationship between the store clothing and the booth. According to the method, the problems that a traditional display mode is low in efficiency and the sales potential is difficult to optimize are solved by ingeniously utilizing the synergistic effect of the graph generator and the graph neural network discriminator model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more particularly to an artificial intelligence-based clothing display method and device. Background Art

[0002] As a core area of ​​fashion consumption, the apparel retail industry is deeply influenced by consumer preferences and market trends. Store display design is directly related to the visual appeal of merchandise and purchasing decisions. It typically involves placing various types of clothing in specific displays, creating spatial connections to guide customers' browsing and purchasing. Traditional display practices rely on manual experience and intuitive judgment, combined with store layout, such as display layout and aisle design, and clothing attributes such as color and style. However, in practice, store space is limited, layouts are complex, and clothing combinations are diverse, making it difficult to fully evaluate the impact of different arrangements on sales. Especially under the influence of external factors such as fluctuating customer traffic and seasonal changes, existing display methods often fail to effectively capture potential sales opportunities and are prone to problems such as unreasonable layouts or combination conflicts, thereby limiting the room for overall sales growth. These challenges highlight the need for a more intelligent approach to handle the complex interactions between space and combination to achieve more precise display optimization. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a clothing display method and device based on artificial intelligence to solve the problems mentioned in the background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A clothing display method based on artificial intelligence comprises the following steps: Constructing a clothing display graph, wherein the clothing display graph uses the clothing to be displayed as nodes, and establishing connecting edges between the nodes based on spatial adjacency relationships; Using the graph generator, multiple candidate clothing display graphs are automatically generated under the constraints of store space layout and clothing combination; Using historical clothing display graphs and their corresponding time period total sales as training samples, we trained a graph neural network discriminator model to predict the impact of any clothing display graph on total sales. A collaborative optimization framework is constructed based on the graph generator and the graph neural network discriminator model. The prediction results of the graph neural network discriminator model are used as feedback signals to continuously update the graph generator parameters so that it can gradually generate clothing display graphs with higher sales potential. After the optimization process converges, the optimal clothing display map is output, and the optimal clothing display map is mapped to the binding relationship between store clothing and exhibition booths to guide the actual clothing display layout.

[0005] Furthermore, the specific steps of automatically generating a plurality of candidate clothing display maps include: Initialize the graph generator model, use a random noise vector or a preset seed as input, combine the constraints of the store space layout and the clothing combination, and gradually build the connection edges between clothing nodes through a graph convolutional network or an attention mechanism; In each iteration, Monte Carlo sampling or genetic algorithm mutation operations are applied to generate diverse candidate graph variants, and invalid graphs that do not meet the store space layout or clothing combination constraints are filtered out through the constraint verification module; Output multiple candidate clothing display graphs that meet the constraints as the input basis of the collaborative optimization framework.

[0006] Furthermore, the information of the store space layout includes any one or more of the following: location coordinates of all booths, booth types, and booth sizes; The clothing combination constraint information includes any one or more of the following: clothing type matching rules, clothing color matching rules, and clothing style compatibility rules.

[0007] In some embodiments, the construction of the constraint verification module includes the following steps: Define a constraint rule library, including store space layout constraint rules and clothing combination constraint rules; A rule engine is constructed and the constraint rule library is integrated using a decision tree algorithm to implement node-by-node and edge-by-edge verification of the candidate clothing display graph.

[0008] Furthermore, the training target of the graph neural network discriminator model is the sales impact component corresponding to the clothing display graph, and the sales impact component is obtained by the following method: Select sample pairs with multiple different clothing display graphs in historical samples but with basically the same external factors; For each sample pair, based on the principle of control variables, calculate the sales difference corresponding to different clothing display maps; The sales difference is used as the training target of the graph neural network discriminator model to learn the relative impact of the display graph structure on sales.

[0009] Furthermore, the external factors include any one or more of the following: store location, customer flow, marketing campaign information, weather data, and seasonal attributes; The basically consistent determination method is: Multi-dimensional attribute encoding of historical samples; Use similarity matching algorithm to pair historical samples; On the premise that the similarity threshold meets the preset conditions, the matched sample pairs are regarded as basically consistent sample pairs.

