An e-commerce order recovery method and system based on artificial intelligence and a storage medium

Abstract

The application discloses an e-commerce order recovery method and system based on artificial intelligence, and a storage medium, and belongs to the technical field of artificial intelligence and data processing. The method triggers the order recovery process in response to a return and refund application event: multi-dimensional features such as user behavior, order, refund reason, user portrait, commodity life cycle, logistics time limit and effect, and social communication are collected, standardized and coded to splice a unified feature vector; a pre-trained machine learning model is input, and an order recovery probability and a malicious refund identification probability are output. If the malicious probability is above the threshold, the order recovery is rejected and risk control is triggered; otherwise, based on the order recovery probability and the interest strategy pool, the optimal strategy is selected through a comprehensive benefit evaluation function, and the execution is performed through a pop-up window, an external call, a private domain or an artificial method. The execution callback feedback is received, and the model is continuously updated. The application realizes intelligent and automatic order recovery decision, improves the order recovery success rate, reduces the refund rate, accurately identifies malicious refunds, and balances short-term benefits and long-term user value.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and e-commerce, specifically to the field of data processing technology, and in particular to an artificial intelligence-based method, system, and storage medium for managing e-commerce orders. Background Technology

[0002] With the rapid development of the e-commerce industry, online shopping has become an important part of consumers' daily lives. However, the problem of reverse logistics (returns and refunds) in e-commerce is becoming increasingly prominent. Industry data shows that the annual growth rate of e-commerce reverse shipments has reached 22.7%, bringing huge operational pressure and profit losses to merchants.

[0003] In current e-commerce operations, returns and refunds present multi-dimensional complexities. From a consumer behavior perspective, the reasons for returns have expanded from traditional product quality issues to include price-sensitive returns (such as returns after price reductions during major sales events), impulse purchase returns (such as returns related to live-streaming e-commerce), and returns due to discrepancies in experience (such as size or color differences). From a merchant's operational perspective, returns not only lead to direct sales losses but also generate hidden costs such as reverse logistics costs, quality inspection costs, restocking costs, and inventory losses. Industry estimates suggest that the total cost of a single returned order is approximately 15% to 30% of the product's selling price.

[0004] Current return and refund processing methods primarily rely on human customer service intervention, which presents the following problems: First, slow response time; consumers often have to wait a long time to receive a solution after initiating a refund, leading to a decline in user experience. Second, order recovery strategies lack data support; customer service personnel typically rely on experience to determine whether to recover an order and what strategy to use, making it difficult to achieve accurate order recovery. Third, it cannot effectively identify malicious refund behavior (such as fraudulently obtaining shipping insurance, maliciously refunding only, bulk refund fraud, etc.), resulting in damage to merchants' rights. Fourth, order recovery costs are difficult to control, lacking a scientific cost-benefit assessment mechanism. Fifth, it lacks consideration for the long-term value of users, focusing excessively on recovering single orders while neglecting user lifetime value management. Sixth, it does not fully utilize private domain traffic and social media data for order recovery decisions.

[0005] Therefore, there is an urgent need for a systematic solution that can automatically assess the value of order recovery, intelligently recommend the best recovery strategy, accurately identify malicious refund behavior, and take into account user lifecycle value management. Summary of the Invention

[0006] This invention aims to solve the technical problems in existing e-commerce technologies, such as low efficiency in handling returns and refunds, inaccurate order recovery strategies, weak ability to identify malicious refunds, lack of user lifetime value assessment, and weak private domain operation capabilities.

[0007] This invention provides an artificial intelligence-based method for recovering e-commerce orders, comprising the following steps:

[0008] S1: In response to the detection of a consumer's return and refund request for an order on the e-commerce platform, the order recovery process is triggered;

[0009] S2: Collect multi-dimensional feature data, which includes at least user behavior features, order features, refund reason features, user profile features, product lifecycle features, logistics timeliness features, and social dissemination features. Standardize and encode the feature data of each dimension and splice them together to form a feature vector of a unified dimension.

[0010] S3: Input the feature vector into a pre-trained machine learning model, and the machine learning model outputs the probability of winning the sale and the probability of identifying malicious refunds;

[0011] S4: Determine whether the probability of identifying malicious refunds exceeds a preset threshold;

[0012] If the limit is exceeded, the attempt to recover the order will be rejected and an account association risk control investigation will be triggered.

[0013] If the probability of closing the deal is not exceeded, the benefit evaluation score of each candidate strategy is calculated by using a comprehensive benefit evaluation function based on the aforementioned probability of closing the deal and the preset equity strategy pool. The strategy with the highest benefit evaluation score is selected as the optimal strategy for closing the deal.

[0014] S5: Execute the optimal order recovery strategy through pop-up interaction, intelligent outbound calling, private domain outreach, or manual intervention;

[0015] S6: Receive the callback feedback of the order recovery execution result, and use the callback feedback as a training sample to update the machine learning model.

