Clothing e-commerce product selection system self-learning method based on closed-loop feedback mechanism

Abstract

This invention discloses a self-learning method for a product selection system in apparel e-commerce based on a closed-loop feedback mechanism. The method includes: real-time collection and structuring of user behavior, operational, and product data; decomposing product selection decisions into strategy units, determining dynamic weights of indicators using the entropy weight method, and generating a comprehensive score, thereby outputting fine-tuning, reconstruction, or elimination signals; under elimination signals, using a linear combination of the score, historical mean, and standard deviation of fluctuation as the efficiency index, eliminating the strategy when it continuously falls below the benchmark, and cross-combining the features of surviving strategies to generate a new strategy for replacement; under fine-tuning or reconstruction signals, retrieving the historical base model most similar to the features of the strategy to be deployed from the model repository, performing parameter fine-tuning or structural reconstruction according to the signal type, and deploying the updated strategy. This invention enables the system to autonomously generate new strategies while eliminating inefficient strategies through a strategy regeneration mechanism, achieving full-link automated closed-loop self-evolution.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a self-learning method for a product selection system in the apparel e-commerce sector based on a closed-loop feedback mechanism. Background Technology

[0002] Currently, product selection systems in the apparel e-commerce sector generally employ static machine learning models based on historical data. Their core limitation lies in the lack of dynamic self-evolution capabilities. Such systems typically rely on periodic batch data training, resulting in model updates lagging significantly behind market changes. When sudden fashion trends or supply chain fluctuations occur, the system cannot perceive and adjust its decision-making logic in real time, leading to a mismatch between recommended products and users' real-time needs. More importantly, the traditional architecture separates decision execution from feedback: user behavior data, such as click-through rates, conversion rates, return reasons, inventory turnover rates, and gross profit margins, are not effectively converted into model optimization signals, trapping the system in a passive cycle of decision-making failure and manual intervention.

[0003] To address the aforementioned issues, existing technologies attempt to incorporate collaborative filtering, time series forecasting, and other methods to improve prediction accuracy. These methods have improved the accuracy of product selection to some extent, attempting to compensate for the shortcomings of static models through algorithm optimization.

[0004] However, these improvements remain at the level of local optimization and fail to fundamentally solve the system's dynamic adaptability problem. First, model updates require manually defined triggering conditions and parameter adjustment strategies, consuming significant operational resources and resulting in response delays exceeding 48 hours. Second, feedback data is only used for fine-tuning local parameters, lacking a global evaluation of decision-making strategies and a mechanism for weeding out underperforming products. Finally, in cold-start scenarios, new product decisions rely on manual experience rules, leading to an accuracy rate of less than half for identifying potential products. These shortcomings make the existing system ill-suited to the highly volatile and rapidly iterating nature of the apparel industry, necessitating an intelligent decision-making architecture with closed-loop self-evolutionary capabilities. Summary of the Invention

[0005] This invention proposes a self-learning method for apparel e-commerce product selection systems based on a closed-loop feedback mechanism. This method solves the problem that existing apparel e-commerce product selection systems lack closed-loop self-evolution capabilities, resulting in model updates lagging far behind market changes and failing to effectively convert real-time feedback data into signals for autonomous system optimization.

[0006] To address the aforementioned technical problems, this invention provides a self-learning method for a product selection system in apparel e-commerce based on a closed-loop feedback mechanism, comprising the following steps:

[0007] Step S1: Collect and aggregate user behavior data, business operation data, and product performance data from the apparel e-commerce product selection system in real time to obtain structured indicators;

[0008] Step S2: Decompose the product selection decision into multiple strategy units, associate a set of structured indicators with each strategy unit, determine the dynamic weights of the structured indicators through the entropy weight method, and obtain a comprehensive evaluation score by weighted summation, and generate fine-tuning signals, reconstruction signals or elimination signals based on the comprehensive evaluation score;

[0009] Step S3: In response to the elimination signal, the linear combination of the comprehensive evaluation score of the strategy unit in the current period, the average score of the historical periods, and the standard deviation of the score fluctuation is used as the strategy effectiveness index. When the strategy effectiveness index is lower than the industry benchmark value for a consecutive preset number of periods, the corresponding strategy unit in the strategy pool is eliminated, and the strategy features are extracted from the surviving strategy units and cross-combined to generate a new strategy unit to be added to the strategy pool.

[0010] In response to fine-tuning or reconstruction signals, retrieve historical base models from the model version repository, select the matching model with the highest cosine similarity to the feature vector of the strategy unit to be deployed, fine-tune the parameters or reconstruct the structure of the matching model according to the signal type, update it, store it in the repository, and deploy it for execution.

[0011] Preferably, the user behavior data in step S1 includes click-through rate, conversion rate, and return reason analysis results; the business operation data includes inventory turnover rate, gross profit margin, and slow-moving goods rate; and the product performance data includes new product identification accuracy and trend response delay.

[0012] The real-time acquisition and aggregation includes: stamping a unified timestamp as a time series benchmark when the original event is generated by the event collector; aggregating high-frequency data at the second and minute levels in real time through a stream processing engine; performing micro-batch processing on hourly data; and calculating low-frequency data at daily and weekly levels according to a fixed period to obtain the structured indicators.

[0013] Preferably, the return reason analysis results are obtained through the following steps:

[0014] Data cleaning and text normalization are performed on the original return text submitted by users. Key entities are extracted from the original return text using matching rules based on the business thesaurus and named entity recognition model.

[0015] The text after key entities are extracted is classified using a pre-trained text classification model and a business keyword library. When the classification result of the text classification model matches the matching result of the business keyword library, a high confidence level is assigned and the classification result is adopted. When the classification result of the text classification model conflicts with the matching result of the business keyword library, the classification result of the text classification model is adopted first and the confidence level is reduced.

[0016] In each calculation period, the classification results with confidence levels higher than a preset threshold are aggregated, and the proportion of each reason for return within the calculation period is calculated.

