Method and system for interactive decision of smart shopping guide robot in clothing sales scene

Abstract

The application provides a clothing sales scene intelligent shopping guide robot interaction decision method and system, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring customer body state data and historical behavior trajectory, determining joint representation and clothing commodity adaptability score; constructing a relation graph based on candidate commodities and determining a recommended path through a graph diffusion algorithm and an attention mechanism; performing Markov chain Monte Carlo sampling and genetic algorithm to generate a dialogue action sequence and determine a recommended commodity list and personalized shopping guide rhetoric. The application realizes accurate commodity adaptation and dynamic interaction decision, and improves the intelligent level of shopping guide service and the sales conversion efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an interactive decision-making method and system for an intelligent shopping guide robot in a clothing sales scenario. Background Technology

[0002] In the apparel retail sector, the application of intelligent shopping guide systems is becoming increasingly widespread. These systems aim to improve the customer shopping experience and sales efficiency through automation technology. Existing technical solutions acquire basic body shape data of customers through image recognition or sensors, and then perform simple size or style matching in the product database to generate a preliminary product recommendation list. The recommendation logic is often based on rule-based filtering or shallow collaborative filtering algorithms, and provides shopping guidance suggestions to customers in the form of static lists or basic scripts.

[0003] However, existing technologies still have shortcomings, such as failing to effectively integrate multi-dimensional dynamic information about customers, resulting in an inability to accurately capture shopping motivations, ignoring the inherent connections between products and the guiding value of the recommendation sequence itself for the customer's decision-making path, and causing the recommendation results to appear fragmented and lacking in strategic depth. Summary of the Invention

[0004] This invention provides an intelligent shopping guide robot interactive decision-making method and system for clothing sales scenarios, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides an interactive decision-making method for an intelligent shopping guide robot in an apparel sales scenario, comprising: Acquire customer body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform suitability scoring on clothing products in a preset product library to obtain a candidate product set. Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated and the weights of each product node are updated through a graph diffusion algorithm. The matching degree between different product nodes and the joint representation is calculated by combining an attention mechanism to obtain a personalized weight distribution. Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to the key product nodes are extracted and multiple recommendation paths are determined by combining a path enumeration algorithm. Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. The comprehensive path value is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. A recommended product list is determined based on the dialogue action sequence. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

[0006] In one alternative implementation, Acquiring customer body posture data and historical behavior trajectories, extracting features from the body posture data to determine body shape characteristics, and extracting behavioral pattern features corresponding to the historical behavior trajectories and constructing a time dependency matrix include: Customer body shape data and historical behavior trajectory are obtained from the customer information collection terminal. Principal component analysis is performed on the body shape data to reduce dimensionality and obtain body shape principal component vectors. Combined with a preset standard body shape library, the body shape classification label of the current customer is determined. Based on the body shape classification label and the body shape principal component vector, body shape features are determined. One-hot encoding is performed on different behavior types in the historical behavior trajectory to obtain a behavior encoding matrix. Combined with the corresponding timestamp information, the behavior encoding matrix is ​​time-weighted to obtain a weighted behavior feature matrix. The weighted behavioral feature matrix is ​​segmented into multiple time segments according to a fixed time window. Statistical analysis is performed on the behavioral features in each time segment to obtain a behavioral frequency vector and a behavioral transition probability vector. The behavioral frequency vector and the behavioral transition probability vector are merged by linear weighted fusion to obtain the behavioral pattern feature vector of each time segment. The Pearson correlation coefficient is calculated based on the behavioral pattern feature vectors of adjacent time segments to obtain the association strength value of each time step. The association strength values ​​between all time steps are arranged in chronological order to construct the time dependency matrix.

[0007] In one alternative implementation, Based on the body shape characteristics and the time dependency matrix, a joint representation is determined and its suitability is scored with clothing items in a pre-set product library to obtain a candidate product set, including: The body shape features are converted into a multi-scale body shape tensor representation. A graph decomposition operation is performed on the time dependency matrix to obtain the temporal pattern eigenvectors and eigenvalue sequences. A high-order feature interaction is performed on the multi-scale body shape tensor and the temporal pattern eigenvectors through a tensor shrinkage operation to obtain a multi-dimensional interaction tensor. Based on the eigenvalue sequences, dynamic weights are assigned to each dimension of the multi-dimensional interaction tensor and combined with tensor decomposition dimensionality reduction to obtain a joint representation. Multidimensional attribute information of each clothing product is extracted from a pre-defined product database and heterogeneous attribute subgraphs corresponding to each dimension are constructed. The heterogeneous attribute subgraphs are then merged into a unified product semantic graph through a cross-modal alignment mechanism. Based on the joint representation, a subgraph matching algorithm is executed in the product semantic graph to obtain a set of candidate matching subgraphs. For each candidate matching subgraph, graph isomorphism and node feature similarity are calculated to obtain a structural matching score and a semantic matching score. The structural matching score and the semantic matching score are nonlinearly fused to obtain a comprehensive suitability score. An adaptive threshold segmentation is performed on the overall suitability score of all clothing products, and high-score regions are identified using a density clustering algorithm. Product nodes corresponding to the high-score regions are extracted, and the extracted product nodes and their corresponding adjacency relationships are used to form a candidate product set.

[0008] In one alternative implementation, Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated using a graph diffusion algorithm, and the weights of each product node are updated. A personalized weight distribution is obtained by calculating the matching degree between different product nodes and the joint representation using an attention mechanism. The attribute features of each product are extracted from the candidate product set and decomposed into appearance attribute vectors and functional attribute vectors in multiple dimensions. The appearance similarity between products is calculated based on the appearance attribute vectors, and the functional similarity between products is calculated based on the functional attribute vectors. Using products as graph nodes, edge connections are constructed based on the appearance similarity and the functional similarity. The appearance similarity and the functional similarity are nonlinearly combined to obtain a comprehensive similarity, which is used as the edge weight to construct a product relationship graph. The joint representation is injected as an initial preference signal into each node of the product relationship graph. Preference information is transmitted to adjacent nodes along the edge connection. In each diffusion process, the preference information is attenuated and modulated according to the edge weight and accumulated to the target node. After multiple diffusion iterations, the accumulated preference value of each product node is extracted as the updated node weight. The joint representation is decomposed into a query vector, and the attribute features of each product node are decomposed into key vectors and value vectors. The semantic relevance between the joint representation and each product node is calculated through multi-head attention. The updated node weights are combined to obtain a weighted fusion score. The comprehensive matching score of all products is normalized to obtain a personalized weight distribution.

[0009] In one alternative implementation, Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to these key product nodes are extracted, and multiple recommendation paths are determined using a path enumeration algorithm, including: The personalized weight distribution is subjected to probability density analysis to obtain the weight distribution function, and local maxima are identified by peak detection. The product nodes corresponding to the local maxima are extracted as candidate key nodes. The weight gradient and neighborhood density of each candidate key node are calculated and jointly evaluated to obtain key product nodes. Using each key product node as the center node, traverse all adjacent nodes and connecting edges within a preset number of hops to construct a local neighborhood subgraph containing the center node and its corresponding adjacency relationship. Recalibrate the edge weights in the neighborhood subgraph to obtain a local weight matrix. Within each neighborhood subgraph, a depth-first traversal operation is performed based on the local weight matrix. Starting from the center node, a path search is performed along the direction of decreasing edge weights. The sequence of nodes traversed during the traversal and the corresponding cumulative weight values ​​are recorded. When the path length reaches a preset upper limit, the current path search is terminated to obtain an initial path set. Valid paths are determined by threshold filtering based on the corresponding cumulative weight values. The valid paths are sorted according to their cumulative weight values ​​to obtain a candidate recommended path set. Path diversity and path quality indicators are extracted from the candidate recommended path set of each neighborhood subgraph, and multiple recommended paths are obtained by screening the candidate recommended paths through multi-objective optimization criteria.

[0010] In one alternative implementation, Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. Using this comprehensive path value as the optimization objective, a genetic algorithm is executed to search for the optimal recommendation strategy and generate a dialogue action sequence, including: The recommended path is decomposed into a node sequence to obtain a path state transition matrix. An initial state distribution and a transition probability distribution are set based on the path state transition matrix. The sampling start node is determined according to the initial state distribution, and multiple sampling paths are generated by step-by-step state jumps according to the transition probability distribution. The access frequency and transition probability of each sampling path are counted to obtain the path stability index. Extract the path length and edge weight information corresponding to the recommended path, calculate the revenue and cost of each recommended path in combination with the path stability index, and solve for the comprehensive path value. Use the comprehensive path value as the fitness value to encode the recommended path set to obtain a chromosome population. Filter the chromosome population according to the fitness value, and iteratively evolve the filtered chromosome population by combining crossover and mutation operations. Repeat the iteration until the preset maximum number of iterations is reached. Extract the chromosome with the highest fitness value after the last iteration and decode it to obtain the optimal recommendation strategy. Based on the path node order and node attribute information in the optimal recommendation strategy, each node is converted into a corresponding recommended action instruction, and the recommended action instructions are concatenated in the path order to obtain a dialogue action sequence.