[0016] Furthermore, in step S2:

[0017] The user behavior characteristics include: browsing depth, time to add to cart, number of times to browse products, comparison behavior, collection behavior, customer service consultation records, order time, payment method, sharing behavior, and live broadcast interaction behavior;

[0018] The order characteristics include: actual payment amount, product category, product gross profit margin, discount rate, whether it is a limited-time special offer order, whether it is a live broadcast order, the time from order placement to shipment, whether it is the user's first order, order source channel, and number of related orders;

[0019] The product lifecycle characteristics include: product shelf life, inventory turnover rate, seasonality index, competitor price fluctuation range, and product sentiment score.

[0020] The logistics timeliness characteristics include: delivery timeliness, transit time, remoteness of the delivery area, number of logistics anomalies, and deviation in the signing time.

[0021] Furthermore, step S2 also includes the collection of multimodal sentiment analysis features: the multimodal sentiment analysis features include: the sentiment tags and sentiment intensity scores obtained by encoding the refund message text using the BERT model, the voice sentiment vector constructed after the voice customer service call content is transcribed by ASR and the acoustic features are extracted, and the video conversation sentiment time-series curve generated after the video customer service facial expressions are recognized by the CNN model;

[0022] The emotional labels are divided into five categories: “anger,” “disappointment,” “hesitation,” “calm,” and “satisfaction,” and the emotional intensity score ranges from 0 to 1. The acoustic features include speech rate, volume, pitch variation, and pause frequency. The CNN model is trained on the FER2013 dataset and recognizes seven expressions: anger, disgust, fear, happiness, sadness, surprise, and neutral.

[0023] Furthermore, the multimodal sentiment analysis features are dynamically weighted and fused with the sentiment scores of each modality through an attention mechanism to output a comprehensive sentiment state feature, which is used to correct the probability of winning the sale.

[0024] Furthermore, step S2 also includes collecting product image quality assessment features:

[0025] The product image quality assessment features include: damage score of returned product images output by the ResNet-50 model, heat map of damaged area location, and image shooting quality assessment results;

[0026] The damage rating range is from 0 to 1, which is used to cross-validate logistics anomaly records and assist in calculating partial refund amounts.

[0027] Furthermore, step S2 also includes the collection of cross-platform user identity fusion features and time-series behavioral sequence features:

[0028] The cross-platform user identity fusion features include: cross-platform purchase frequency, cross-platform return rate, cross-platform average order value comparison, cross-platform category preference consistency, and cross-platform coupon sensitivity; the cross-platform user identity is identified through device fingerprint technology and behavioral pattern matching.

[0029] The temporal behavioral sequence features include: the trend of product page views within a preset time period before refund, the behavior of deleting items from the shopping cart, the frequency of comparing prices of similar products, the number of times customer service inquires about refund policies, the change in the dwell time on the order details page, and the ignore rate of App push messages. The temporal behavioral sequence features are statistically calculated using a sliding window to determine the behavioral mutation index.

[0030] Furthermore, in step S3:

[0031] The machine learning model adopts an integrated architecture of gradient boosting tree model and deep neural network model;

[0032] The gradient boosting tree model processes the structured numerical features in the feature vector and outputs a first predicted probability.

[0033] The deep neural network model processes the high-dimensional sparse features in the feature vector and outputs a second prediction probability; the first prediction probability and the second prediction probability are weighted and fused, and then mapped to the order recovery probability through the Sigmoid activation function.

[0034] Furthermore, the structure of the deep neural network model includes:

[0035] The input layer receives the feature vector;

[0036] The Embedding layer maps sparse class features into dense vectors;

[0037] Layer normalization layer, which performs layer normalization on the Embedding output;

[0038] A bidirectional LSTM layer with 64 hidden units each for the forward and backward directions;

[0039] A multi-head attention layer with 8 heads performs self-attention weighting on the bidirectional LSTM output;

[0040] The fully connected layer has 128 neurons and uses the GELU activation function.

[0041] The output layer consists of a single neuron and uses the Sigmoid activation function to output the second predicted probability.

[0042] Furthermore, step S3 also includes: The probability of identifying malicious refunds is calculated by an independent risk control model, which integrates a rule engine and an anomaly detection algorithm. The rule engine outputs binary decision results based on preset hard rules; The anomaly detection algorithm uses the isolated forest algorithm to output anomaly scores; The probability of identifying malicious refunds is calculated using a fusion formula: , in, The binary decision result of the rule engine. Let λ be the abnormal score, and λ be the abnormal score weighting coefficient.