[0017] Preferably, step S2 involves determining the dynamic weights of the structured indicators using the entropy weight method, and then weighting and summing them to obtain the comprehensive evaluation score of the strategy unit, including the following steps:

[0018] Using the extreme value method to the first Item structured indicators in Observations on a sample Standardize to obtain Calculate the first Information entropy of structured indicators :

[0019] ;

[0020] ;

[0021] ;

[0022] In the formula, This represents the total number of samples; j represents the index of the j-th sample. ;

[0023] The basic weights are determined based on the information entropy:

[0024] ;

[0025] The structured indicators are weighted and summed based on the aforementioned basic weights to obtain the comprehensive evaluation score of the strategy unit:

[0026] ;

[0027] ;

[0028] In the formula, Indicates the first The overall evaluation score of each strategy unit; Indicates the relationship with the first The total number of structured metrics associated with each strategy unit; Indicates the first The dynamic weights of the structured indicators after being processed by the confidence decay factor; Indicates the confidence decay factor; Indicates the first The basic weights of the structured indicators are calculated using the entropy weighting method. This represents the normalized index value; This represents the index conversion function.

[0029] Preferably, the index transformation function This includes at least one of the following: a logarithmic transformation function for exponential growth characteristic indicators, a linear normalization function for percentage indicators, a sigmoid compression function for suppressing abnormal fluctuations, and a piecewise decay function for distinguishing between target values ​​and out-of-target values.

[0030] Preferably, step S2 generates a Pareto solution set based on the comprehensive evaluation score of each strategy unit, and then performs a weighted score and sorting of the Pareto solution set using the business objective weight matrix to select the champion strategy.

[0031] A / B testing is performed on the champion strategy and the current online strategy. The significance level of the performance improvement of the champion strategy is determined by t-test. If the p-value is less than 0.05, the champion strategy is fully deployed; otherwise, the strategy is rolled back to the current online strategy.

[0032] Preferably, the expression for calculating the strategy effectiveness index in step S3 is:

[0033] ;

[0034] In the formula, Indicates the first The strategy effectiveness index of each strategy unit; Indicates the first The overall evaluation score of each strategy unit in the current evaluation period; Indicates the first The comprehensive evaluation score of each strategy unit in the t-th historical period; Indicates the total number of historical cycles; Indicates the first The standard deviation of the volatility of the overall evaluation score of each strategy unit; , , Indicates the weighting coefficient;

[0035] The industry benchmark value is obtained by sorting the strategy effectiveness indices of all strategy units in the strategy pool and taking the average of the strategy effectiveness indices of the top 30% of strategy units.

[0036] Preferably, the corresponding policy unit in the elimination policy pool in step S3 further includes:

[0037] Determine whether the eliminated strategy unit in the strategy pool has a dependent strategy. If a dependent strategy exists, initiate strategy reorganization. If no dependent strategy exists, release the weight resources occupied by the eliminated strategy unit and allocate the released weight resources to other strategy units in the strategy pool with similar functions and located at the Pareto front.

[0038] Extracting policy features from the policy pool and combining them to generate new policy units to supplement the policy pool includes: selecting at least two surviving policy units with the highest policy effectiveness index from the policy pool as parent policies; extracting the policy feature vectors of each parent policy; exchanging segments according to random crossover points to generate a mixed feature vector; applying random perturbation to some dimensions of the mixed feature vector with a preset mutation probability; and initializing new candidate policy units based on the perturbated mixed feature vector to supplement the policy pool.

[0039] Preferably, the formula for calculating the cosine similarity of feature vectors in step S3 is:

[0040] ;

[0041] In the formula, Let represent the cosine similarity of the feature vectors, and sim∈[0,1]; The feature vector representing the strategy unit to be deployed; The feature vector representing the historical basis model; and They represent and The modulus length;

[0042] The feature vectors of the strategy unit to be deployed and the historical base model are both generated by encoding the weight coefficients and decision preference parameters of the associated strategy unit through an embedding layer; the matching model is the historical base model with the highest cosine similarity to the feature vector of the strategy unit to be deployed that is greater than a preset similarity threshold; when the cosine similarity of the feature vectors of all historical base models is lower than the preset similarity threshold, the strategy unit to be deployed is stored as a new base model in the model version repository.

[0043] Preferably, updating the matching model based on the fine-tuning signal and the reconstructed signal in step S3 includes the following steps:

[0044] The feature offset is calculated by subtracting the cosine similarity of the feature vector from 1. When the feature offset corresponding to the fine-tuning signal is less than 0.7, online knowledge distillation technology is used to fine-tune the parameters of the matching model based on a loss function. The expression for the loss function is as follows:

[0045] ;

[0046] In the formula, This represents the total loss from knowledge distillation; Represents the task loss function; Indicates the output distribution of the teacher model With student model output distribution KL divergence between them; Indicates temperature parameter; , Here are the weight parameters, where and ;

[0047] When the feature offset corresponding to the reconstructed signal is greater than or equal to 0.7, at least one of the following is performed on the matching model: attention mechanism addition, full-connected layer dimension adjustment, and residual connection introduction to perform structural reconstruction;

[0048] When a dramatic change in the market environment is detected, a dedicated sub-network is constructed by extracting general features from the matching model through knowledge transfer, and the dedicated sub-network is dynamically activated to work in collaboration with the main network of the matching model through a gating mechanism.

[0049] The advantages of this invention include at least the following:

[0050] 1. Based on an event-driven architecture, it can process thousands of data streams per second, and can capture core indicators of user behavior, business operations and product performance in real time. It can also push structured indicators to subsequent steps in real time through the API gateway. At the same time, the branch expansion mechanism can quickly adjust strategies for emerging market trends, greatly reduce trend response latency, and enable the system to adapt to the rapidly changing market rhythm of the apparel industry.

[0051] 2. Built-in multi-objective Pareto front optimization algorithm can effectively balance the conflict between multiple business objectives such as gross profit margin and inventory turnover rate, avoiding the situation in traditional systems where the overall interests are sacrificed for a single objective; through global evaluation and dynamic adjustment of product selection strategy, the proportion of slow-moving products can be reduced, inventory turnover efficiency can be improved, and at the same time, a reasonable proportion of high-margin products can be ensured, so as to achieve two-way optimization of inventory health and profitability. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0053] 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 protection scope of the present invention.

[0054] like Figure 1 As shown, this embodiment of the invention provides a self-learning method for a product selection system in apparel e-commerce based on a closed-loop feedback mechanism, including the following steps:

[0055] Step S1: Collect and aggregate user behavior data, business operation data, and product performance data from the apparel e-commerce product selection system in real time to obtain structured indicators.

[0056] Specifically, this step connects the decision-making and feedback stages through an automated data pipeline, monitors user behavior and business metrics in real time, and transforms raw data into structured indicators, which are then input into the subsequent evolutionary effect evaluation stage. This mechanism eliminates the need for manual intervention, ensuring that the system can respond to market changes in a short period of time.