[0011] In one alternative implementation, Based on the dialogue action sequence, a recommended product list is determined; based on the dialogue action sequence, a script generation action is determined; and a personalized shopping guide script is determined using a conditional text generation algorithm. The recommended product list and the personalized shopping guide script are then integrated to obtain the shopping guide response content, which is output as follows: The product identifier and recommendation weight of each action node in the dialogue action sequence are analyzed. The basic information of the product is retrieved from the preset product database through the product identifier and sorted by importance in combination with the recommendation weight. The core recommended products are determined according to the sorting result and the corresponding detailed attribute information is organized according to the recommendation weight order to obtain a list of recommended products. Extract the dialogue action sequence of each action node’s dialogue tag and personalized parameters, match the dialogue tag with the preset dialogue generation rule, input the personalized parameters as conditional variables into the dialogue generation rule for variable substitution and semantic expansion to obtain the exclusive introduction text and guiding statement for each recommended product, and concatenate the introduction text and the guiding statement to obtain personalized shopping guide dialogue. Based on the order of the products in the recommended product list, the detailed information of each recommended product is sequentially embedded into the preset placeholder position corresponding to the description text in the personalized shopping guide, generating the shopping guide response content and transmitting it to the user interface through the output interface.

[0012] A second aspect of this invention provides an intelligent shopping guide robot interactive decision-making system for apparel sales scenarios, comprising: The feature fusion module is used to acquire customers' body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform a suitability score with clothing products in a preset product library to obtain a candidate product set. The path generation module is used to construct a product relationship graph based on the attribute features of products in the candidate product set, propagate customer preference information and update the weights of each product node through a graph diffusion algorithm, calculate the matching degree between different product nodes and the joint representation to obtain a personalized weight distribution by combining an attention mechanism, determine key product nodes based on the personalized weight distribution, extract the neighborhood subgraphs corresponding to the key product nodes and determine multiple recommendation paths by combining a path enumeration algorithm. The decision output module is used to perform Markov chain Monte Carlo sampling on the recommended path and solve for the comprehensive value of the path. The comprehensive value of the path is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. Based on the dialogue action sequence, a list of recommended products is determined. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

[0013] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] This invention integrates real-time customer body posture data with historical behavioral trajectories to construct a joint representation encompassing body shape characteristics and time-dependent behavioral patterns. This achieves unified modeling of customers' static body characteristics and dynamic behavioral preferences, significantly improving the accuracy and comprehensiveness of product suitability scoring and ensuring that the candidate product set better matches customers' individual needs. Utilizing product relationship graphs and graph diffusion algorithms, customer preference information is effectively propagated within the product association network. Combined with an attention mechanism to calculate matching degree, the invention accurately identifies key products most relevant to the customer's current state from the candidate set, overcoming the limitations of isolated product evaluation and enhancing the relevance and interpretability of recommendation results. Through path enumeration and Markov chain Monte Carlo sampling, multiple recommendation paths are comprehensively evaluated for value. A genetic algorithm is used for global optimization search, efficiently determining the optimal recommendation strategy from a massive number of possible interaction sequences. This not only considers the immediate effect of a single recommendation but also takes into account long-term interaction goals such as guiding customer exploration and promoting cross-selling, making the shopping process more strategic and guiding. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the interactive decision-making method of the intelligent shopping guide robot in a clothing sales scenario according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the personalized product recommendation path generation process of the intelligent shopping guide robot interactive decision-making method in the clothing sales scenario according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the interactive decision-making method of the intelligent shopping guide robot in a clothing sales scenario according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire customer body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform suitability scoring on clothing products in a preset product library to obtain a candidate product set. Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated and the weights of each product node are updated through a graph diffusion algorithm. The matching degree between different product nodes and the joint representation is calculated by combining an attention mechanism to obtain a personalized weight distribution. Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to the key product nodes are extracted and multiple recommendation paths are determined by combining a path enumeration algorithm. Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. The comprehensive path value is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. A recommended product list is determined based on the dialogue action sequence. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

[0020] In one alternative implementation, Acquiring customer body posture data and historical behavior trajectories, extracting features from the body posture data to determine body shape characteristics, and extracting behavioral pattern features corresponding to the historical behavior trajectories and constructing a time dependency matrix include: Customer body shape data and historical behavior trajectory are obtained from the customer information collection terminal. Principal component analysis is performed on the body shape data to reduce dimensionality and obtain body shape principal component vectors. Combined with a preset standard body shape library, the body shape classification label of the current customer is determined. Based on the body shape classification label and the body shape principal component vector, body shape features are determined. One-hot encoding is performed on different behavior types in the historical behavior trajectory to obtain a behavior encoding matrix. Combined with the corresponding timestamp information, the behavior encoding matrix is ​​time-weighted to obtain a weighted behavior feature matrix. The weighted behavioral feature matrix is ​​segmented into multiple time segments according to a fixed time window. Statistical analysis is performed on the behavioral features in each time segment to obtain a behavioral frequency vector and a behavioral transition probability vector. The behavioral frequency vector and the behavioral transition probability vector are merged by linear weighted fusion to obtain the behavioral pattern feature vector of each time segment. The Pearson correlation coefficient is calculated based on the behavioral pattern feature vectors of adjacent time segments to obtain the association strength value of each time step. The association strength values ​​between all time steps are arranged in chronological order to construct the time dependency matrix.

[0021] Customer body shape data and historical behavior data are collected from the customer information collection terminal. Body shape data includes multi-dimensional quantitative indicators such as height, weight, shoulder width, chest circumference, waist circumference, and hip circumference. Historical behavior data records the customer's activity trajectory in the clothing sales scenario, such as browsing products, trying on clothes, consulting with sales assistants, and other behaviors and their timestamp information.

[0022] Principal component analysis (PCA) is performed to reduce the dimensionality of the acquired body posture data. An original data matrix is ​​constructed based on the pre-acquired body posture data. This original data matrix, containing multiple body posture features, is standardized. The covariance matrix between features is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The top principal components with a cumulative contribution rate exceeding 85% are selected in descending order of eigenvalues ​​to form the body shape principal component vector. For example, if the original body posture data contains 15 feature dimensions, PCA reduces this to 4 principal components, resulting in a 4-dimensional body shape principal component vector. The first principal component reflects overall body size, the second principal component reflects the upper and lower body proportions, the third principal component reflects the front-to-back body difference, and the fourth principal component reflects shoulder features.

[0023] The system uses a pre-defined standard body shape database to determine the current customer's body shape classification label. This database includes various typical body shapes, such as pear-shaped, apple-shaped, hourglass-shaped, and rectangular. The system calculates the Euclidean distance between the customer's principal component vector and each typical body shape in the database, selecting the body shape with the smallest distance as the customer's body shape classification label. For example, if a customer's principal component vector has an Euclidean distance of 0.23 with an hourglass-shaped body shape, which is less than the distances to other body shapes, then the customer's body shape classification label is determined to be hourglass-shaped.

[0024] Body type features are determined based on body type classification labels and body type principal component vectors. The body type classification labels are one-hot encoded to form a category feature vector, which is then concatenated with the body type principal component vector to form a complete body type feature representation. If the body type database contains 6 typical body types and the body type principal component vector is 4-dimensional, then the complete body type feature is a 10-dimensional vector.

[0025] One-hot encoding is used to obtain a behavior encoding matrix for different behavior types in historical behavior trajectories. Customer behaviors are categorized into various types, such as browsing women's clothing, browsing men's clothing, trying on outerwear, inquiring about prices, and taking photos to share. A unique encoding bit is assigned to each behavior type, and each behavior event forms a one-hot encoding vector. Assuming there are 12 behavior types, each behavior event corresponds to a 12-dimensional one-hot encoding vector, and all behavior events form the behavior encoding matrix, where the number of rows in the matrix equals the total number of behavior events.

[0026] The behavior encoding matrix is ​​time-weighted by incorporating timestamp information. A time decay function is used to calculate the time weight of each behavior event, assigning higher weights to more recent behaviors and lower weights to earlier behaviors. The time weights are calculated using an exponential decay function with a decay coefficient of 0.05. When the time difference between the current time and the behavior occurrence time is 30 days, the weight is approximately 0.22. Multiplying the time weights by the behavior encoding vector yields the weighted behavior feature matrix.

[0027] The weighted behavioral feature matrix is ​​segmented according to a fixed time window. The time window size is set to 7 days, dividing the entire behavioral history into multiple consecutive time segments. For example, if a customer's historical behavioral records span 42 days, it is divided into 6 time segments.