[0043] Furthermore, step S3 also includes calculating the policy sensitivity probability: The strategy sensitivity probability is calculated based on uplift modeling in the causal inference framework. The average conditional treatment effect of each user is estimated by CausalForest, and users are divided into four groups: users who are bound to win orders, users who are bound to churn, strategy-sensitive users, and strategy-averse users.

[0044] Furthermore, step S3 also includes knowledge graph-enhanced order recovery probability correction: Construct a knowledge graph that includes product nodes, user nodes, and refund reason nodes; Message passing is performed on the knowledge graph using a graph neural network to learn a high-order interactive representation of user-product-refund reason; When a high-weighted association path in the knowledge graph is detected between a user's historical refund reasons and the current order items, the order recovery probability is adjusted according to the weight of the association path.

[0045] Furthermore, in the message passing mechanism of the graph neural network, the hidden state update formula for the k-th layer node v is: , in Let v be the hidden representation of node v at the k-th layer. Let v be the set of neighboring nodes. σ is a learnable aggregate weight matrix, σ is the activation function, and k is the preset message passing layer.

[0046] Furthermore, the comprehensive benefit evaluation function in step S4 is: Benefit evaluation score = (Probability of closing the deal × Customer lifetime value coefficient × Customer acceptance willingness coefficient × Deal closing amount) / (Cost of closing the deal + Risk cost + Customer churn cost); The user lifetime value coefficient is calculated based on the user's historical spending amount, repurchase frequency, and referral value; the user acceptance willingness coefficient is the historical acceptance rate statistics of each strategy; the risk cost is the product of the probability of malicious refund identification and the potential loss; and the user churn cost is the expected future revenue loss caused by the failure to recover the order.

[0047] Furthermore, step S4 also includes Stackelberg game-based dynamic pricing: The decision-making process of merchants and users is modeled as a dynamic game, with merchants as leaders deciding the type and amount of the sales recovery strategy, and users as followers deciding whether to accept the sales recovery based on the merchant's strategy. Solving the subgame perfect Nash equilibrium using backward induction; The dynamic pricing engine adjusts the strategy amount based on the user's real-time response.

[0048] Furthermore, the mathematical modeling of the Stackelberg game is as follows: the payoff function of the merchant as the leader is... , in x is the actual amount paid for the order, and x is the amount of compensation provided by the merchant to salvage the order. The reputation loss coefficient, This is an indicator function for whether a product has quality problems. The utility function of a user as a follower is: , in To retain users' subjective value assessment of the product, The amount that can be refunded. Increase in psychological satisfaction; Nash equilibrium solution process: First, solve for the user's optimal response. This refers to the minimum compensation amount a user is willing to accept for salvaging the deal. Then the merchant is under control. Maximizing the revenue function of merchants under the constraint of cost ceiling ,in As the upper limit of merchant costs, when If the merchant is unable to provide compensation that meets the user's minimum requirements, the order is deemed irrecoverable and will proceed directly to the refund process.

[0049] Furthermore, step S4 also includes deep reinforcement learning strategy optimization: The recommended strategy for closing deals is modeled as a Markov decision process. The state space of the Markov decision process includes user feature vector, order feature vector, current order recovery round, historical strategy response record and time decay factor; The action space includes discrete strategy type selection and continuous strategy amount; The reward function takes into account both immediate and long-term rewards. A proximal policy optimization algorithm is used to train the policy network to maximize long-term cumulative returns.

[0050] Furthermore, the equity strategy pool in step S4 includes at least two of the following strategy types: Coupons, product vouchers, return shipping fees, price difference refunds, exchanges, gifts or replacements, partial refunds, full refunds without returns, intelligent outbound calling, private community outreach, member points compensation, and personalized recommendation alternatives.

[0051] Furthermore, the triggering conditions for intelligent outbound calling in step S5 are as follows: when the probability of closing the deal is greater than a preset threshold, the actual payment amount of the order is greater than a preset amount threshold, and the user's lifetime value level is higher than a preset level, intelligent voice outbound calling is automatically triggered, and a deal-closing plan is pushed to the user before the outbound call.

[0052] Furthermore, in step S3, the machine learning model also outputs a refund risk warning probability: The refund risk warning probability is output based on the sequence modeling of the temporal behavioral sequence features using a bidirectional LSTM network. Step S5 also includes pre-emptive warning intervention: when the refund risk warning probability exceeds a preset threshold, the warning intervention is proactively triggered before the user formally initiates a refund application; The early warning and intervention methods include: proactively pushing care messages, automatically triggering dedicated customer service follow-ups, issuing targeted coupons or gifts in advance, and optimizing logistics and delivery plans; The early warning intervention adopts a tiered response mechanism, triggering different levels of early warning measures based on the numerical range of the refund risk warning probability.