[0057] The data collection scope covers three major dimensions: user behavior data, business operation data, and product performance data. Each dimension is further subdivided into multiple indicators to ensure a holistic evaluation rather than localized fine-tuning. User behavior data includes click-through rate (CTR), conversion rate, and return reason analysis results. Business operation data includes inventory turnover rate, gross profit margin, and slow-moving inventory rate. Product performance data includes new product identification accuracy and trend response latency. The formula for calculating CTR is as follows:

[0058] .

[0059] The formula for calculating conversion rate (CVR) is:

[0060] .

[0061] Inventory turnover ratio, calculated by dividing cost of goods sold by average inventory, is used to assess supply chain efficiency. The formula is as follows:

[0062] .

[0063] Gross profit margin is calculated by dividing sales revenue by cost of goods sold, and is used to reflect profitability health. The slow-moving inventory rate is the number of slow-moving SKUs divided by the total number of SKUs, used to identify product selection errors. SKU stands for Inventory Unit, referring to the smallest independently manageable inventory unit for a product, differentiated by attributes such as color and size. New product identification accuracy is the percentage of system-recommended new products that actually become bestsellers, used to address the cold start problem. Trend response latency is the average time from the emergence of a trend to the system's response, used to quantify agility. The monitoring frequency and specific meaning of each indicator are shown in Table 1.

[0064] Table 1. Core Dimensions and Specific Components of Structured Indicators

[0065]

[0066] For data with different monitoring frequencies, this embodiment of the invention employs a hierarchical, event-driven data pipeline for collaborative processing, rather than demanding complete physical synchronization of all data. This mechanism begins with the establishment of a unified time series benchmark. All original events, such as user clicks or completed purchases, are instantly tagged with a high-precision unified timestamp by the event collector, providing a unique and reliable time series benchmark for the alignment of all subsequent data.

[0067] Building upon this time-series benchmark, the system employs a tiered buffering and on-demand alignment intelligent strategy. Data of different frequencies enters different processing channels: high-frequency data at the second and minute levels, such as click-through rates and conversion rates, are aggregated in real-time via a stream processing engine; hourly data, such as return reason analysis results, undergoes micro-batch processing; and low-frequency data at the daily and weekly levels, such as gross profit margin and slow-moving inventory rate, are calculated on a fixed schedule. When the upper-level step S2 needs to evaluate the effectiveness of a strategy unit, it sends a request to the engine responsible for data aggregation. This engine performs a signal readiness check, assembling the most complete and relevant dataset for the current moment for that evaluation. This dataset might include real-time click-through rates from the past five minutes or gross profit margin data from the entire previous day. This on-demand assembly mechanism ensures that evaluations are conducted promptly based on the best available information.

[0068] The final decisive step is a weighted fusion based on time-sensitive confidence. The evaluation process does not treat data from different time periods equally; in the core formula for calculating the comprehensive evaluation score of the strategy unit, the dynamic weight of each indicator is multiplied by a confidence decay factor.

[0069] ;

[0070] Among them, the most recently updated high-frequency data enjoys the highest confidence level. A confidence level close to 1 fully reflects the impact; while low-frequency data marked as yesterday or earlier has its confidence level appropriately lowered to balance its influence in decision-making. This mechanism enables the system to react quickly based on high-frequency signals while ensuring that these reactions do not deviate from the long-term strategic direction anchored by low-frequency core business indicators, achieving a balance between agility and robustness.

[0071] To obtain the results of the return reason analysis, this embodiment designs a three-step processing flow based on NLP classification technology. The first step is text preprocessing and key information extraction. The system receives the original return text submitted by users in batches through the data interface, such as the actual color being much darker than the picture or the sleeve length being unsuitable. Then, a standardized text preprocessing pipeline is started. First, data cleaning is performed to remove meaningless special symbols and perform text normalization. Next, using matching rules based on the business lexicon and the named entity recognition model, key entities are accurately extracted from the text, such as color, size, sleeve, fabric, and loose threads. This initially structures the messy natural language description, laying the foundation for subsequent accurate classification.

[0072] The second step is multi-label classification and confidence assessment based on a hybrid model, which is the core of the entire process. Its goal is to accurately determine the return reason category for each piece of text. The system employs a hybrid classification strategy to balance accuracy and efficiency. First, a pre-trained text classification model performs semantic understanding of the text, providing preliminary classification results and probabilities. Simultaneously, a continuously updated business keyword database containing industry-specific terms and synonyms performs rapid matching. Finally, the decision fusion module processes the results of these two methods. When the classification result of the text classification model matches the matching result of the business keyword database, it is directly adopted and assigned a high confidence level. When results conflict, the classification result of the text classification model is prioritized, but the confidence level is appropriately lowered, and this data is recorded as a potential special case for manual review to optimize the system.

[0073] The third step is dynamic aggregation, trend analysis, and signal output. After classification, the system enters the statistics and insight generation stage. In each calculation cycle, all high-confidence classification results are aggregated to calculate the specific proportion of each return reason. The calculation formula is the number of categories for that reason divided by the total number of categories in the cycle and then multiplied by 100%. Finally, these structured results are pushed to step S2 in real time via API as key feedback signals to evaluate the accuracy of product descriptions, size recommendation algorithms, and other related product selection strategies, thereby driving the system's self-optimization.

[0074] As a key indicator of product performance data, the accuracy of new product identification relies on the objective determination of best-selling products. In this embodiment, a product is judged as a best-seller by the system based on a simple and quantifiable combination of dynamic indicators, specifically including the following three core dimensions: First, the sales explosiveness dimension, where the product's sales revenue or sales volume growth rate within a specific time window, such as 24 hours, must exceed 300%, indicating a sharp increase in market demand; second, the traffic conversion efficiency dimension, where the product's click-through rate and conversion rate must consistently be more than twice the average level of similar products, indicating high attention and effective conversion of traffic into purchases; and finally, the inventory health dimension, where the product's daily sales turnover rate must be greater than 5% and the sell-through rate must reach more than 80% within a specific period, reflecting its strong sales speed and good inventory turnover, thus avoiding misjudging clearance items as best-sellers. When a product simultaneously meets the above indicator thresholds within a continuous monitoring period, the system automatically marks it as a best-seller, thereby ensuring the objectivity and timeliness of the determination.