[0028] Statistical analysis of behavioral characteristics within each time segment yields a behavior frequency vector and a behavior transition probability vector. The behavior frequency vector records the frequency of occurrence of each type of behavior within that time segment, with a sum of 1. The behavior transition probability vector records the transition probability between different behaviors; for example, the probability of moving from browsing the women's clothing section to trying on a dress is 0.35, and the probability of moving from trying on a dress to inquiring about the price is 0.68. The behavior frequency vector is obtained by counting the number of occurrences of each type of behavior and dividing by the total number of behaviors within that time segment. A behavior transition matrix is ​​constructed, counting the number of transitions between behavior types, normalizing by row to obtain conditional probabilities, and then flattening the matrix to form the behavior transition probability vector.

[0029] The behavior frequency vector and behavior transition probability vector are merged using linear weighted fusion to obtain the behavior pattern feature vector for each time segment. The weight of the behavior frequency vector is set to 0.4, and the weight of the behavior transition probability vector is set to 0.6. The two are then weighted and added together to obtain the behavior pattern feature vector. If the behavior frequency vector is 12-dimensional and the behavior transition matrix is ​​144-dimensional after flattening, then the behavior pattern feature vector is 156-dimensional.

[0030] The Pearson correlation coefficient is calculated based on the behavioral pattern feature vectors of adjacent time segments to obtain the time-step association strength value. For any two time segments, the corresponding Pearson correlation coefficient is calculated, with a value ranging from -1 to 1, representing the similarity of the behavioral patterns of the two time segments. The closer the correlation coefficient is to 1, the more similar the behavioral patterns of the two time segments are; the closer the correlation coefficient is to -1, the more opposite the behavioral patterns are; and the closer the correlation coefficient is to 0, the more unrelated the behavioral patterns are.

[0031] Construct a time dependency matrix by arranging the association strength values ​​between all time steps in chronological order. For historical data containing six time segments, construct a 6×6 time dependency matrix, where each element represents the Pearson correlation coefficient between two corresponding time segments. All elements on the diagonal of the dependency matrix are 1, indicating that each time segment is perfectly correlated with itself. For example, the association strength between the first and second time segments is 0.73, indicating that the behavioral patterns in these two periods are highly similar; the association strength between the first and sixth time segments is 0.21, indicating that the customer's behavioral patterns have changed significantly over time.

[0032] In this embodiment, principal component analysis (PCA) is used to reduce the dimensionality of customer body shape data. High-dimensional, redundant, and highly correlated body shape features are compressed into representative body shape principal component vectors. Combined with a standard body shape library, body shape classification is performed, achieving an accurate characterization of customer body shape structure. This effectively reduces noise interference and the impact of redundant information, improving the stability and consistency of body shape recognition. By using one-hot encoding of historical behavior trajectories and introducing timestamps for weighted processing, the behavioral features not only reflect the distribution of behavior types but also the degree of time influence of behavior occurrence. This overcomes the problem of only counting behavior frequency while ignoring time decay factors, improving the ability to characterize the dynamic changes in customer behavior. By segmenting the data into fixed time windows and extracting behavior frequency vectors and behavior transition probability vectors separately, and then performing linear fusion, the behavioral features simultaneously contain static distribution information and dynamic transition relationship information. This compensates for the insufficiency of single statistical methods in reflecting the correlation of behavioral structure, improving the completeness and discriminative ability of behavioral pattern expression.

[0033] In one alternative implementation, Based on the body shape characteristics and the time dependency matrix, a joint representation is determined and its suitability is scored with clothing items in a pre-set product library to obtain a candidate product set, including: The body shape features are converted into a multi-scale body shape tensor representation. A graph decomposition operation is performed on the time dependency matrix to obtain the temporal pattern eigenvectors and eigenvalue sequences. A high-order feature interaction is performed on the multi-scale body shape tensor and the temporal pattern eigenvectors through a tensor shrinkage operation to obtain a multi-dimensional interaction tensor. Based on the eigenvalue sequences, dynamic weights are assigned to each dimension of the multi-dimensional interaction tensor and combined with tensor decomposition dimensionality reduction to obtain a joint representation. Multidimensional attribute information of each clothing product is extracted from a pre-defined product database and heterogeneous attribute subgraphs corresponding to each dimension are constructed. The heterogeneous attribute subgraphs are then merged into a unified product semantic graph through a cross-modal alignment mechanism. Based on the joint representation, a subgraph matching algorithm is executed in the product semantic graph to obtain a set of candidate matching subgraphs. For each candidate matching subgraph, graph isomorphism and node feature similarity are calculated to obtain a structural matching score and a semantic matching score. The structural matching score and the semantic matching score are nonlinearly fused to obtain a comprehensive suitability score. An adaptive threshold segmentation is performed on the overall suitability score of all clothing products, and high-score regions are identified using a density clustering algorithm. Product nodes corresponding to the high-score regions are extracted, and the extracted product nodes and their corresponding adjacency relationships are used to form a candidate product set.

[0034] The process of converting body shape features into a multi-scale body shape tensor representation begins with the reorganization of the original body shape features. In the previous example, the original body shape features were 10-dimensional vectors containing a one-hot encoded part of the body shape classification label and a body shape principal component vector part. Feature expansion methods are used to extend the original features to multiple scale levels. Specifically, the 10-dimensional feature vector is reshaped into a 2×5 second-order tensor, and then an autocorrelation operation is performed to obtain a 2×2 local feature tensor and a 5×5 global feature tensor. The tensors at different scales are merged into a single multi-scale body shape tensor, with a final scale of 2×5×7, where 7 is the merged channel dimension.

[0035] A spectral decomposition operation is performed on the time dependency matrix, which is a 6×6 real symmetric matrix representing the association strength between different time segments. The time dependency matrix is ​​treated as the adjacency matrix of a weighted undirected graph, and spectral decomposition is performed to obtain eigenvectors and eigenvalues. The eigenvectors represent the basic patterns of the time series, and the eigenvalues ​​represent the importance of each basic pattern. Spectral decomposition of the 6×6 time dependency matrix yields six eigenvectors and their corresponding eigenvalues. Each eigenvector has a dimension of 6 and reflects a basic feature of the time series pattern. For example, the eigenvalues, ordered from largest to smallest, are 0.92, 0.78, 0.65, 0.41, 0.24, and 0.13, indicating that the first three eigenvectors contain most of the temporal structure information.

[0036] By performing high-order feature interaction on the multi-scale body shape tensor and the temporal pattern feature vector through tensor contraction operation, a multi-dimensional interaction tensor reflecting the interaction relationship between body shape features and temporal behavior patterns is generated. Tensor contraction is then performed on the 2×5×7 multi-scale body shape tensor and the first 3 important feature vectors in the temporal pattern feature vector to obtain a 2×5×3 multi-dimensional interaction tensor.

[0037] Dynamic weights are assigned to each dimension of the multidimensional interaction tensor based on the eigenvalue sequence. The eigenvalue sequence is 0.92, 0.78, and 0.65, representing the importance of each temporal pattern basis. Correspondingly, the weights for the third dimension are 0.39, 0.33, and 0.28. These weights are applied to the corresponding dimensions of the interaction tensor to highlight the contributions of important patterns. Combined with tensor decomposition dimensionality reduction, the weighted 2×5×3 interaction tensor is decomposed into a low-rank approximation with rank 4, resulting in a 12-dimensional joint representation vector.

[0038] The system extracts multi-dimensional attribute information for each clothing item from a pre-defined product database and constructs a heterogeneous attribute subgraph. The product database contains 3000 clothing items, each with attributes such as category, style, material, function, occasion, and season. For each attribute dimension, an attribute subgraph is constructed, where nodes represent products and attribute values, and edges represent a product possessing a specific attribute value. For example, the category attribute subgraph contains category nodes such as outerwear, trousers, and skirts; the style attribute subgraph contains style nodes such as casual, formal, and sporty. Each product node is connected to its associated attribute value node, forming a bipartite graph structure.

[0039] Heterogeneous attribute subgraphs are fused into a unified product semantic graph through a cross-modal alignment mechanism. Low-dimensional embedding representations are learned for nodes in each attribute subgraph, and a graph neural network with shared parameters is used to process different subgraphs, ensuring that the representations of the same product in different subgraphs are close to each other. During the fusion process, a weight matrix is ​​set to adjust the importance of different attributes: category and style attributes have higher weights (0.25 and 0.20, respectively); material and function have moderate weights (0.15 and 0.15, respectively); and occasion and season have lower weights (0.15 and 0.10, respectively). The fused product semantic graph contains 3000 product nodes and approximately 350 attribute value nodes, with about 18000 edges, forming a densely connected heterogeneous graph structure.