[0053] Furthermore, in step S2: Numerical features are normalized using the Z-score standardization method, and missing values ​​are filled using the median imputation strategy. In categorical features, those with a cardinality less than a preset value are encoded using One-Hot encoding, while those with a cardinality greater than or equal to a preset value are encoded using Target Encoding. The refund reason text is encoded into a semantic vector using a pre-trained Chinese BERT model; Product titles are used to extract n-gram features using the TextCNN model; User behavior sequences are encoded using Position Encoding and Transformer Encoder.

[0054] Furthermore, the feature acquisition in step S2 adopts a dual-channel architecture of real-time stream processing and batch processing: The real-time channel extracts real-time features from real-time data sources based on a stream processing engine, with processing latency controlled within 200ms; The batch processing channel pre-calculates user profile features and historical statistical features periodically based on the offline computing engine, and stores them in the feature storage service for real-time querying; The dual-channel features are merged in the feature splicing layer to form the feature vector.

[0055] Furthermore, in step S6: The callback feedback is received via a RESTful API through a callback interface, and the request body is in JSON format. The callback interface adopts a timeout retry strategy: after the first callback times out, the exponential backoff algorithm is used for retry, with retry intervals of 1 second, 2 seconds, and 4 seconds, and a maximum number of retries of 3. The callback interface uses the combination of the order's unique identifier and the callback trigger type as an idempotent key.

[0056] Furthermore, step S6 also includes an A / B testing framework: The A / B testing framework randomly assigns different strategies to similar user groups; By comparing the success rate of order recovery, user satisfaction, short-term repurchase rate and long-term repurchase rate of each group, the strategy recommendation model is continuously optimized. The proportional indicators were statistically significant using the chi-square test, and these proportional indicators included the success rate of order recovery. The continuous indicators, including order amount and user satisfaction score, were tested for statistical significance using an independent samples t-test.

[0057] Furthermore, step S6 also includes federated learning optimization: Each participating merchant trains its local model on the premise that its local data does not leave the domain; Each merchant uploads encrypted gradients to the central coordination server; The central coordination server aggregates the local model weights of each merchant through a federated averaging algorithm to generate a global model. The aggregation formula for the federated average algorithm is: , in The global model weights for round t+1 are... Let K be the local model weights for merchant k after training round t. Let K be the local training sample size for merchant k. K represents the total sample size of all merchants, and K represents the total number of merchants participating in federated learning. Differential privacy techniques are used to add calibration noise to the gradient.

[0058] This invention also provides an artificial intelligence-based e-commerce order recovery system, comprising: The trigger module is used to respond to the detection of a user's return and refund request for an order on the e-commerce platform, and to trigger the order recovery process. The feature acquisition module is used to collect multi-dimensional feature data in response to the detection of a return and refund application event. The multi-dimensional feature data includes at least user behavior features, order features, refund reason features, user profile features, product lifecycle features, logistics timeliness features, and social dissemination features. The feature data of each dimension is standardized and encoded and then concatenated to form a feature vector. The probability evaluation module is used to input the feature vector into a pre-trained machine learning model to calculate the probability of winning the sale and the probability of identifying malicious refunds. The strategy recommendation module is used to determine whether the probability of malicious refund identification exceeds a preset threshold. If it exceeds the threshold, the order recovery is rejected and an account association risk control investigation is triggered. If it does not exceed the threshold, the efficiency evaluation score of each candidate strategy is calculated through a comprehensive efficiency evaluation function based on the order recovery probability and the preset equity strategy pool. The strategy with the highest efficiency evaluation score is selected as the optimal order recovery strategy. The execution module is used to execute the optimal order-saving strategy through pop-up interaction, intelligent outbound calling, private domain outreach, or manual intervention. The feedback optimization module is used to receive callback feedback of the order recovery execution result and use the callback feedback as training samples to update the machine learning model.

[0059] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0060] The present invention also provides a computer program product, including a computer program or computer instructions, which, when executed by a processor, implement the method described in any of the preceding claims.

[0061] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any of the preceding claims.

[0062] Beneficial effects Compared with the prior art, the present invention has the following advantages: First, it achieves intelligent and automated decision-making for closing deals. By comprehensively evaluating seven categories of features through a multi-layered machine learning model, it replaces traditional human experience-based judgment, increasing the success rate of closing deals by approximately 21%, and the return on investment for intelligent deal closing is about twice that of manual configuration. The system response time has been reduced from the average 15 minutes of human customer service to milliseconds, significantly improving the user experience.

[0063] Secondly, a scientific cost-benefit assessment mechanism has been established. The comprehensive benefit assessment function not only considers the probability and cost of retaining a single order, but also incorporates the user lifetime value coefficient and user churn cost, ensuring that order retention decisions balance short-term gains with long-term value. Industry data shows that every 1 yuan cost of retaining an order can save approximately 100 yuan of actual payment, while the retained user contributes an average of 35 yuan in repurchase amount over the next 6 months.