[0075] Trend response delay specifically refers to the point in time when a quantifiable emerging consumer trend first reaches a significant level in the platform's overall data. The time from when the system first successfully lists or promotes related products in response to this trend. The time interval between these steps quantifies the system's agility in moving from trend perception to effective decision-making. Its calculation relies on an automated trend detection and matching engine, consisting of three steps: Step one is determining the starting point of the trend. The system continuously monitors aggregated data from search terms, browsing behavior, and social media topics across the entire platform. When the growth rate of the volume of a specific topic, such as a checkerboard element, exceeds a preset threshold within a continuous time window, and its absolute value reaches its minimum, that point in time is determined to be the starting point of the trend. Step two involves determining the system's response point. The system simultaneously monitors its product selection and recommendation decisions. When the system first elevates a product matching this trend to a core traffic position and significantly increases its exposure, this time point is recorded as [the response point]. Step three is the delay calculation; the trend response delay is the difference between the two, usually expressed in minutes or hours.

[0076] ;

[0077] Step S1 employs a distributed data pipeline for real-time data acquisition. The architecture comprises three core components: First, an event collector deployed on both the user and server sides, capturing click, purchase, and return events. Data is streamed, processing a peak of 10,000 records per second. Second, a dimension aggregation engine cleanses and aggregates the raw data, outputting structured metrics. Third, a feedback loop pushes metrics to step S2 in real-time via an API gateway, forming a closed loop of monitoring and transmission. This mechanism solves the problem of constructing an automated closed-loop feedback channel, avoiding the lag of traditional batch training.

[0078] Step S2: Decompose the product selection decision into multiple strategy units, associate a set of structured indicators with each strategy unit, determine the dynamic weights of the structured indicators using the entropy weight method, and obtain a comprehensive evaluation score by weighted summation. Based on the comprehensive evaluation score, generate fine-tuning signals, reconstruction signals, or elimination signals.

[0079] A strategy matrix architecture is adopted to decompose product selection decisions into independently evaluable atomic strategy units. A strategy unit is the smallest evaluation unit formed after the system breaks down complex product selection decisions. Each strategy unit represents a specific decision logic; for example, a best-selling product prediction strategy is specifically responsible for identifying potential new products, while an inventory clearance strategy is specifically responsible for handling slow-moving goods. Strategy units are generated primarily from two aspects: First, based on abstract extraction from historically successful decision patterns, the system automatically solidifies validated and effective decision logic into independent strategy units. Second, for newly emerging business scenarios, the system initializes basic strategy units according to preset rule templates, such as creating exclusive product selection strategies for new product categories. Third, they are automatically generated through the strategy regeneration mechanism in step S3. When inefficient strategies are eliminated, the system extracts strategy features from surviving efficient strategies, performs cross-combinations and mutation perturbations, and automatically generates new candidate strategy units to replenish the strategy pool, so that the source of strategy units is no longer limited to manual configuration and historical extraction.

[0080] The decomposition of strategy units follows these operational standards: First, the decision space is divided into several non-overlapping sub-target domains based on the business objectives of product selection decisions. Each sub-target domain corresponds to a specific decision requirement, such as a new product identification sub-target domain, a slow-moving inventory handling sub-target domain, and a high conversion rate sub-target domain. For each sub-target domain, the system extracts decision sequences that have achieved significant results under that sub-target domain from historical decision logs. A sequence pattern mining algorithm is used to extract frequently occurring strategy feature combinations, and these combinations are solidified into strategy units for that sub-target domain. When a new business scenario emerges, the system initializes a strategy unit for that scenario based on a preset rule template. The rule template includes three basic fields: a set of associated indicators, an initial weight vector, and an evaluation period. Strategy units are distinguished by unique identifiers, and each strategy unit is independently stored, evaluated, and iterated in the strategy pool.

[0081] The initialization process of strategy units employs a progressive optimization mechanism. Newly generated strategy units receive a set of basic weight parameters, which are initialized using Bayesian optimization methods to ensure the strategy unit possesses basic decision-making capabilities upon deployment. Simultaneously, the system establishes an independent evaluation profile for each new strategy unit, recording its performance data. The mapping relationship between strategy units and the underlying model is reflected in the fact that strategy units act as the concrete bearers of model capabilities; each strategy unit corresponds to the performance of the underlying model in a specific decision-making dimension. For example, the "hot item prediction" strategy unit actually evaluates the model's ability in trend prediction. The performance evaluation results of the strategy units, in turn, guide the parameter optimization of the underlying model, forming a closed-loop feedback mechanism from strategy evaluation to model optimization. Through this design, the system can perform fine-grained evaluation of overall decision-making capabilities while ensuring that each strategy unit can be independently optimized and iterated, thereby achieving precise and efficient self-evolution.

[0082] In calculating the overall evaluation score of each strategy unit, each strategy unit is associated with a set of structured indicators, and its overall evaluation score is calculated through a weighted evaluation function:

[0083] ;

[0084] In the formula, Indicates the first The overall evaluation score of each strategy unit; Indicates the relationship with the first The total number of structured metrics associated with each strategy unit; Indicates the first The dynamic weights of the structured indicators after being processed by the confidence decay factor are calculated in real time using the entropy weight method. This represents the normalized metric value, for example, mapping the click-through rate (CTR) to the [0,1] interval; This represents the indicator transformation function, such as using a logarithmic transformation function to eliminate the influence of dimensions. This model breaks through the limitations of traditional static evaluation. For example, when the new product identification accuracy indicator fluctuates abnormally, the system can automatically reduce the weight of relevant strategy units to avoid global misjudgment caused by the distortion of a single indicator.

[0085] Basic weights The calculation uses the entropy weight method, and the specific steps are as follows: First, the original index data is standardized. For the first... Each indicator in Observations on a sample Standardization is performed using the extreme value method:

[0086] ;

[0087] in ∈[0,1], representing the standardized index value, when the... When the maximum and minimum values ​​of a structured indicator are equal across all samples, the standardized value of the current indicator is... Set all values ​​to 0; the corresponding information entropy Set it to 1, thus increasing the weight of this indicator. It will automatically be reduced to 0 and will not be included in the calculation of the overall evaluation score.

[0088] Then calculate the first... The probability distribution and information entropy value of each indicator:

[0089] ;

[0090] ;

[0091] In the formula, This represents the total number of samples; j represents the index of the j-th sample. .