[0040] A subgraph matching algorithm is performed in the product semantic graph based on joint representations to obtain a set of candidate matching subgraphs. A 12-dimensional joint representation vector is mapped to the feature space of the product semantic graph as the query vector. A random walk algorithm is used to sample subgraphs from the product semantic graph, with the initial point selected as the node most similar to the query vector. In each random walk step, the node moves along the edge to an adjacent node with a probability of 0.85 and returns to the initial node with a probability of 0.15. The walk step is set to 10 steps, and this process is repeated 200 times to generate 200 candidate subgraphs. Each subgraph contains approximately 8 to 15 nodes, covering various products and attribute values.

[0041] For each candidate matching subgraph, graph isomorphism and node feature similarity are calculated to obtain a structural matching score and a semantic matching score. Graph isomorphism is measured by graph edit distance, which calculates the minimum number of operations required to transform a candidate subgraph into the target structure. Operations include adding / deleting nodes, adding / deleting edges, and changing node labels. Node feature similarity is obtained by calculating the cosine similarity between the product node features and the joint representation vector in the subgraph. For example, a graph isomorphism of 0.83 for a subgraph indicates a high degree of structural matching, and a node feature similarity of 0.76 indicates a high degree of semantic relevance.

[0042] A comprehensive suitability score is obtained by nonlinearly fusing the structural matching score and the semantic matching score. A weighted harmonic mean method is used for fusion, with a structural weight of 0.65 and a semantic weight of 0.35. The weighted harmonic mean is then calculated, and the result is mapped to a range of 0 to 1 using the sigmoid function as the comprehensive suitability score. For example, a structural score of 0.83 and a semantic score of 0.76 result in a comprehensive suitability score of 0.81. Comprehensive suitability scores were calculated for 3000 products, with scores ranging from 0.12 to 0.92.

[0043] An adaptive threshold segmentation was performed on the overall suitability scores of all clothing items. The Otsu method was used to automatically determine the optimal threshold, dividing the items into high-scoring and low-scoring groups. For example, a threshold of 0.68 was determined, with approximately 420 items scoring above this threshold. High-score regions were identified using the DBSCAN density clustering algorithm, with a distance threshold of 0.1 and a minimum sample size of 5, resulting in the identification of 8 high-density regions.

[0044] Product nodes and their adjacency relationships corresponding to high-scoring regions are extracted to form a candidate product set. Product nodes are extracted from 8 high-density regions, containing approximately 85 products. The products are organized according to the adjacency relationships in the original product semantic graph to form a coherent product recommendation network. For example, a certain high-scoring region contains 10 casual style outerwear products. These products are connected to each other in the original graph and are also connected to specific style attribute nodes such as "casual" and "simple".

[0045] In this embodiment, by converting body shape features into a multi-scale body shape tensor representation and combining it with a time-dependent relation matrix for graph decomposition, customer body shape structure information and behavioral temporal evolution features are coupled and modeled in a unified high-order space. This improves the sensitivity of the joint representation to individual differences and behavioral changes, enhances the discriminative power and stability of feature expression, and overcomes the problem that feature vectorization representation is difficult to characterize complex attribute relationships and semantic associations by constructing heterogeneous attribute subgraphs for multi-dimensional product attributes and using a cross-modal alignment mechanism to fuse them into a unified product semantic graph. This allows for the explicit expression of the association structure between product attributes at the structural level, improving the completeness and scalability of product knowledge expression and providing a higher-quality semantic foundation for subsequent accurate matching. By performing subgraph matching in the product semantic graph and calculating graph isomorphism and node feature similarity respectively, followed by nonlinear fusion, the matching results simultaneously consider structural consistency and semantic similarity, effectively reducing the probability of mismatch, improving the accuracy and robustness of matching results, and enhancing the adaptability to complex combinations of product attributes.

[0046] In one alternative implementation, Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated using a graph diffusion algorithm, and the weights of each product node are updated. A personalized weight distribution is obtained by calculating the matching degree between different product nodes and the joint representation using an attention mechanism. The attribute features of each product are extracted from the candidate product set and decomposed into appearance attribute vectors and functional attribute vectors in multiple dimensions. The appearance similarity between products is calculated based on the appearance attribute vectors, and the functional similarity between products is calculated based on the functional attribute vectors. Using products as graph nodes, edge connections are constructed based on the appearance similarity and the functional similarity. The appearance similarity and the functional similarity are nonlinearly combined to obtain a comprehensive similarity, which is used as the edge weight to construct a product relationship graph. The joint representation is injected as an initial preference signal into each node of the product relationship graph. Preference information is transmitted to adjacent nodes along the edge connection. In each diffusion process, the preference information is attenuated and modulated according to the edge weight and accumulated to the target node. After multiple diffusion iterations, the accumulated preference value of each product node is extracted as the updated node weight. The joint representation is decomposed into a query vector, and the attribute features of each product node are decomposed into key vectors and value vectors. The semantic relevance between the joint representation and each product node is calculated through multi-head attention. The updated node weights are combined to obtain a weighted fusion score. The comprehensive matching score of all products is normalized to obtain a personalized weight distribution.

[0047] The attribute features of each product in the candidate product set are extracted and decomposed into multiple dimensions to construct an accurate product relationship graph. The candidate product set contains approximately 85 products, each with approximately 24 attribute features, covering information such as color, texture, material, cut, style, functionality, and occasion suitability. Principal component analysis and non-negative matrix factorization techniques are applied to the aforementioned 24 attribute features for dimensionality reduction and decomposition, decomposing the original features into two subspaces: a 14-dimensional appearance attribute vector and a 10-dimensional functional attribute vector. The appearance attribute vector focuses on expressing the visual characteristics of the product. For example, in the appearance attribute vector of a beige cotton-linen suit jacket, the color-related dimension value is 0.82, the texture-related dimension value is 0.75, and the cut-related dimension value is 0.68. The functional attribute vector focuses on expressing the practical characteristics of the product. For example, in the functional attribute vector of this suit jacket, the formal occasion suitability dimension value is 0.92, the warmth dimension value is 0.45, and the breathability dimension value is 0.78.

[0048] The appearance similarity between products is calculated based on appearance attribute vectors. Cosine similarity is used to measure the cosine of the angle between the appearance attribute vectors of two products. This value ranges from -1 to 1, with a higher value indicating greater similarity. For example, two light-colored cotton-linen tops have an appearance similarity of 0.93, while a dark suit and a light-colored casual trousers have an appearance similarity of only 0.28. Similarly, functional similarity between products is calculated based on functional attribute vectors. Two shirts made of the same business-appropriate, wrinkle-resistant fabric have a functional similarity of 0.88, while a sports T-shirt and a formal dress trousers have a functional similarity of only 0.15.

[0049] Using products as graph nodes, edges are constructed based on appearance and functional similarity. When the appearance or functional similarity of two products exceeds a threshold of 0.7, an edge is established between the corresponding product nodes. A comprehensive similarity is obtained by non-linearly combining appearance and functional similarity, which serves as the edge weight. The non-linear combination uses a weighted geometric mean method, with an appearance similarity weight of 0.65 and a functional similarity weight of 0.35. For example, if two products have an appearance similarity of 0.85 and a functional similarity of 0.75, the comprehensive similarity after non-linear combination is 0.81. The final constructed product relationship graph contains 85 nodes and approximately 298 edges, with each node connecting an average of 7 neighboring nodes.

[0050] The joint representation is injected as the initial preference signal into each node of the product relationship graph. The joint representation is a 12-dimensional vector containing user body shape features and behavioral pattern information. A mapping function is used to calculate the similarity between the joint representation and the attribute features of each product node, resulting in an initial preference score. For example, a user's joint representation has a similarity of 0.76 with a beige suit jacket and 0.45 with a sweatshirt. The initial preference score is used as the initial weight for each node, ranging from 0 to 1.

[0051] Preference information is propagated to neighboring nodes along the edges, realizing a graph diffusion process of preferences. In each diffusion process, a node distributes its own weight to neighboring nodes according to the weights of its outgoing edges. Simultaneously, a decay coefficient of 0.85 is applied to modulate the propagation process, preventing distant nodes from receiving excessively high weights. For example, a node with an initial weight of 0.76 (e.g., a suit jacket) propagates a weight of 0.76 × 0.81 × 0.85 = 0.52 to a neighboring node with a similarity of 0.81. After each round of diffusion, each node accumulates the received weights and updates its own node weight. The diffusion process iterates for 5 rounds and then stabilizes, at which point the node weights reflect the preference scores considering the relational network structure. For example, a node with an initial weight of 0.45 (e.g., a sports sweatshirt) might increase its weight to 0.68 after diffusion iterations.

[0052] The joint representation is decomposed into query vectors. A multilayer perceptron is used to decompose the 12-dimensional joint representation into four 3-dimensional query vectors, each focusing on a different semantic aspect. The 24-dimensional attribute features of each product node are decomposed into key vectors and value vectors. Specifically, the attribute features are projected into four 3-dimensional key vectors and four 3-dimensional value vectors through two independent multilayer perceptrons.