[0064] Third, it accurately identifies malicious refund behavior. An independent risk control model integrates a rule engine and anomaly detection algorithms to effectively identify malicious behavior patterns such as fraudulent shipping insurance claims, malicious refunds without due cause, bulk refund fraud, false damage claims, and repeatedly obtaining new goods by exploiting the no-reason return policy. The risk control interception rate is approximately 8% to 12%, accurately excluding malicious refund requests from about 30% of recoverable orders, saving medium-sized merchants tens to hundreds of thousands of yuan in losses annually.

[0065] Fourth, it supports multiple execution methods and flexible rights and benefits strategy configurations. Four execution methods (pop-up, outbound call, private domain, and manual) cover different scenarios and user groups, while 12 strategy types meet diverse needs. Merchants can configure a rights and benefits strategy pool and cost cap according to their own business characteristics. The system automatically selects the optimal solution within the strategy pool, balancing sales retention and cost control.

[0066] Fifth, the system introduces user lifetime value (LTV) assessment to upgrade from "order-level lead recovery" to "user-level lead recovery." For users with high LTV, even if the ROI of a single lead recovery is slightly lower, the system may still recommend lead recovery to maintain the long-term relationship; for users with low LTV, the system strictly controls lead recovery costs to avoid wasting resources.

[0067] Sixth, fully leverage private domain traffic and social media data to expand lead generation channels. Private domain outreach costs are only 1 / 10 to 1 / 5 of outbound call costs, and user acceptance rates are typically higher than with pop-up ads. Social media reach helps identify high-influence users, allowing for better lead generation conditions in exchange for positive word-of-mouth.

[0068] Seventh, continuous optimization is achieved through an A / B testing framework. The system does not rely on one-time model training, but rather continuously experiments to compare the effects of different strategies, automatically iterating and optimizing the recommendation algorithm to ensure that the order recovery strategy always remains in sync with the market environment and user preferences.

[0069] Eighth, multimodal sentiment analysis significantly improves the accuracy of sales recovery. By integrating refund message text, the acoustic features of voice customer service representatives, and the facial expressions of video customer service representatives, the system can comprehensively perceive the user's emotional state and dynamically adjust the sales recovery strategy. For users with extremely negative emotional states, the system prioritizes recommending high-compensation strategies or transferring the case to human intervention, avoiding automated sales recovery that could exacerbate user dissatisfaction. Multimodal sentiment fusion increases the personalization of sales recovery strategies by approximately 30% and user satisfaction by approximately 15%.

[0070] Ninth, the uplift modeling framework enables precise strategy targeting. By differentiating between four groups—"inevitable customers," "inevitable churn customers," "strategy-sensitive customers," and "strategy-averse customers"—the system avoids excessive incentives for eligible customers (saving 25% to 35% in costs), avoids actively bothering strategy-averse customers (reducing complaint rates by approximately 40%), and concentrates limited resources on strategy-sensitive customers (improving resource utilization efficiency by approximately 50%).

[0071] Tenth, knowledge graphs enhance the discovery of deep relationships. By constructing a product-user-refund reason knowledge graph and utilizing GNNs for message passing, the system can discover high-order interaction patterns that are difficult to capture with traditional feature engineering (such as "the return rate of a certain size of a certain brand is abnormally high"), thereby allowing for early adjustment of order recovery strategies for corresponding products or promoting product improvements. Knowledge graphs improve the AUC of order recovery probability prediction by approximately 3% to 5%.

[0072] Eleventh, Stackelberg's game-theoretic dynamic pricing strategy optimizes pricing. By modeling the decision-making process of merchants and users as a dynamic game, the system can adjust its pricing strategy based on real-time user responses, avoiding a fixed, one-size-fits-all approach. Game-theoretic dynamic pricing increases the success rate of closing deals by approximately 12% to 18%, while keeping cost overruns below 3%.

[0073] Twelfth, deep reinforcement learning (PPO) maximizes long-term returns. Compared to greedy strategies that only optimize single order recovery, PPO selects strategies by estimating long-term value (user repeat purchases, word-of-mouth marketing), achieving approximately 22% improvement in long-term return metrics over 180 days. Reinforcement learning is particularly suitable for multi-round order recovery scenarios, as it can learn the optimal combination of timing, channels, and amounts for order recovery.

[0074] Thirteenth, the timing of re-initiating after-sales service is moved forward to before the refund request. By predicting the probability of refund risk and triggering tiered warnings through a Bi-LSTM model, the system can proactively intervene when a user intends to request a refund but has not yet taken action, reducing the refund initiation rate by approximately 18% to 25%, and the success rate of re-initiating the refund after warning is 30% to 40% higher than that of re-initiating the refund after the refund.