[0010] Furthermore, the attribute information of each node in the clothing display graph includes any one or more of the following: clothing category, style label, color code, price range; The connection edge attributes of the clothing display graph include any one or more of the following: physical distance, display angle difference, and visual occlusion relationship.

[0011] Furthermore, the graph neural network discriminator model is constructed using a graph convolutional network or a graph attention network or a Transformer model based on a graph structure.

[0012] Furthermore, the optimization process adjusts the parameters of the atlas generator through reinforcement learning, genetic algorithm or gradient-based method, so that the clothing display atlas generated by it is gradually optimized in the score of the graph neural network discriminator model.

[0013] The present invention also discloses a device for implementing the artificial intelligence-based clothing display method, comprising: A graph generator is used to automatically generate multiple clothing display graphs under the constraints of store space layout and clothing combination. The clothing display graphs use the clothing to be displayed as nodes, and establish connecting edges between nodes based on spatial adjacency; A graph neural network discriminator model is used to predict the impact of any clothing display graph on total sales, using historical clothing display graphs and their corresponding time periods as training samples; A collaborative optimization framework is used to use the prediction results of the graph neural network discriminator model as feedback signals to continuously update the parameters of the graph generator, so that the graph generator gradually generates clothing display graphs with higher sales potential; The mapping output unit is used to output the optimal clothing display map after the optimization process converges, and map the optimal clothing display map into a binding relationship between store clothing and exhibition booths to guide the actual clothing display layout.

[0014] The advantage of the present invention over the prior art is that it constructs a clothing display map and uses a map generator to automatically generate candidate solutions under the constraints of store space layout and clothing combination, and then combines the graph neural network discriminator model to predict sales and build a collaborative optimization framework, and finally outputs the optimal map mapping as the actual binding relationship, which effectively solves the inefficiency problem caused by the traditional clothing display relying on manual experience, and can automatically generate layouts with high sales potential, improve the overall performance of the store and reduce human errors. During the optimization process, the discriminator model is trained based on historical data, providing a reliable feedback signal, so that the generator gradually tends to a better solution, thereby significantly improving the scientific nature and adaptability of the display and further reducing operating costs. The present invention also introduces a constraint verification module to ensure that the generated map strictly complies with the spatial and combination rules, avoiding invalid solutions from interfering with the optimization; at the same time, by matching the training objectives of the sales influencing components with external factors, the model's accurate prediction of relative sales is enhanced, avoiding confusion caused by external interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a flow chart of the generation process of the clothing display atlas of the present invention. DETAILED DESCRIPTION

[0016] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0017] The present invention provides an artificial intelligence-based clothing display method and device, which generates a clothing display plan that can significantly increase store sales by constructing a clothing display graph and utilizing a collaborative optimization framework of a graph generator and a graph neural network discriminator model.

[0018] More specifically, Figure 1 Shown is a flow chart of the method of the present invention, comprising: Constructing a clothing display graph, wherein the clothing display graph uses the clothing to be displayed as nodes, and establishing connecting edges between the nodes based on spatial adjacency relationships; Using the graph generator, multiple candidate clothing display graphs are automatically generated under the constraints of store space layout and clothing combination; Using historical clothing display graphs and their corresponding time period total sales as training samples, we trained a graph neural network discriminator model to predict the impact of any clothing display graph on total sales. A collaborative optimization framework is constructed based on the graph generator and the graph neural network discriminator model. The prediction results of the graph neural network discriminator model are used as feedback signals to continuously update the graph generator parameters so that it can gradually generate clothing display graphs with higher sales potential. After the optimization process converges, the optimal clothing display map is output, and the optimal clothing display map is mapped to the binding relationship between store clothing and exhibition booths to guide the actual clothing display layout.

[0019] More specifically, the clothing display map is used to represent the spatial and logical relationships between clothing to be displayed.

[0020] The graph uses clothing items as nodes, with each item acting as a separate node. The edges connecting the nodes are determined based on the spatial proximity of the items within the store. For example, if two items are placed adjacent to each other on a display, a connecting edge is created in the graph. To ensure the graph fully reflects the actual display of clothing, each node's attribute information includes clothing category (e.g., tops, pants, skirts), style label (e.g., casual, formal, sporty), color coding (represented in RGB or Lab color space), and price range (e.g., 0-100 yuan, 100-300 yuan, etc.).