[0053] The semantic relevance between the joint representation and each product node is calculated using multi-head attention. The multi-head attention mechanism comprises four attention heads. Each head independently calculates the dot product of the query vector and the key vector, then performs scaling and softmax normalization to obtain attention weights. These attention weights are then weighted and summed with the value vectors to obtain the output of that head. The outputs of the four heads are concatenated and subjected to a linear transformation to obtain the final semantic relevance. For example, the semantic relevance between the joint representation and a beige suit jacket is 0.82, indicating a high degree of semantic matching.

[0054] The updated node weights are combined and weighted to obtain the overall matching score. The semantic relevance weight is 0.6, and the node weight is 0.4, using a weighted average method for fusion. For example, a product with a semantic relevance of 0.82 and a node weight of 0.71 will have an overall matching score of 0.82 × 0.6 + 0.71 × 0.4 = 0.77. The overall matching scores of all products are then normalized to a sum of 1, resulting in a personalized weight distribution. Normalization uses the softmax function, adjusting the temperature coefficient to 0.8 to enhance the weight difference of high-scoring products. The personalized weight distribution of the 85 products ranges from 0.0003 to 0.0512, with a total weight sum of 1.

[0055] In this embodiment, by performing multi-dimensional decomposition of the attribute features of candidate products, appearance attributes and functional attributes are modeled and their similarity is calculated separately. This expands the association between products from single feature similarity to multi-dimensional semantic structure association, which can more realistically reflect the strength of the composite association between products, thereby improving the granularity and expressive power of product relationship modeling. By injecting the joint representation as the initial preference signal into the product relationship graph and performing multiple rounds of weighted diffusion along the edge connection, user preferences are no longer limited to directly matched product nodes, but can propagate to structurally related adjacent nodes along high similarity paths. This can fully utilize the potential associations in the graph structure, enhance the ability to mine implicit interests, and improve the coverage and coherence of recommendation results. By performing vector decomposition on the joint representation and product node features, and introducing a multi-head attention mechanism to calculate semantic association, the matching process can capture semantic alignment relationships of different dimensions in multiple subspaces, which can improve the ability to model complex semantic relationships and enhance the recognition accuracy of fine-grained preference differences.

[0056] In one alternative implementation, Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to these key product nodes are extracted, and multiple recommendation paths are determined using a path enumeration algorithm, including: The personalized weight distribution is subjected to probability density analysis to obtain the weight distribution function, and local maxima are identified by peak detection. The product nodes corresponding to the local maxima are extracted as candidate key nodes. The weight gradient and neighborhood density of each candidate key node are calculated and jointly evaluated to obtain key product nodes. Using each key product node as the center node, traverse all adjacent nodes and connecting edges within a preset number of hops to construct a local neighborhood subgraph containing the center node and its corresponding adjacency relationship. Recalibrate the edge weights in the neighborhood subgraph to obtain a local weight matrix. Within each neighborhood subgraph, a depth-first traversal operation is performed based on the local weight matrix. Starting from the center node, a path search is performed along the direction of decreasing edge weights. The sequence of nodes traversed during the traversal and the corresponding cumulative weight values ​​are recorded. When the path length reaches a preset upper limit, the current path search is terminated to obtain an initial path set. Valid paths are determined by threshold filtering based on the corresponding cumulative weight values. The valid paths are sorted according to their cumulative weight values ​​to obtain a candidate recommended path set. Path diversity and path quality indicators are extracted from the candidate recommended path set of each neighborhood subgraph, and multiple recommended paths are obtained by screening the candidate recommended paths through multi-objective optimization criteria.

[0057] Probability density analysis was performed on the personalized weight distribution to obtain the weight distribution function, and local maxima were identified using peak detection. The personalized weight distribution includes weight values ​​for 85 items, ranging from 0.0003 to 0.0512. A kernel density estimation method was used to construct the weight distribution function, selecting a Gaussian kernel function with a bandwidth parameter set to 0.005. The constructed weight distribution function exhibits multimodal characteristics, reflecting the weight clustering of different types of items. Local maxima were identified using a one-dimensional peak detection algorithm, with a minimum peak height of 0.2 and a minimum peak spacing of 0.015. Four local maxima were detected, corresponding to weight values ​​of 0.0512, 0.0458, 0.0362, and 0.0285, respectively. The item nodes corresponding to these four local maxima were extracted as candidate key nodes: beige suit jacket, white business shirt, dark gray trousers, and light blue shirt.

[0058] The weight gradient and neighborhood density of each candidate key node are calculated, and a joint evaluation is performed to screen and obtain key product nodes. The weight gradient is obtained by calculating the average weight difference between a candidate node and its directly connected nodes, reflecting the steepness of the node in the weight space. For example, the weight of the beige suit jacket is 0.0512, the average weight of its 7 directly connected nodes is 0.0386, and the weight gradient is 0.0126. The neighborhood density is obtained by calculating the product of the connectivity of the candidate node in the graph and the weights of its neighboring nodes, reflecting the degree of weight clustering in the area surrounding the node. The connectivity of the beige suit jacket is 7, the sum of the weights of its neighboring nodes is 0.2702, and the neighborhood density is 1.8914. The joint evaluation uses a weighted sum of the weight gradient and the neighborhood density, with a weight gradient weight of 0.6 and a neighborhood density weight of 0.4. The calculated results are: beige suit jacket 0.7634, white business shirt 0.7328, dark gray trousers 0.6976, and light blue shirt 0.6512. With a threshold set at 0.7, the final selection yielded three key product categories: beige suit jackets, white business shirts, and dark gray trousers.

[0059] Using each key product node as the center node, all adjacent nodes and connecting edges within a preset hop count range are traversed to construct a local neighborhood subgraph containing the center node and its corresponding adjacency relationships. The preset hop count is set to 2, meaning that all nodes reachable from the center node via a maximum of 2 edges are included in the local neighborhood subgraph. The local neighborhood subgraph centered on the beige suit jacket contains 1 center node, 7 one-hop neighbor nodes, and 18 two-hop neighbor nodes, for a total of 26 nodes and approximately 62 edges. The edge weights in the neighborhood subgraph are recalibrated to obtain a local weight matrix. The recalibration considers three factors: the original edge weights, the distance decay from the node to the center node, and the individual weights of the nodes. The distance decay coefficient is set to 0.8, meaning that for each additional hop, the edge weight decays to 0.8 times its original value; the node weight influence factor is set to 0.3, meaning that the individual weights of the nodes affect the edge weights by a ratio of 0.3. For example, the original edge weight between the central node "Beige Suit Jacket" and its one-hop neighbor node "Beige Casual Suit" is 0.92, and the recalibrated edge weight is 0.92×1×(1+0.3×0.0412)=0.9314.

[0060] Within each neighborhood subgraph, a depth-first traversal is performed based on the local weight matrix, starting from the center node and searching for paths along decreasing edge weights. Taking the neighborhood subgraph centered on the beige suit jacket as an example, starting from the center node, the node connected by the edge with the highest weight (beige casual suit, edge weight 0.9314) is visited first; then, the node connected by the edge with the second highest weight (light gray casual pants, edge weight 0.8968) is visited; finally, the node connected by the white simple T-shirt (edge ​​weight 0.8576) is visited. The sequence of nodes visited during the traversal and their corresponding cumulative weight values ​​are recorded. The cumulative weight is calculated by multiplication; the cumulative weight of the aforementioned path is 0.9314 × 0.8968 × 0.8576 = 0.7172. The current path search terminates when the path length reaches a preset upper limit. The preset upper limit for path length is 4, meaning each path contains a maximum of 4 nodes. Approximately 42 initial paths are generated starting from the beige suit jacket using depth-first traversal. A threshold of 0.6 is used to filter valid paths based on the cumulative weight values, retaining 15 valid paths after filtering. The effective paths were sorted in descending order of their cumulative weight values ​​to obtain the candidate recommended path set. The top 3 candidate recommended paths after sorting are: beige suit jacket - beige casual suit - light gray casual pants - white simple T-shirt, with a cumulative weight of 0.7172; beige suit jacket - white business shirt - dark gray trousers - black leather shoes, with a cumulative weight of 0.7058; and beige suit jacket - beige casual suit - khaki casual pants - blue Oxford shirt, with a cumulative weight of 0.6921.

[0061] Path diversity and path quality indices are extracted from the candidate recommendation path sets of each neighborhood subgraph. The path diversity index is obtained by calculating the entropy value of the node categories in the path; a higher entropy value indicates a more diverse range of product categories covered by the path. For example, the path "beige suit jacket - white business shirt - dark gray trousers - black leather shoes" includes four categories: tops, trousers, and shoes, with a diversity index of 1.5. The path quality index is obtained by calculating the average personalized weight of the nodes in the path, reflecting the overall recommendation strength of the path. The personalized weights of the nodes in the above path are 0.0512, 0.0498, 0.0475, and 0.0382, respectively, with an average of 0.0467, which is used as the path quality index. Multiple recommendation paths are obtained by screening the candidate recommendation paths using a multi-objective optimization criterion. The multi-objective optimization adopts the Pareto optimality principle, considering both the diversity and quality indices simultaneously, and selecting paths that are not surpassed by any other path in both indices. After screening, 5 recommendation paths with balanced diversity and quality are selected from the 15 candidate paths of the "beige suit jacket" central node. Similarly, four recommended paths were selected from the two central nodes: white business shirt and dark gray trousers.