[0075] Fourteenth, federated learning architecture empowers small and medium-sized businesses. By enabling multi-business joint modeling with data remaining within the same domain, small and medium-sized businesses can quickly acquire high-quality order retention capabilities by leveraging the platform and shared models with other businesses, solving the problems of cold start and insufficient data. Federated learning improves the order retention success rate of small and medium-sized businesses with a sample size of less than 1,000 by approximately 15% to 20%.

[0076] Fifteenth, cross-platform user identity fusion solves the data silo problem. By identifying the same user across different platforms through device fingerprinting and behavioral pattern matching, the system can comprehensively evaluate a user's cross-platform consumption behavior and credit level, avoiding misjudgments caused by data bias from a single platform. Cross-platform features increase the success rate of retaining new users by approximately 10% to 15%.

[0077] Sixteenth, product image quality assessment improves the efficiency of processing refunds for damaged goods. By automatically identifying the extent and area of ​​damage to returned goods through computer vision, the system can quickly determine the reasonableness of the refund and automatically calculate part of the refund amount, reducing manual review costs by approximately 60% to 70%, while also reducing the success rate of malicious damage claims by approximately 25%. Attached Figure Description

[0078] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0079] Figure 1 is a schematic diagram of the process of the e-commerce order recovery method based on artificial intelligence of the present invention; Figure 2 is a schematic diagram of the module structure of the AI-based e-commerce order recovery system of the present invention. Detailed Implementation

[0080] This invention provides an artificial intelligence-based method for recovering e-commerce orders, comprising the following steps: Step S1 (Triggering Phase): When the e-commerce platform detects that a consumer has initiated a return and refund request for an order on the platform, the order recovery process is automatically triggered. The system obtains basic order information, the reason for the refund, the list of rights and benefits policies configured by the merchant, and the user's historical interaction data.

[0081] Technical Implementation of Refund Request Event Listening: This invention employs a real-time event listening mechanism based on a message queue. A consumer group subscribes to the "refund_apply" topic. When the e-commerce platform's refund service module generates a return and refund request event, it publishes the event to the message queue in Avro serialization format. The consumer group achieves horizontal scaling through a partitioning allocation strategy. Each consumer instance is responsible for processing messages in a specific partition, ensuring that events from the same user are processed sequentially by the same consumer. The message queue retention time is set to 24 hours, supporting message retries in case of consumption failure. Event consumption latency is controlled within 500ms, and the total latency from event generation to consumer reception and processing does not exceed 500ms.

[0082] The decision tree logic for triggering condition judgments: After receiving a refund request, an e-commerce platform executes the following three levels of judgments sequentially. Level 1: Determine if the refund type is "return and refund" (excluding "refund only"). If it's "refund only," skip the order recovery process, as this usually means the user no longer wants to keep the item. Level 2: Determine if it's the first order and the order amount is below a preset threshold (e.g., 50 yuan). If so, skip the order recovery process to avoid excessive investment in recovering low-value new customers. Level 3: Determine if the merchant has enabled and configured the order recovery function. If not, skip this step. After all three levels of judgments pass, the system officially enters the order recovery process. The total time for triggering judgments is controlled within 100ms to ensure it doesn't affect the response time of the main refund process.

[0083] Trigger Delay Control Parameters: To ensure the real-time performance of the order recovery process, this invention sets strict delay constraints on each stage of the trigger phase. Event consumption delay (from message enqueuing to consumer retrieval) is less than 500ms, trigger judgment time (three-level decision tree judgment and data preloading) is less than 100ms, and the total trigger delay is controlled within 600ms. When the delay exceeds the threshold, the system automatically triggers a degradation strategy: if the event consumption delay exceeds 500ms, the number of consumer instances is automatically increased; if the trigger judgment time exceeds 100ms, local caching is enabled to accelerate the query of user profiles and merchant configurations.

[0084] Step S2 (Feature Acquisition Stage): Acquire multi-dimensional feature data and construct feature vectors. The feature system includes seven basic categories and four multimodal feature categories: (1) User behavior characteristics: browsing depth (duration and scrolling depth of the product page, collected by tracking user mouse trajectory, scrolling speed and hot spot click distribution through page embedding), add-to-cart time (time interval from adding to cart to placing an order, reflecting the prudence of the purchase decision), number of product views, comparison behavior (whether to browse other products in the same category and the duration of comparison), collection behavior, customer service consultation records (number of consultations, keywords of consultation content and customer service response time), order placement time (distinguishing between weekdays / weekends and daytime / nighttimes, analyzing the return tendency of users who place orders at different times), payment method (Alipay / installment payment reflects high intention, cash on delivery reflects low intention) and sharing behavior (actively sharing to social platforms is a strong signal of high intention). In addition, for the live e-commerce scenario, new live room interaction behavior characteristics have been added, including live room dwell time, number of interactive comments, whether to click on the shopping cart, whether to participate in flash sales and the frequency of interaction with the host.