[0021] These attributes are extracted from clothing images through manual annotation or image recognition technology. For example, color coding can be done by analyzing the main color of the clothing image and extracting RGB values ​​using the k-means clustering algorithm.

[0022] The attributes of connecting edges reflect the spatial relationships between garments, including physical distance (based on the actual measured distance between display booths), display angle difference (measured in degrees, representing the difference in the angle of the garments on the display relative to the customer's line of sight), and visual occlusion (a Boolean value indicating whether a garment is partially obscured by another garment). For example, if a top is placed at the front of a display rack and a pair of pants is behind it, the visual occlusion relationship might be recorded as "true." The specific values ​​of these attributes are obtained through digital scanning of the store layout or manual measurement, ensuring that the atlas accurately reflects the actual display environment.

[0023] In order to generate diverse candidate clothing display graphs, the method of the present invention adopts a graph generator model, combined with the constraints of store space layout and clothing combination, to automatically construct multiple feasible graph variants.

[0024] The specific steps are as follows: First, initialize the graph generator model. The graph generator takes a random noise vector or a preset seed as input. Combined with store layout constraints (such as booth location coordinates, type, and size) and clothing combination constraints (such as type matching rules, color matching rules, and style compatibility rules), it gradually constructs edges connecting clothing nodes using a graph convolutional network or a graph attention network. The random noise vector can be set to a normally distributed vector with a length of 128, with a mean of 0 and a standard deviation of 1. The preset seed can be an encoding vector based on a historically high-selling display scheme, obtained by feature extraction from the historical graph.

[0025] During the generation process, the graph generator learns the potential associations between clothing nodes through multi-layer graph convolution operations. For example, each layer of the graph convolution network updates the node features through the following formula: ; in, Representation node In the The feature vector of the layer, For nodes The set of neighbor nodes of is the normalization coefficient, and are the weight matrix and bias vector respectively, is the ReLU activation function. The atlas generator can contain 3-5 graph convolution layers, with the hidden layer dimension set to 64 or 128 to balance model complexity and computational efficiency.

[0026] To increase graph diversity, Monte Carlo sampling or genetic algorithms are used to mutate the generated graphs in each iteration. Monte Carlo sampling generates variants by randomly adjusting edge weights or adding or removing edges, with the weight adjustment range controlled within ±10%. Genetic algorithms use crossover and mutation operations, iterating with a population size (e.g., 50 candidate graphs) and a mutation probability set between 0.1 and 0.3. Generated candidate graphs are screened using a constraint verification module to ensure that all constraints are met.

[0027] The constraint validation module, built on a rules engine, incorporates store layout constraints, such as booth size limits and spacing requirements, as well as clothing combination constraints, such as prioritizing similarly styled clothing adjacent to each other and complementary color combinations. The rules engine utilizes a decision tree algorithm to verify each node and connecting edge one by one. For example, if a rule requires that "casual tops and athletic pants should not be adjacent to each other," the decision tree checks the graph for edges that violate this rule and, if so, removes the graph. Validated candidate graphs (e.g., 50-100 per round) serve as input to the collaborative optimization framework.

[0028] The graph neural network discriminator model is used to predict the impact of clothing display graphs on total sales. Its training objective is the sales impact component. Training samples are derived from historical clothing display graphs and their corresponding time period total sales data. To accurately extract the sales impact component, it is necessary to control for the influence of external factors (such as store location, customer flow, marketing campaign information, weather data, and seasonal attributes). The specific steps are as follows: First, sample pairs with largely consistent external factors were screened from historical data. External factors were represented using multidimensional attribute encodings, such as store location encoded as latitude and longitude coordinates, customer flow as hourly visitors, and weather data including temperature. Cosine similarity or Euclidean distance was used as a similarity matching algorithm to calculate the similarity between sample pairs. A similarity threshold was set at 0.9 (range 0-1). When the similarity of a sample pair exceeded this threshold, the external factors were considered largely consistent.