[0062] In this embodiment, by performing probability density analysis on the personalized weight distribution and combining it with peak detection to identify local maxima, the recommendation decision no longer relies on a single global maximum value. This avoids the recommendation results being overly concentrated in a single region, improves the ability to identify multi-interest preference structures, and thus enhances the hierarchy and coverage of the recommendation results. By constructing a local neighborhood subgraph centered on key product nodes and recalibrating the edge weights, the recommendation analysis focuses on locally highly relevant structural regions, reducing the computational burden caused by the complexity of the global graph structure. While ensuring the integrity of structural associations, it improves computational efficiency and the relevance of results. In the neighborhood subgraph, a depth-first traversal is performed along the direction of decreasing edge weights, and filtering is performed by combining path length and cumulative weights. This ensures that the recommendation results not only reflect the direct correlation between nodes but also depict the potential matching relationships reflected by multi-hop associated paths, thereby improving the combination rationality and scenario adaptability of the recommendation results.

[0063] Figure 2 This is a flowchart illustrating the personalized product recommendation path generation process of the intelligent shopping guide robot interactive decision-making method in the clothing sales scenario according to an embodiment of the present invention.

[0064] In one alternative implementation, Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. Using this comprehensive path value as the optimization objective, a genetic algorithm is executed to search for the optimal recommendation strategy and generate a dialogue action sequence, including: The recommended path is decomposed into a node sequence to obtain a path state transition matrix. An initial state distribution and a transition probability distribution are set based on the path state transition matrix. The sampling start node is determined according to the initial state distribution, and multiple sampling paths are generated by step-by-step state jumps according to the transition probability distribution. The access frequency and transition probability of each sampling path are counted to obtain the path stability index. Extract the path length and edge weight information corresponding to the recommended path, calculate the revenue and cost of each recommended path in combination with the path stability index, and solve for the comprehensive path value. Use the comprehensive path value as the fitness value to encode the recommended path set to obtain a chromosome population. Filter the chromosome population according to the fitness value, and iteratively evolve the filtered chromosome population by combining crossover and mutation operations. Repeat the iteration until the preset maximum number of iterations is reached. Extract the chromosome with the highest fitness value after the last iteration and decode it to obtain the optimal recommendation strategy. Based on the path node order and node attribute information in the optimal recommendation strategy, each node is converted into a corresponding recommended action instruction, and the recommended action instructions are concatenated in the path order to obtain a dialogue action sequence.

[0065] The recommended paths are decomposed into node sequences to obtain the path state transition matrix. Based on 13 recommended paths, each containing 4 product nodes, a total of 32 different product nodes are involved. A 32×32 path state transition matrix is ​​constructed, where each element represents the number of times a node transitions to another. For example, the transition from a beige suit jacket to a white business shirt occurs 3 times in the recommended path, corresponding to a state transition matrix element value of 3; the transition from a white business shirt to dark gray trousers occurs 2 times, corresponding to an element value of 2. By normalizing the elements in each row of the state transition matrix, the transition probability from each node to other nodes is obtained. For example, starting from a beige suit jacket, the probability of transitioning to a white business shirt is 0.38, the probability of transitioning to a beige casual suit is 0.42, and the probability of transitioning to a light blue shirt is 0.20.

[0066] The initial state distribution and transition probability distribution are set based on the path state transition matrix. The initial state distribution is determined based on the frequency of each node as the starting node in the recommended path. The beige suit jacket appears 5 times as the starting node, the white business shirt appears 4 times, and the dark gray trousers appear 4 times, with normalized initial state probabilities of 0.38, 0.31, and 0.31, respectively. The transition probability distribution directly uses the normalized row vectors in the state transition matrix. The sampling starting node is determined according to the initial state distribution, and sampling is performed using a roulette wheel selection method. In one sampling, a random number between 0 and 1, 0.27, is generated. This random number is less than the cumulative probability of the beige suit jacket (0.38), so the beige suit jacket is selected as the starting node. Multiple sampling paths are generated by progressively jumping states according to the transition probability distribution. Starting from the starting node beige suit jacket, a random number, 0.52, is generated. This number is greater than the transition probability of the white business shirt (0.38) but less than the sum of the transition probabilities of the white business shirt and the beige casual suit (0.38 + 0.42 = 0.80), so the next node is the beige casual suit. Continuing from the beige casual suit, proceed to the next jump based on the corresponding transition probability distribution until the path length reaches 4 or a termination state with no further transitions is encountered. Repeat the above process to generate 200 sampled paths, and calculate the access frequency and transition probability of each sampled path to obtain the path stability index.

[0067] Extract the path length and edge weight information corresponding to the recommended paths. The path length of each recommended path is the number of nodes minus one; the length of all 13 recommended paths is 3. Extract the weights of each edge in the path. For example, in the path "beige suit jacket - white business shirt - dark gray trousers - black leather shoes", the weights of the three edges are 0.9226, 0.8943, and 0.8546, respectively. Calculate the revenue and cost values ​​of each recommended path based on the path stability index. The revenue value is determined by the weighted sum of the personalized weights of each node in the path and the path stability index, with weights of 0.6 and 0.4, respectively. The revenue value of the aforementioned path is 0.0772. The cost value is determined by the weighted difference between the path length and the mean edge weights, with weights of 0.35 and 0.65, respectively. The cost value of the aforementioned path is 0.3815. The overall path value is calculated by dividing the revenue value by the cost value; the overall value of the aforementioned path is 0.2024. The comprehensive value was calculated for each of the 13 recommended paths, and the fitness values ​​ranged from 0.1823 to 0.2253.

[0068] The recommended path set is encoded to obtain a chromosome population. Each recommended path is encoded as a chromosome, with a length twice the number of nodes in the path. The first half encodes the node numbers, and the second half encodes the connection methods between nodes. The chromosome population is screened based on fitness values ​​using a roulette wheel selection algorithm. Chromosomes with higher fitness values ​​have a greater probability of being selected. For example, a chromosome with a comprehensive value of 0.2253 has a selection probability of 0.14, while a chromosome with a comprehensive value of 0.1823 has a selection probability of only 0.05. The screened chromosome population is iteratively evolved using crossover and mutation operations. The crossover operation selects two chromosomes and swaps segments of the two chromosomes at random positions, with a crossover probability of 0.75. The mutation operation performs gene mutations at random positions on the chromosomes, with a mutation probability of 0.08.

[0069] The iteration process is repeated until the preset maximum number of iterations (50) is reached. In each iteration, the fitness value of the newly generated chromosome is evaluated and compared with the original chromosome, retaining the chromosome with higher fitness. The chromosome with the highest fitness value after the last iteration is extracted, with a fitness value of 0.2612. This chromosome is decoded to obtain the optimal recommendation strategy. The decoded result is the path "beige suit jacket - white business shirt - dark gray trousers - dark blue silk tie", where the connection mode between each node is encoded as 1, indicating a direct recommendation relationship.

[0070] Based on the path node order and node attribute information in the optimal recommendation strategy, each node is converted into a corresponding recommendation action instruction. Each product node contains attribute information such as identifier, category, recommendation weight, and verbal tag. For example, the attributes of the beige suit jacket node include: identifier 10101, category "outerwear", recommendation weight 0.85, and verbal tag "classic". The node is converted into a concise recommendation action instruction: "Recommend 10101". The white business shirt is converted into the action instruction: "Recommend 20505". The dark gray trousers are converted into the action instruction: "Recommend 31202". The dark blue silk tie is converted into the action instruction: "Recommend 40803". These simple recommendation action instructions are concatenated according to the path order to obtain the dialogue action sequence: "Recommend 10101; Recommend 20505; Recommend 31202; Recommend 40803".

[0071] In this embodiment, the recommended path is decomposed into a node sequence and a path state transition matrix is ​​constructed. Based on this, the initial state distribution and transition probability distribution are set, and multiple sampling is performed to generate paths. The access frequency and transition probability are statistically analyzed to form a path stability index. This transforms path evaluation from a one-time static calculation to a dynamic evaluation based on a probabilistic evolution process. It can quantify the stability and structural reliability of the path in multiple random evolution processes, thereby reducing the uncertainty risk brought by accidental high-weight paths and improving the robustness of recommended path selection. By encoding the recommended path into a chromosome population and performing iterative evolution through screening, crossover, and mutation based on fitness values, the search for the recommendation strategy is expanded from a local optimum to a global optimization process. This allows the strategy to escape local optima, improves the strategy space exploration capability, and enhances the global optimality and adaptability of the recommendation strategy. According to the optimal recommendation strategy, the path nodes are converted into recommended action instructions and chained together to form a dialogue action sequence. This upgrades the recommendation result from a static product list output to a structured and executable interactive process, which can enhance the logical coherence and interactive guidance capability of the recommendation, and improve user decision-making efficiency and experience consistency.