[0092] when When =0, it is stipulated that =0. Finally, the weights of each indicator are calculated based on the information entropy:

[0093] ;

[0094] In the formula This represents the total number of indicators, and each weight satisfies the normalization constraint. Basic weights The entropy value of a certain indicator is recalculated every 4 hours. When the value is close to 1, it indicates that the indicator has very little difference across all samples, and the weight... The importance of indicators will automatically decrease; conversely, the greater the difference, the higher the weight given to the indicator. This method can automatically identify changes in the importance of each indicator, ensuring that the evaluation results objectively reflect the actual business situation.

[0095] Indicator Conversion Function This includes at least one of the following: a logarithmic transformation function for indicators with exponential growth characteristics; a linear normalization function for percentage-based indicators; a sigmoid compression function for indicators that suppress abnormal fluctuations; and a piecewise decay function for distinguishing between target and exceedance values. Specifically, the logarithmic transformation function is suitable for indicators with exponential growth characteristics, such as click volume and sales revenue, and its expression is:

[0096] .

[0097] This function can compress the scale of data, reduce the impact of extreme values ​​on the overall evaluation, and make the indicator distribution closer to a normal distribution. For example, when processing product pageviews, using logarithmic transformation can prevent a few best-selling products from excessively dominating the evaluation results. The linear normalization function is suitable for percentage-based indicators with clear upper and lower bounds, such as conversion rate and gross profit margin. Its expression is:

[0098] ;

[0099] In the formula and These are the maximum and minimum values ​​of the indicator within the statistical period, respectively. This function maps the indicator value to the [0,1] interval, eliminating the influence of dimensions and facilitating weighted comparisons between different indicators. The Sigmoid compression function is suitable for indicators that need to suppress the influence of extreme values, such as return rate and inventory turnover rate. Its expression is:

[0100] ;

[0101] Parameters in the formula Control function slope, parameters By setting a center point, this function smoothly compresses the original indicator value to the (0,1) interval, making it particularly suitable for handling operational indicators that may exhibit abnormal fluctuations. The piecewise decay function is suitable for indicators that require differentiated treatment of different numerical ranges, such as trend response delay; its expression is:

[0102] ;

[0103] In the formula The ideal delay threshold; The weighting coefficient assigns full marks to compliant values ​​and exponentially decays scores to values ​​exceeding the target, reflecting the business's stringent requirements for timeliness. The parameters of the indicator conversion function are determined through training on historical data and are automatically updated every 24 hours to ensure the conversion rules match the current business environment. The evaluation dimensions for typical strategy units are shown in Table 2.

[0104] Table 2. Evaluation Dimension Design for Typical Strategy Units

[0105]

[0106] Regarding the multi-objective evolutionary mechanism, the core responsibility of step S2 is to resolve conflicts between multiple business objectives and decide on the final strategy to be implemented from a set of non-dominated solutions. This mechanism is based on the comprehensive evaluation score of the strategy units, but the focus shifts from the absolute effectiveness score of a single strategy to the comparison and selection of the relative merits of multiple strategies.

[0107] In practice, step S2 generates a Pareto solution set based on the comprehensive evaluation scores of each strategy unit. The Pareto solution set consists of a group of excellent strategies with different focuses, and the final selection is made using a business objective weight matrix. The strategy score is obtained by multiplying the gross profit margin weight, turnover rate weight, and slow-moving inventory rate weight by their respective performance values ​​and then summing the results. The generation of the Pareto solution set uses a non-dominated ranking method: for all strategy units in the strategy pool, a non-dominated relationship comparison is performed based on their comprehensive evaluation scores across all business objective dimensions. If strategy A performs no worse than strategy B on all objectives, and is superior to strategy B on at least one objective, then strategy A is said to dominate strategy B. All strategy units not dominated by any other strategy constitute the first-level Pareto front, i.e., the Pareto solution set.

[0108] The weights of business targets have a dynamic adjustment mechanism. During peak sales seasons, the gross profit margin weight automatically increases to 0.7, while the turnover rate weight decreases to 0.3. During clearance seasons, the turnover rate weight increases to 0.6, while the gross profit margin weight decreases to 0.2. During normal periods, the weights of each target remain balanced, for example, allocated in the ratio of 0.4, 0.4, and 0.2.

[0109] The winning strategy selected in the previous step will not be immediately deployed to the entire system. To address potential performance differences between the strategy in simulated and real-world environments, the system will conduct A / B testing on it against the current best online strategy. A small portion, such as 5%, of user traffic will be redirected to the new strategy, and key metrics will be continuously monitored. Statistical tests, such as t-tests, will be used to determine whether the performance improvement is significant, with a p-value less than 0.05 as the significance threshold. Only strategies that pass the tests and are verified to be effective will be officially deployed. Otherwise, the system will roll back to the original strategy and record the experience gained from this exploration to optimize the decision function, thereby ensuring the robustness of the system's evolution.

[0110] Based on continuous performance monitoring, step S2 automatically generates three types of optimization signals to drive the downstream policy layer: fine-tuning signals are used to identify parameter-level optimization opportunities and trigger weight coefficient adjustments; reconstruction signals are used to detect structural defects and initiate model architecture updates; and elimination signals are used to identify persistently inefficient policies and recommend their removal from the policy pool. The final decisions generated by this mechanism and the policy performance change trends are transmitted in real time to the downstream step S3 via a standardized signal through an API interface, driving policy pool updates and iterative model optimization.

[0111] Step S3: In response to the elimination signal, the strategy unit's comprehensive evaluation score in the current period, the average score of the historical periods, and the standard deviation of the score fluctuation are used as the strategy performance index. When the strategy performance index is lower than the industry benchmark value for a preset number of consecutive periods, the corresponding strategy unit in the strategy pool is eliminated, and the strategy features are extracted from the surviving strategy units and cross-combined to generate new strategy units to be added to the strategy pool.

[0112] In response to fine-tuning or reconstruction signals, retrieve historical base models from the model version repository, select the matching model with the highest cosine similarity to the feature vector of the strategy unit to be deployed, fine-tune the parameters or reconstruct the structure of the matching model according to the signal type, update it, store it in the repository, and deploy it for execution.