[0029] For each sample pair, based on the principle of control variables, the sales difference corresponding to different clothing display maps is calculated. The sales difference is calculated using the following formula: ; in, and are the total sales corresponding to different graphs within the group. The sales difference is used as the training target of the graph neural network discriminator model to learn the relative impact of graph structure on sales.

[0030] The graph neural network discriminator model is constructed using a graph convolutional network, a graph attention network, or a graph-based Transformer model. Taking the graph attention network as an example, the model can include 3-5 graph attention layers, with hidden layer dimensions of 64 or 128, and 4-8 attention heads. Training uses the Adam optimizer, with a learning rate range of 0.001-0.005, a batch size of 32-64, and 100-200 training epochs. The loss function is mean squared error, and the goal is to minimize the difference between the predicted sales volume and the actual sales volume.

[0031] The collaborative optimization framework gradually generates clothing display graphs with higher sales potential through the interaction between the graph generator and the graph neural network discriminator model.

[0032] The framework uses the prediction results of the graph neural network discriminator model (i.e., the sales impact component) as feedback to update the parameters of the graph generator. The specific optimization process uses reinforcement learning, genetic algorithms, or gradient-based methods.

[0033] In the reinforcement learning method, the graph generator is regarded as an agent whose action space is to generate different candidate graphs, and the reward function is the prediction score of the graph neural network discriminator model.

[0034] The Proximal Policy Optimization (PPO) algorithm can be used, with a discount factor γ set to 0.99 and a learning rate range of 0.0001-0.001 for the policy network. Genetic algorithms use population evolution to retain high-scoring graph variants, with a population size of 50-100, a crossover probability of 0.8, and a mutation probability of 0.1-0.3. Gradient-based methods directly optimize the parameters of the graph generator through backpropagation, with the loss function being the negative of the discriminator's prediction score.

[0035] The optimization process converges when the discriminator's prediction score improves by less than 0.01 over 10 consecutive iterations, or when the maximum number of iterations, such as 500, is reached. After the optimization is complete, the clothing display map with the highest score is output as the optimal solution.

[0036] The optimal clothing display graph is converted into a binding relationship between store clothing and exhibition booths through the mapping output unit. Specifically, each node, i.e., clothing, is mapped to a specific exhibition booth.

[0037] In a specific embodiment, the device corresponding to the method of the present invention includes the following modules: The graph generator is built on a graph convolutional network or a graph attention network and runs on a GPU-accelerated server, such as the NVIDIA A100, with a memory requirement of 16-32GB. It is implemented in Python and relies on the PyTorch Geometric library.

[0038] The graph neural network discriminator model is also implemented using PyTorch Geometric. The training data is stored in a distributed database such as MongoDB, which supports real-time updates of historical data.

[0039] The collaborative optimization framework is implemented through reinforcement learning or genetic algorithms and runs on cloud computing platforms such as AWS or Azure to ensure high concurrent processing capabilities.

[0040] The mapping output unit uses Python, and the output results are stored in JSON format, including the mapping relationship between clothing ID and booth ID, for direct use by store staff.

[0041] For example, a clothing store located in a city business district has 10 booths, ranging in size from 0.5 to 2 square meters. The displayed garments include 20 tops, 15 pairs of trousers, and 10 jackets. When initializing the clothing display graph, node attributes include category (tops, trousers, jackets), style (casual, formal), and color (RGB encoding). A graph generator generates 100 candidate graphs. Monte Carlo sampling is used to adjust edge weights in each iteration, and a constraint verification module eliminates graphs that do not meet the rule that casual tops and formal trousers should not be adjacent. A graph neural network discriminator, trained on the past six months' sales data, predicts that the highest-scoring graph indicates that placing casual tops next to trousers of the same style can increase sales by 10%. The resulting bindings assign the 20 garments to the 10 booths, resulting in an approximately 8% increase in the store's average daily sales.

[0042] In summary, through the above-mentioned implementation methods, the present invention can efficiently generate clothing display plans that meet actual needs, significantly increase store sales, and at the same time has good scalability and practicality.