[0085] (2) Order characteristics: actual payment amount, product category, product gross profit margin, discount rate, whether it is a limited-time special offer order, whether it is a live broadcast order, the time from order placement to shipment, whether it is the user's first order, order source channel (natural search, paid promotion, social fission, live broadcast traffic, etc.) and the number of related orders (whether there are multiple related orders in the same period).

[0086] (3) Characteristics of Refund Reasons: The reason for a refund directly affects the likelihood of retaining the customer. Among them, refund reasons such as "I don't want it anymore / I ordered too much", "I ordered the wrong color / size", and "I saw something cheaper" have a high probability of retaining the customer (baseline value of retaining the customer 60% to 80%); "7-day no-reason return" requires further assessment of the purchase intention (baseline value of retaining the customer 40% to 60%); "Not good / Not suitable" can be resolved through a partial refund (baseline value of retaining the customer 30% to 50%); "Damaged goods" are difficult to retain the customer (baseline value of retaining the customer 10% to 30%); serious quality problems such as "Not as described / Fake goods" are not recommended for retaining the customer (baseline value of retaining the customer 0% to 10%).

[0087] (4) User profile characteristics: user level, historical cumulative order count, historical return rate, historical refund count, whether the user is a repeat customer, account registration duration, repeat customer loyalty score, and account fraud risk score. The system processes users in layers according to their type: high-loyalty old customers (historical return rate less than 5%, repeat customer loyalty score higher than 80) have the highest retention priority and can appropriately increase the discount; ordinary repeat customers (historical return rate 5% to 15%) receive discounts according to normal standards; new users (no historical data) receive discounts cautiously, with the main focus on building trust; users with high return rates (historical return rate higher than 30%) have their order retention costs strictly controlled; users suspected of malicious refunds (fraud risk score higher than the threshold) directly enter the risk control interception process.

[0088] (5) Product lifecycle characteristics: product shelf life, inventory turnover rate, seasonality index (e.g., winter down jackets returned in summer need to consider seasonal mismatch factors), competitor price fluctuation range (monitoring the price changes of major competitors in the past 7 days), and product evaluation sentiment score (based on NLP technology to conduct sentiment analysis on evaluations in the past 30 days to identify the concentration of quality complaints). When a product is in different stages of its lifecycle (new product stage, growth stage, maturity stage, decline stage), the order recovery strategy should be different: the new product stage focuses on collecting user feedback to optimize the product, the growth stage focuses on maintaining sales ranking, the maturity stage focuses on maximizing profits, and the decline stage focuses on clearing inventory.

[0089] (6) Logistics timeliness characteristics: delivery time (time from order placement to delivery), transit time (time from delivery to receipt), remoteness of the delivery area (based on the distance between the delivery address and the warehouse and the density of logistics coverage), number of logistics anomalies (abnormal scanning records during transportation), and deviation in receipt timeliness (difference between the actual receipt time and the expected receipt time). Refunds caused by logistics problems have a clear causal relationship, and order recovery strategies should prioritize solving logistics pain points rather than simply offering discounts. For example, for refunds caused by logistics delays, a combination strategy of refunding shipping costs and prioritizing delivery is recommended; for delivery problems in remote areas, it is recommended to change to a nearby warehouse for delivery or provide logistics compensation.

[0090] (7) Social Communication Characteristics: User social influence index (based on number of followers, interaction rate and historical sales data), product social communication popularity (number of times it has been shared / forwarded in the past 7 days), community activity (frequency of user comments and interaction quality in the merchant's private community) and KOC / KOL identification (whether they are key opinion consumers or key opinion leaders). For users with high social influence, even if the cost of retaining a single order is high, the word-of-mouth value they bring should be considered, and the conditions for retaining the order should be appropriately relaxed.

[0091] (8) Multimodal sentiment analysis features: This invention innovatively introduces multimodal sentiment analysis technology to extract user emotional state features from three dimensions: refund message text, voice customer service communication, and video customer service interaction. The refund message sentiment analysis adopts a BERT-based Chinese sentiment classification model to classify user refund messages into five sentiment tags: "anger," "disappointment," "hesitation," "calm," and "satisfaction," and outputs sentiment intensity scores (0 to 1). The voice customer service communication analysis converts the call content into text through speech recognition (ASR) and performs sentiment analysis. At the same time, it extracts acoustic features (speech rate, volume, pitch change, and pause frequency) to construct a voice sentiment vector and identify the trend of user emotional escalation. The video customer service communication analysis uses computer vision technology to detect user facial expressions in real time (based on a CNN model trained on the FER2013 dataset, identifying seven expressions: anger, disgust, fear, happiness, sadness, surprise, and neutral), and combines voice intonation analysis to generate a video conversation sentiment time-series curve. The multimodal emotion fusion layer uses an attention mechanism to dynamically weight the emotion scores of each modality and outputs comprehensive emotion state features to correct the probability of winning back a partner (the more negative the emotion, the greater the reduction in the probability of winning back a partner, up to a maximum reduction of 20%).