[0113] Specifically, the fine-tuning signal is triggered when the overall evaluation score of the strategy unit is higher than the industry benchmark but shows a downward trend for two consecutive cycles; the reconstruction signal is triggered when the overall evaluation score is lower than the industry benchmark but does not meet the continuous elimination cycle condition; and the elimination signal is triggered when the strategy effectiveness index is lower than the industry benchmark for a preset number of consecutive cycles. The preset number of cycles is set according to the business scenario, and is usually in the range of 2 to 5 cycles. In this embodiment, 3 cycles are used.

[0114] A quantitative elimination threshold system is adopted, and the strategy pool is dynamically managed through a strategy effectiveness index. The strategy effectiveness index integrates short-term performance and long-term stability, and its calculation formula is as follows:

[0115] ;

[0116] In the formula, Indicates the first The strategy effectiveness index of each strategy unit; Indicates the first The overall evaluation score of each strategy unit in the current evaluation period; Indicates the first The comprehensive evaluation score of each strategy unit in the t-th historical period; Indicates the total number of historical cycles; Indicates the first The standard deviation of the volatility of the overall evaluation score of each strategy unit; , , This represents the weighting coefficient, with default values ​​of 0.7, 0.2, and 0.1. When a strategy unit's strategy performance index falls below the industry benchmark for three consecutive periods, the system automatically triggers a termination process. The industry benchmark is automatically generated by analyzing the historical performance data of all strategy units in the strategy pool. The system calculates the distribution of the strategy performance index daily and takes the average of the strategy performance indices of the top 30% of strategy units as the benchmark for that day. This industry benchmark is updated periodically and dynamically adjusted according to the performance of the strategy pool.

[0117] The elimination decision also considers strategy relevance. If eliminating a strategy would cause dependent strategies to fail, the system will initiate strategy reorganization instead of direct deletion. The adjustment targets are the dynamic weights in the comprehensive evaluation score formula for strategy units. Dynamic weights It is the final weight after processing with the time-dependent confidence decay factor, i.e. The core of the adjustment is the basic weight. The specific adjustment process includes: first, weight release, in which the weight resources occupied by the eliminated strategy unit are released; second, weight redistribution, in which the system will prioritize the allocation of these released weights to other strategy units that perform well on the Pareto front and have similar functions. For example, if a best-selling product prediction strategy is eliminated, the weight share of its original new product identification accuracy and trend response delay indicators will be partially transferred to other efficient best-selling product prediction strategies in the pool to ensure that such decision-making capabilities are not weakened.

[0118] The typical elimination scenarios in step S3 include three types: continuous inefficiency, goal conflict, and environment mismatch. The specific judgment logic is shown in Table 3.

[0119] Table 3. Judgment Logic for Typical Elimination Scenarios

[0120]

[0121] After the elimination process is completed, the system starts the policy regeneration mechanism to prevent the policy pool from shrinking due to continuous elimination.

[0122] Specifically, the system selects at least two surviving policy units with the highest policy effectiveness index from the policy pool as parent policies, and extracts the policy feature vectors of each parent policy.

[0123] Before performing the cross operation, the system semantically groups the policy feature vectors, marking dimensions with strong coupling relationships, such as the weight coefficients and preference parameters of the same indicator, as the same indivisible feature block. The selection of cross points is only performed on the boundaries of the feature blocks, avoiding the splitting of coupled dimensions and ensuring that the generated mixed feature vectors maintain logical internal consistency.

[0124] The system randomly selects an intersection point between two parent feature vectors and swaps the feature segments before and after the intersection point between the two parents to generate a hybrid feature vector. This hybrid feature vector inherits the advantageous features of each parent strategy while introducing the diversity generated by feature combinations. The random intersection point is uniformly and randomly selected within the effective dimension range of the strategy feature vector, avoiding the selection of the beginning or end endpoints to ensure the effective length of the swapped segments. The preset mutation probability range is 0.01 to 0.1; in this embodiment, it is set to 0.05. The mutation probability is set to a low value to avoid drastic distortion of the strategy features and ensure the usability of the new strategy. The random perturbation amplitude is a random offset within ±10% of the current dimension value.

[0125] Based on this, the system applies random perturbations to some dimensions of the hybrid feature vector with a preset mutation probability. The perturbation magnitude can be set as a random offset within a preset proportion range of the current dimension value. The purpose of the mutation operation is to enhance the breadth of policy exploration and prevent the policy pool from getting trapped in local optima. Finally, the system initializes a new candidate policy unit based on the perturbated hybrid feature vector, assigns it an initial evaluation file, and adds it to the policy pool. The newly generated candidate policy unit enjoys a cold start protection exemption in the first week and is incorporated into the regular policy performance evaluation process after accumulating sufficient evaluation data.

[0126] Through the strategy regeneration mechanism, step S3 is upgraded from simple strategy elimination to strategy ecosystem management that combines elimination and generation. This enables the strategy pool to continuously eliminate inefficient strategies while autonomously replenishing candidate strategies with new feature combinations, ensuring the stability of the strategy pool's size and the diversity of its features.

[0127] Efficient iteration is achieved through a dynamic model repository and intelligent matching strategies, strictly adhering to the core principle of lightweight updates based on historical models. The system constructs a two-layer version repository: the base model layer stores verified stable versions, and the incremental version layer records derived versions. When a strategy unit to be deployed arrives, the feature matching engine automatically parses its strategy feature vector and calculates the cosine similarity between the feature vector of the strategy unit to be deployed and the feature vector of the historical base model.

[0128] ;

[0129] In the formula, The feature vector representing the strategy unit to be deployed; The feature vector representing the historical basis model; and They represent and The modulus length is used for normalization. ∈[0,1] represents the similarity value; a larger value indicates greater feature similarity. The feature vectors of the historical basis model are automatically generated using feature vectorization encoding technology. When the policy unit runs online, the system extracts its key policy features, including weight coefficients and decision preference parameters, and transforms them into fixed-dimensional feature vectors through an embedding layer. This feature vector contains core information such as strategy weight configuration, category preference, and parameter sensitivity. Ultimately, it serves as a unique identifier, bound to the model version, and stored in the model repository, forming a searchable model feature mapping relationship. The system prioritizes selecting the optimal historical baseline model with a similarity greater than 0.8 as the matching model.

[0130] The update process is dynamically driven by the feature offset, which is obtained by subtracting the cosine similarity from 1. These three features correspond to three update methods, and each of them corresponds to the fine-tuning signal, the reconstruction signal, and the drastic environmental change conditions output in step S2.