[0043] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A clothing display method based on artificial intelligence, characterized in that: The steps include: Constructing a clothing display graph, wherein the clothing display graph uses the clothing to be displayed as nodes, and establishing connecting edges between the nodes based on spatial adjacency relationships; Using the graph generator, multiple candidate clothing display graphs are automatically generated under the constraints of store space layout and clothing combination; Using historical clothing display graphs and their corresponding time period total sales as training samples, we trained a graph neural network discriminator model to predict the impact of any clothing display graph on total sales. A collaborative optimization framework is constructed based on the graph generator and the graph neural network discriminator model. The prediction results of the graph neural network discriminator model are used as feedback signals to continuously update the graph generator parameters so that it can gradually generate clothing display graphs with higher sales potential. After the optimization process converges, the optimal clothing display map is output, and the optimal clothing display map is mapped to the binding relationship between store clothing and exhibition booths to guide the actual clothing display layout.

2. The method according to claim 1, wherein The specific steps of automatically generating multiple candidate clothing display maps include: Initialize the graph generator model, use a random noise vector or a preset seed as input, combine the constraints of the store space layout and the clothing combination, and gradually build the connection edges between clothing nodes through a graph convolutional network or an attention mechanism; In each iteration, Monte Carlo sampling or genetic algorithm mutation operations are applied to generate diverse candidate graph variants, and invalid graphs that do not meet the store space layout or clothing combination constraints are filtered out through the constraint verification module; Output multiple candidate clothing display graphs that meet the constraints as the input basis of the collaborative optimization framework.

3. The method according to claim 1 or 2, characterized in that: The information of the store space layout includes any one or more of the following: location coordinates of all booths, booth types, and booth sizes; The clothing combination constraint information includes any one or more of the following: clothing type matching rules, clothing color matching rules, and clothing style compatibility rules.

4. The method according to claim 2, wherein: The construction of the constraint verification module includes the following steps: Define a constraint rule library, including store space layout constraint rules and clothing combination constraint rules; A rule engine is constructed and the constraint rule library is integrated using a decision tree algorithm to implement node-by-node and edge-by-edge verification of the candidate clothing display graph.

5. The method according to claim 1, wherein The training target of the graph neural network discriminator model is the sales impact component corresponding to the clothing display graph, and the sales impact component is obtained by the following method: Select sample pairs with multiple different clothing display graphs in historical samples but with basically the same external factors; For each sample pair, based on the principle of control variables, calculate the sales difference corresponding to different clothing display maps; The sales difference is used as the training target of the graph neural network discriminator model to learn the relative impact of the display graph structure on sales.

6. The method according to claim 5, characterized in that The external factors include any one or more of the following: store location, customer flow, marketing campaign information, weather data, and seasonal attributes; The basically consistent determination method is: Multi-dimensional attribute encoding of historical samples; Use similarity matching algorithm to pair historical samples; On the premise that the similarity threshold meets the preset conditions, the matched sample pairs are regarded as basically consistent sample pairs.

7. The method according to claim 1, characterized in that The attribute information of each node in the clothing display map includes any one or more of the following: clothing category, style label, color code, price range; The connection edge attributes of the clothing display graph include any one or more of the following: physical distance, display angle difference, and visual occlusion relationship.

8. The method according to claim 1, characterized in that The graph neural network discriminator model is constructed using a graph convolutional network or a graph attention network or a Transformer model based on a graph structure.

9. The method according to claim 1, characterized in that The optimization process adjusts the parameters of the atlas generator through reinforcement learning, genetic algorithm or gradient-based method, so that the clothing display atlas generated by it is gradually optimized in the score of the graph neural network discriminator model.

10. A device for implementing the artificial intelligence-based clothing display method according to claim 1, characterized in that: include: A graph generator is used to automatically generate multiple clothing display graphs under the constraints of store space layout and clothing combination. The clothing display graphs use the clothing to be displayed as nodes, and establish connecting edges between nodes based on spatial adjacency; A graph neural network discriminator model is used to predict the impact of any clothing display graph on total sales, using historical clothing display graphs and their corresponding time periods as training samples; A collaborative optimization framework is used to use the prediction results of the graph neural network discriminator model as feedback signals to continuously update the parameters of the graph generator, so that the graph generator gradually generates clothing display graphs with higher sales potential; The mapping output unit is used to output the optimal clothing display map after the optimization process converges, and map the optimal clothing display map into a binding relationship between store clothing and exhibition booths to guide the actual clothing display layout.

Images