[0072] In one alternative implementation, Based on the dialogue action sequence, a recommended product list is determined; based on the dialogue action sequence, a script generation action is determined; and a personalized shopping guide script is determined using a conditional text generation algorithm. The recommended product list and the personalized shopping guide script are then integrated to obtain the shopping guide response content, which is output as follows: The product identifier and recommendation weight of each action node in the dialogue action sequence are analyzed. The basic information of the product is retrieved from the preset product database through the product identifier and sorted by importance in combination with the recommendation weight. The core recommended products are determined according to the sorting result and the corresponding detailed attribute information is organized according to the recommendation weight order to obtain a list of recommended products. Extract the dialogue action sequence of each action node’s dialogue tag and personalized parameters, match the dialogue tag with the preset dialogue generation rule, input the personalized parameters as conditional variables into the dialogue generation rule for variable substitution and semantic expansion to obtain the exclusive introduction text and guiding statement for each recommended product, and concatenate the introduction text and the guiding statement to obtain personalized shopping guide dialogue. Based on the order of the products in the recommended product list, the detailed information of each recommended product is sequentially embedded into the preset placeholder position corresponding to the description text in the personalized shopping guide, generating the shopping guide response content and transmitting it to the user interface through the output interface.

[0073] The product identifier and recommendation weight of each action node in the dialogue action sequence are analyzed. For the action "Recommend 10101", the product identifier is 10101, and the corresponding recommendation weight is 0.85; for the action "Recommend 20505", the product identifier is 20505, and the recommendation weight is 0.76; for the action "Recommend 31202", the product identifier is 31202, and the recommendation weight is 0.68; for the action "Recommend 40803", the product identifier is 40803, and the recommendation weight is 0.62. Basic product information is retrieved from a pre-set product database using the product identifier. The basic information for product 10101 includes: name "Beige Suit Jacket", price 1299.00 yuan, material "Cotton Linen Blend", style "Business Casual", size "S / M / L / XL", and quantity in stock 42 pieces. Based on the recommendation weight, the importance is ranked as follows: 10101 > 20505 > 31202 > 40803. Based on the ranking results, the core recommended products are determined to be 10101 and 20505, and the corresponding detailed attribute information is organized according to the recommendation weight order to obtain a list of recommended products.

[0074] The dialogue action sequence extracts the verbal tags and personalized parameters for each action node. The verbal tag for 10101 is "Classic Style," with personalized parameters including suitable scenarios ("Business Meetings / Formal Occasions") and key selling points ("High-Quality Fabric, Classic Cut"). The verbal tag for 20505 is "Basic Style," with personalized parameters including suitable scenarios ("Daily Office / Business Activities") and key selling points ("Wrinkle-Resistant Fabric, Slim Fit"). Pre-defined verbal generation rules are matched based on the verbal tags. The verbal generation rule corresponding to the "Classic Style" tag is: "This [Product Name] is made of [Material], with [Key Selling Points], making it very suitable for [Suitable Scenarios]. What do you think?" Personalized parameters are input as conditional variables into the verbal generation rule for variable substitution and semantic expansion to obtain the exclusive introductory text for each recommended product.

[0075] For example, the description text for 10101 is: "This beige suit jacket is made of a cotton-linen blend fabric, a high-quality material with a classic cut, perfect for business meetings and formal occasions. What do you think?" The guiding statement is: "We can try it with a shirt." The description text for 20505 is: "This white business shirt is made of wrinkle-resistant fabric with a slim fit, especially suitable for everyday office work and business activities, perfectly showcasing your professional image." The guiding statement is: "It would look even better with dress pants." The description text for 31202 is: "These dark gray dress pants are made of high-grade wool, with a three-dimensional cut that flatters the figure, making them an ideal choice for formal wear." The guiding statement is: "For a more formal occasion, we can add a tie." The description text for 40803 is: "This dark blue silk tie has a soft texture and a classic pattern design, adding a professional touch to the overall look." The guiding statement is: "This whole outfit suits your style perfectly. Would you like to try it on?" By combining these introductory texts and guiding statements, personalized sales pitches can be created.

[0076] Based on the order of the recommended products in the list, the detailed information of each recommended product is sequentially embedded into the preset placeholder positions in the personalized sales guide text. For example, price information is embedded into the price placeholder in the description text, generating "This beige suit jacket is priced at 1299.00 yuan and is made of cotton-linen blend fabric." Size information is embedded into the guiding statement, generating "We have multiple sizes available, S / M / L / XL. You can try it with a shirt to see the effect." The complete sales guide response is generated as follows: "This beige suit jacket is priced at 1299.00 yuan and is made of cotton-linen blend fabric. It features high-quality fabric and a classic cut, making it perfect for business meetings and formal occasions. What do you think? We have multiple sizes available, S / M / L / XL. You can try it with a shirt to see the effect." The sales guide response is then passed to the user interface through the output interface and displayed to the customer.

[0077] In this embodiment, by parsing product identifiers and recommendation weights in the dialogue action sequence and combining them with product database information for importance ranking, the recommendation results are no longer simple path outputs or rating lists. Instead, products are hierarchically divided and prioritized based on recommendation weights, improving the logical clarity and hierarchy of the recommendation display, thereby enhancing the user decision-making guidance effect. By extracting dialogue tags and personalized parameters from the dialogue action sequence and performing variable substitution and semantic expansion based on preset dialogue generation rules, the recommendation content is upgraded from a fixed template expression to a personalized text generation mechanism that can be dynamically adjusted according to user characteristics. This enhances the pertinence and contextual adaptability of language expression, improves the naturalness and affinity of the recommendation content, and thus improves the user interaction experience and acceptance. By embedding the detailed information of recommended products into preset placeholders in the personalized shopping guide dialogue according to weight order, the structured integration of product information and language content is achieved, making the recommendation display more coherent in content organization, enhancing semantic consistency and reading fluency, reducing the sense of information fragmentation, and improving information transmission efficiency.

[0078] A second aspect of this invention provides an intelligent shopping guide robot interactive decision-making system for apparel sales scenarios, comprising: The feature fusion module is used to acquire customers' body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform a suitability score with clothing products in a preset product library to obtain a candidate product set. The path generation module is used to construct a product relationship graph based on the attribute features of products in the candidate product set, propagate customer preference information and update the weights of each product node through a graph diffusion algorithm, calculate the matching degree between different product nodes and the joint representation to obtain a personalized weight distribution by combining an attention mechanism, determine key product nodes based on the personalized weight distribution, extract the neighborhood subgraphs corresponding to the key product nodes and determine multiple recommendation paths by combining a path enumeration algorithm. The decision output module is used to perform Markov chain Monte Carlo sampling on the recommended path and solve for the comprehensive value of the path. The comprehensive value of the path is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. Based on the dialogue action sequence, a list of recommended products is determined. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

[0079] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0080] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0081] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An interactive decision-making method for intelligent shopping guide robots in apparel sales scenarios, characterized in that, include: Acquire customer body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform suitability scoring on clothing products in a preset product library to obtain a candidate product set. Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated and the weights of each product node are updated through a graph diffusion algorithm. The matching degree between different product nodes and the joint representation is calculated by combining an attention mechanism to obtain a personalized weight distribution. Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to the key product nodes are extracted and multiple recommendation paths are determined by combining a path enumeration algorithm. Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. The comprehensive path value is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. A recommended product list is determined based on the dialogue action sequence. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

2. The method according to claim 1, characterized in that, Acquiring customer body posture data and historical behavior trajectories, extracting features from the body posture data to determine body shape characteristics, and extracting behavioral pattern features corresponding to the historical behavior trajectories and constructing a time dependency matrix include: Customer body shape data and historical behavior trajectory are obtained from the customer information collection terminal. Principal component analysis is performed on the body shape data to reduce dimensionality and obtain body shape principal component vectors. Combined with a preset standard body shape library, the body shape classification label of the current customer is determined. Based on the body shape classification label and the body shape principal component vector, body shape features are determined. One-hot encoding is performed on different behavior types in the historical behavior trajectory to obtain a behavior encoding matrix. Combined with the corresponding timestamp information, the behavior encoding matrix is ​​time-weighted to obtain a weighted behavior feature matrix. The weighted behavioral feature matrix is ​​segmented into multiple time segments according to a fixed time window. Statistical analysis is performed on the behavioral features in each time segment to obtain a behavioral frequency vector and a behavioral transition probability vector. The behavioral frequency vector and the behavioral transition probability vector are merged by linear weighted fusion to obtain the behavioral pattern feature vector of each time segment. The Pearson correlation coefficient is calculated based on the behavioral pattern feature vectors of adjacent time segments to obtain the association strength value of each time step. The association strength values ​​between all time steps are arranged in chronological order to construct the time dependency matrix.