[0131] The first approach involves parameter fine-tuning, specifically for cases where the fine-tuning signal corresponds to a feature offset less than 0.7. When the features of the strategy unit to be deployed have a high similarity to the historical base model, the system initiates parameter-level fine-tuning, employing online knowledge distillation technology to fuse the old and new decision-making logics. Specifically, a matching model validated through A / B testing is used as the teacher model, and a student model carrying the newly added strategy features of the strategy unit to be deployed is used as the student model. During distillation, the validated decision-making capabilities of the teacher model are retained, while gradient signals related to the newly added features in the strategy unit to be deployed are absorbed. The weights of the student model are then optimized using the loss function. The expression for the core loss function is:

[0132] ;

[0133] In the formula, This represents the total loss from knowledge distillation; Represents the task loss function; Indicates the output distribution of the teacher model With student model output distribution KL divergence between them; Indicates temperature parameter; , Here are the weight parameters, where and ;

[0134] In the embodiments of the present invention , We use values ​​of 0.7 and 0.3 respectively to balance the weights of task performance and knowledge transfer. This fine-tuning method is suitable for scenarios with relatively stable market environments, such as weight adjustments caused by minor seasonal fluctuations, and can achieve precise optimization while maintaining the overall stability of the model architecture.

[0135] The second approach is structural reconstruction, corresponding to the reconstructed signal and a feature offset greater than or equal to 0.7. When the feature offset exceeds the threshold of 0.7, it indicates a significant difference between the strategy unit to be deployed and the historical base model, requiring structural evolution. The system dynamically reconstructs neural network branches, making targeted adjustments to regions of feature mutation. Specifically, this includes at least one of the following: adding an attention mechanism to enhance the ability to capture key features, adjusting the dimensions of fully connected layers to adapt to the new feature space, or introducing residual connections to improve gradient flow. This evolutionary mechanism employs a gradual update strategy, reconstructing only the affected branches while maintaining the stability of the main structure, avoiding the resource consumption of retraining the entire model. This mechanism is particularly suitable for scenarios where the market environment undergoes moderate changes, such as changes in feature distribution caused by the emergence of new trends.

[0136] The model matching process consists of two stages: The first stage is the candidate selection stage, where the system selects candidate models from the model repository whose feature vector cosine similarity to the policy unit to be deployed is greater than a preset similarity threshold, and selects the model with the highest similarity. When the similarity of all candidate models is lower than the preset similarity threshold, the policy unit to be deployed is registered as a new base model. The second stage is the update method determination stage. After successful matching, the feature offset is calculated. When the offset is less than 0.7, parameter fine-tuning is performed; when the offset is greater than or equal to 0.7, structural reconstruction is performed. It should be noted that when the preset similarity threshold is 0.8, the offset of a successfully matched model will necessarily be less than or equal to 0.2, primarily triggering parameter fine-tuning. Reconstruction signals are more likely to appear in scenarios where the preset similarity threshold is set low or the system downgrades the matching.

[0137] The third approach is branch expansion, which addresses drastic environmental changes. When facing significant market shifts or entirely new business scenarios, the system activates a branch expansion mechanism to build lightweight, dedicated sub-networks for specific scenarios. It extracts general features from the matching model through knowledge transfer to quickly adapt to new environments and uses a gating mechanism to dynamically activate the dedicated sub-networks to work collaboratively with the main network. These three mechanisms together constitute a complete self-evolving system: parameter fine-tuning ensures the system's sensitive response to subtle changes, structural reconstruction handles moderate environmental changes, and branch expansion addresses major market shifts. The gating mechanism is implemented through a trainable gating network that takes the current input features as input and outputs an activation weight between 0 and 1. When the activation weight is higher than a preset threshold, the dedicated sub-network is activated and processes the input data in parallel with the main network, ultimately outputting a weighted fusion of the main network's output and the dedicated sub-network's output. When the activation weight is lower than the preset threshold, only the main network operates independently. The parameters of the gating network are optimized through end-to-end training, enabling the system to automatically learn when to activate the dedicated sub-network.

[0138] The updated model is stored in the model version repository and deployed to the product selection decision-making execution stage. Through the reuse of historical models and lightweight updates, a self-driven closed loop is ultimately formed, which optimizes strategies, accumulates model capabilities, and generates new strategies, thus completely eliminating the need for manual intervention.

[0139] In summary, the self-learning method for apparel e-commerce product selection systems based on a closed-loop feedback mechanism provided in this invention achieves fully automated operation across the entire chain, from data monitoring and effect evaluation to strategy iteration and model updates, through a closed-loop self-evolutionary architecture. This effectively solves the technical problems of existing product selection systems, such as lag in feedback, lack of self-evolutionary capability, and high reliance on manual intervention. It provides an efficient and robust technical solution for intelligent product selection in apparel e-commerce. The above are merely preferred embodiments of this invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the protection scope of this invention.

[0140] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0141] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A self-learning method for a clothing e-commerce product selection system based on a closed-loop feedback mechanism, characterized by, Includes the following steps: Step S1: Collect and aggregate user behavior data, business operation data, and product performance data from the apparel e-commerce product selection system in real time to obtain structured indicators; Step S2: Decompose the product selection decision into multiple strategy units, associate a set of structured indicators with each strategy unit, determine the dynamic weights of the structured indicators through the entropy weight method, and obtain a comprehensive evaluation score by weighted summation, and generate fine-tuning signals, reconstruction signals or elimination signals based on the comprehensive evaluation score; Step S3: In response to the elimination signal, the linear combination of the comprehensive evaluation score of the strategy unit in the current period, the average score of the historical periods, and the standard deviation of the score fluctuation is used as the strategy effectiveness index. When the strategy effectiveness index is lower than the industry benchmark value for a consecutive preset number of periods, the corresponding strategy unit in the strategy pool is eliminated, and the strategy features are extracted from the surviving strategy units and cross-combined to generate a new strategy unit to be added to the strategy pool. In response to fine-tuning or reconstruction signals, retrieve historical base models from the model version repository, select the matching model with the highest cosine similarity to the feature vector of the strategy unit to be deployed, fine-tune the parameters or reconstruct the structure of the matching model according to the signal type, update it, store it in the repository, and deploy it for execution. 2.The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 1, characterized in that: The user behavior data mentioned in step S1 includes click-through rate, conversion rate, and return reason analysis results; the business operation data includes inventory turnover rate, gross profit margin, and slow-moving goods rate; and the product performance data includes new product identification accuracy and trend response delay. The real-time acquisition and aggregation includes: stamping a unified timestamp as a time series benchmark when the original event is generated by the event collector; aggregating high-frequency data at the second and minute levels in real time through a stream processing engine; performing micro-batch processing on hourly data; and calculating low-frequency data at daily and weekly levels according to a fixed period to obtain the structured indicators. 3.The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 2, characterized in that: The results of the return reason analysis were obtained through the following steps: Data cleaning and text normalization are performed on the original return text submitted by users. Key entities are extracted from the original return text using matching rules based on the business thesaurus and named entity recognition model. The text after key entities are extracted is classified using a pre-trained text classification model and a business keyword library. When the classification result of the text classification model matches the matching result of the business keyword library, a high confidence level is assigned and the classification result is adopted. When the classification result of the text classification model conflicts with the matching result of the business keyword library, the classification result of the text classification model is adopted first and the confidence level is reduced. In each calculation period, the classification results with confidence levels higher than a preset threshold are aggregated, and the proportion of each reason for return within the calculation period is calculated.