3. The method according to claim 1, characterized in that, Based on the body shape characteristics and the time dependency matrix, a joint representation is determined and its suitability is scored with clothing items in a pre-set product library to obtain a candidate product set, including: The body shape features are converted into a multi-scale body shape tensor representation. A graph decomposition operation is performed on the time dependency matrix to obtain the temporal pattern eigenvectors and eigenvalue sequences. A high-order feature interaction is performed on the multi-scale body shape tensor and the temporal pattern eigenvectors through a tensor shrinkage operation to obtain a multi-dimensional interaction tensor. Based on the eigenvalue sequences, dynamic weights are assigned to each dimension of the multi-dimensional interaction tensor and combined with tensor decomposition dimensionality reduction to obtain a joint representation. Multidimensional attribute information of each clothing product is extracted from a pre-defined product database and heterogeneous attribute subgraphs corresponding to each dimension are constructed. The heterogeneous attribute subgraphs are then merged into a unified product semantic graph through a cross-modal alignment mechanism. Based on the joint representation, a subgraph matching algorithm is executed in the product semantic graph to obtain a set of candidate matching subgraphs. For each candidate matching subgraph, graph isomorphism and node feature similarity are calculated to obtain a structural matching score and a semantic matching score. The structural matching score and the semantic matching score are nonlinearly fused to obtain a comprehensive suitability score. An adaptive threshold segmentation is performed on the overall suitability score of all clothing products, and high-score regions are identified using a density clustering algorithm. Product nodes corresponding to the high-score regions are extracted, and the extracted product nodes and their corresponding adjacency relationships are used to form a candidate product set.

4. The method according to claim 1, characterized in that, Based on the attribute features of the products in the candidate product set, a product relationship graph is constructed. Customer preference information is propagated using a graph diffusion algorithm, and the weights of each product node are updated. A personalized weight distribution is obtained by calculating the matching degree between different product nodes and the joint representation using an attention mechanism. The attribute features of each product are extracted from the candidate product set and decomposed into appearance attribute vectors and functional attribute vectors in multiple dimensions. The appearance similarity between products is calculated based on the appearance attribute vectors, and the functional similarity between products is calculated based on the functional attribute vectors. Using products as graph nodes, edge connections are constructed based on the appearance similarity and the functional similarity. The appearance similarity and the functional similarity are nonlinearly combined to obtain a comprehensive similarity, which is used as the edge weight to construct a product relationship graph. The joint representation is injected as an initial preference signal into each node of the product relationship graph. Preference information is transmitted to adjacent nodes along the edge connection. In each diffusion process, the preference information is attenuated and modulated according to the edge weight and accumulated to the target node. After multiple diffusion iterations, the accumulated preference value of each product node is extracted as the updated node weight. The joint representation is decomposed into a query vector, and the attribute features of each product node are decomposed into key vectors and value vectors. The semantic relevance between the joint representation and each product node is calculated through multi-head attention. The updated node weights are combined to obtain a weighted fusion score. The comprehensive matching score of all products is normalized to obtain a personalized weight distribution.

5. The method according to claim 1, characterized in that, Based on the personalized weight distribution, key product nodes are determined. The neighborhood subgraphs corresponding to these key product nodes are extracted, and multiple recommendation paths are determined using a path enumeration algorithm, including: The personalized weight distribution is subjected to probability density analysis to obtain the weight distribution function, and local maxima are identified by peak detection. The product nodes corresponding to the local maxima are extracted as candidate key nodes. The weight gradient and neighborhood density of each candidate key node are calculated and jointly evaluated to obtain key product nodes. Using each key product node as the center node, traverse all adjacent nodes and connecting edges within a preset number of hops to construct a local neighborhood subgraph containing the center node and its corresponding adjacency relationship. Recalibrate the edge weights in the neighborhood subgraph to obtain a local weight matrix. Within each neighborhood subgraph, a depth-first traversal operation is performed based on the local weight matrix. Starting from the center node, a path search is performed along the direction of decreasing edge weights. The sequence of nodes traversed during the traversal and the corresponding cumulative weight values ​​are recorded. When the path length reaches a preset upper limit, the current path search is terminated to obtain an initial path set. Valid paths are determined by threshold filtering based on the corresponding cumulative weight values. The valid paths are sorted according to their cumulative weight values ​​to obtain a candidate recommended path set. Path diversity and path quality indicators are extracted from the candidate recommended path set of each neighborhood subgraph, and multiple recommended paths are obtained by screening the candidate recommended paths through multi-objective optimization criteria.

6. The method according to claim 1, characterized in that, Markov chain Monte Carlo sampling is performed on the recommended path to obtain the comprehensive path value. Using this comprehensive path value as the optimization objective, a genetic algorithm is executed to search for the optimal recommendation strategy and generate a dialogue action sequence, including: The recommended path is decomposed into a node sequence to obtain a path state transition matrix. An initial state distribution and a transition probability distribution are set based on the path state transition matrix. The sampling start node is determined according to the initial state distribution, and multiple sampling paths are generated by step-by-step state jumps according to the transition probability distribution. The access frequency and transition probability of each sampling path are counted to obtain the path stability index. Extract the path length and edge weight information corresponding to the recommended path, calculate the revenue and cost of each recommended path in combination with the path stability index, and solve for the comprehensive path value. Use the comprehensive path value as the fitness value to encode the recommended path set to obtain a chromosome population. Filter the chromosome population according to the fitness value, and iteratively evolve the filtered chromosome population by combining crossover and mutation operations. Repeat the iteration until the preset maximum number of iterations is reached. Extract the chromosome with the highest fitness value after the last iteration and decode it to obtain the optimal recommendation strategy. Based on the path node order and node attribute information in the optimal recommendation strategy, each node is converted into a corresponding recommended action instruction, and the recommended action instructions are concatenated in the path order to obtain a dialogue action sequence.

7. The method according to claim 1, characterized in that, Based on the dialogue action sequence, a recommended product list is determined; based on the dialogue action sequence, a script generation action is determined; and a personalized shopping guide script is determined using a conditional text generation algorithm. The recommended product list and the personalized shopping guide script are then integrated to obtain the shopping guide response content, which is output as follows: The product identifier and recommendation weight of each action node in the dialogue action sequence are analyzed. The basic information of the product is retrieved from the preset product database through the product identifier and sorted by importance in combination with the recommendation weight. The core recommended products are determined according to the sorting result and the corresponding detailed attribute information is organized according to the recommendation weight order to obtain a list of recommended products. Extract the dialogue action sequence of each action node’s dialogue tag and personalized parameters, match the dialogue tag with the preset dialogue generation rule, input the personalized parameters as conditional variables into the dialogue generation rule for variable substitution and semantic expansion to obtain the exclusive introduction text and guiding statement for each recommended product, and concatenate the introduction text and the guiding statement to obtain personalized shopping guide dialogue. Based on the order of the products in the recommended product list, the detailed information of each recommended product is sequentially embedded into the preset placeholder position corresponding to the description text in the personalized shopping guide, generating the shopping guide response content and transmitting it to the user interface through the output interface.

8. An intelligent shopping guide robot interactive decision-making system for apparel sales scenarios, used to implement the method of any one of claims 1-7, characterized in that, include: The feature fusion module is used to acquire customers' body shape data and historical behavior trajectory, extract features from the body shape data and determine body shape features, extract behavioral pattern features corresponding to the historical behavior trajectory and construct a time dependency matrix, determine joint representation based on the body shape features and the time dependency matrix, and perform a suitability score with clothing products in a preset product library to obtain a candidate product set. The path generation module is used to construct a product relationship graph based on the attribute features of products in the candidate product set, propagate customer preference information and update the weights of each product node through a graph diffusion algorithm, calculate the matching degree between different product nodes and the joint representation to obtain a personalized weight distribution by combining an attention mechanism, determine key product nodes based on the personalized weight distribution, extract the neighborhood subgraphs corresponding to the key product nodes and determine multiple recommendation paths by combining a path enumeration algorithm. The decision output module is used to perform Markov chain Monte Carlo sampling on the recommended path and solve for the comprehensive value of the path. The comprehensive value of the path is used as the optimization objective to perform a genetic algorithm search to determine the optimal recommendation strategy and generate a dialogue action sequence. Based on the dialogue action sequence, a list of recommended products is determined. Based on the dialogue action sequence, a speech generation action is determined and a personalized shopping guide speech is determined through a conditional text generation algorithm. The recommended product list and the personalized shopping guide speech are integrated to obtain the shopping guide response content and output it.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images