4. The self-learning method for a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 1, characterized in that: Step S2 involves determining the dynamic weights of the structured indicators using the entropy weight method, and then weighting and summing them to obtain the comprehensive evaluation score of the strategy unit. This includes the following steps: Using the extreme value method to the first Item structured indicators in Observations on a sample Standardize to obtain Calculate the first Information entropy of structured indicators : ； ； ； wherein denotes the total number of samples; j denotes the sequence number of the jth sample, ; The basic weights are determined based on the information entropy: ； The structured indicators are weighted and summed based on the aforementioned basic weights to obtain the comprehensive evaluation score of the strategy unit: ； ； In the formula, Indicates the first The overall evaluation score of each strategy unit; Indicates the relationship with the first The total number of structured metrics associated with each strategy unit; Indicates the first The dynamic weights of the structured indicators after being processed by the confidence decay factor; Indicates the confidence decay factor; Indicates the first The basic weights of the structured indicators are calculated using the entropy weighting method. This represents the normalized index value; This represents the index conversion function.

5. The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 4, characterized in that: Indicator Conversion Function This includes at least one of the following: a logarithmic transformation function for exponential growth characteristic indicators, a linear normalization function for percentage indicators, a sigmoid compression function for suppressing abnormal fluctuations, and a piecewise decay function for distinguishing between target values ​​and out-of-target values.

6. The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 1, characterized in that: Step S2 generates a Pareto solution set based on the comprehensive evaluation score of each strategy unit, and then performs a weighted score and sorting of the Pareto solution set using the business objective weight matrix to select the champion strategy. A / B testing is performed on the champion strategy and the current online strategy. The significance level of the performance improvement of the champion strategy is determined by t-test. If the p-value is less than 0.05, the champion strategy is fully deployed; otherwise, the strategy is rolled back to the current online strategy.

7. The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 1, characterized in that: The expression for calculating the strategy effectiveness index in step S3 is as follows: ； In the formula, Indicates the first The strategy effectiveness index of each strategy unit; Indicates the first The overall evaluation score of each strategy unit in the current evaluation period; Indicates the first The comprehensive evaluation score of each strategy unit in the t-th historical period; Indicates the total number of historical cycles; Indicates the first The standard deviation of the volatility of the overall evaluation score of each strategy unit; , , Indicates the weighting coefficient; The industry benchmark value is obtained by sorting the strategy effectiveness indices of all strategy units in the strategy pool and taking the average of the strategy effectiveness indices of the top 30% of strategy units.

8. The self-learning method for a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 7, characterized in that: The corresponding policy units in the elimination policy pool in step S3 also include: Determine whether the eliminated strategy unit in the strategy pool has a dependent strategy. If a dependent strategy exists, initiate strategy reorganization. If no dependent strategy exists, release the weight resources occupied by the eliminated strategy unit and allocate the released weight resources to other strategy units in the strategy pool with similar functions and located at the Pareto front. Extracting policy features from the policy pool and combining them to generate new policy units to supplement the policy pool includes: selecting at least two surviving policy units with the highest policy effectiveness index from the policy pool as parent policies; extracting the policy feature vectors of each parent policy; exchanging segments according to random crossover points to generate a mixed feature vector; applying random perturbation to some dimensions of the mixed feature vector with a preset mutation probability; and initializing new candidate policy units based on the perturbated mixed feature vector to supplement the policy pool. 9.The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 1, characterized in that: The formula for calculating the cosine similarity of feature vectors in step S3 is as follows: ； In the formula, Let represent the cosine similarity of the feature vectors, and sim∈[0,1]; The feature vector representing the strategy unit to be deployed; The feature vectors representing the historical basis model; and They represent and The modulus length; The feature vectors of the strategy unit to be deployed and the historical base model are both generated by encoding the weight coefficients and decision preference parameters of the associated strategy unit through an embedding layer; the matching model is the historical base model with the highest cosine similarity to the feature vector of the strategy unit to be deployed that is greater than a preset similarity threshold; when the cosine similarity of the feature vectors of all historical base models is lower than the preset similarity threshold, the strategy unit to be deployed is stored as a new base model in the model version repository.

10. The self-learning method of a clothing e-commerce product selection system based on a closed-loop feedback mechanism according to claim 9, characterized in that: Step S3 updates the matching model based on the fine-tuning signal and the reconstructed signal, including the following steps: The feature offset is calculated by subtracting the cosine similarity of the feature vector from 1. When the feature offset corresponding to the fine-tuning signal is less than 0.7, online knowledge distillation technology is used to fine-tune the parameters of the matching model based on a loss function. The expression for the loss function is as follows: ； In the formula, This represents the total loss from knowledge distillation; Represents the task loss function; Indicates the output distribution of the teacher model With student model output distribution KL divergence between them; Indicates temperature parameter; , Here are the weight parameters, where and ; When the feature offset corresponding to the reconstructed signal is greater than or equal to 0.7, at least one of the following is performed on the matching model: attention mechanism addition, full-connected layer dimension adjustment, and residual connection introduction to perform structural reconstruction; When a dramatic change in the market environment is detected, a dedicated sub-network is constructed by extracting general features from the matching model through knowledge transfer, and the dedicated sub-network is dynamically activated to work in collaboration with the main network of the matching model through a gating mechanism.

Images