[0092] (9) Product Image Quality Assessment Features: For returned product images uploaded by users, this invention introduces a computer vision quality assessment model. This model is based on the ResNet-50 backbone network and fine-tuned on the product damage detection dataset, outputting a product damage score (0 to 1, where 0 is intact and 1 is severely damaged) and a heatmap of the damaged area. At the same time, the model assesses the image shooting quality (clarity, lighting, angle) and filters low-quality images to avoid misjudgment. For refund reasons of "product damage", the system combines the image quality assessment results with logistics anomaly records for cross-validation: if the image shows severe damage but there are no logistics anomaly records, the probability of malicious refund identification is increased; if the image shows minor defects and there are logistics anomaly records, the cost of order recovery is reduced and a partial refund strategy is recommended. The image quality assessment features are also used to optimize the calculation of partial refund amounts: for damage scores below 0.3, a compensation of 5% to 10% is recommended; for scores between 0.3 and 0.6, a compensation of 15% to 25% is recommended; and for scores above 0.6, a replacement or full refund is recommended.

[0093] (10) Cross-platform user identity fusion features: This invention identifies the identity of the same user on different e-commerce platforms through device fingerprinting technology (collecting browser Canvas fingerprints, WebGL fingerprints, font lists, time zones, language settings, etc.) and behavioral pattern matching (click hotspot distribution, scrolling speed, page browsing path, order time distribution, etc.). Cross-platform features include: cross-platform consumption frequency (number of orders on other platforms in the past 90 days), cross-platform return rate (return ratio on other platforms), cross-platform average order value comparison (ratio of average order value on this platform to average order value on cross-platform platforms), cross-platform category preference consistency (whether category A is purchased on this platform and the same type of product is purchased on other platforms), and cross-platform coupon sensitivity (whether coupons are frequently used on other platforms). Cross-platform identity fusion solves the problem of data silos on a single platform, enabling the system to more comprehensively evaluate user consumption habits and credit levels. For example, if a user is a new user on this platform but a high-repurchase user on other platforms, the system can appropriately relax the conditions for order retention to build trust.

[0094] (11) Temporal Behavioral Sequence Features: This invention collects temporal data on user behavior before a refund request to construct refund risk warning features. The behavioral sequence includes: the trend of changes in the number of product page views (sudden increase or decrease) within 7 days before the refund, the behavior of deleting items from the shopping cart, the frequency of comparing prices of similar products, the number of times customer service inquires about refund policies, the change in the dwell time on the order details page, and the ignore rate of App push messages. The behavioral mutation index is calculated by using sliding window statistics (1-day, 3-day, and 7-day windows) to identify potential refund intent signals. The temporal features are input into an LSTM network for sequence modeling, and the refund risk probability is output for pre-warning intervention in step S5.

[0095] Technical Architecture of Feature Engineering Pipeline: This invention adopts a dual-channel architecture of real-time stream processing and batch processing. The real-time channel uses a stream processing engine to achieve low-latency feature computation. When a refund request event is triggered, the stream processing engine extracts real-time features (such as browsing behavior in the last hour, page dwell time in the current session, etc.) from real-time data sources (user behavior log stream, order status change stream), with processing latency controlled within 200ms. The batch processing channel uses an offline computing engine to periodically (daily) pre-calculate user profile features and historical statistical features (such as average order value in the last 30 days, historical return rate, etc.), and stores them in the feature storage service for real-time querying. The features from both channels are merged at the feature concatenation layer to ensure a balance between real-time performance and feature richness.

[0096] Standardization Method for Numerical Features: For numerical features such as order payment amount, user's historical spending amount, and product browsing frequency, this invention uses the Z-score standardization method for normalization. The standardization formula is: z = (x - μ) / σ, where x is the original feature value, μ is the mean of the feature on the training set, and σ is the standard deviation. For numerical features with missing values, a median imputation strategy is used, that is, the median of the feature in the training set is used to replace the missing values. Median imputation is more robust than mean imputation and is not affected by extreme values. The standardized feature has a mean of 0 and a standard deviation of 1, ensuring that features of different dimensions have equal contribution to the model.

[0097] Encoding method for categorical features: This invention employs different encoding strategies based on the cardinality (number of values) of the categorical features. For low cardinality features (e.g., when the number of values ​​is less than a preset threshold, which can be 10 values ​​as an optional implementation, such as payment method, refund reason type, product category, etc.), One-Hot encoding is used, mapping each value to an independent binary dimension. For high cardinality features (e.g., when the number of values ​​is greater than or equal to the aforementioned preset threshold, which can be 10 values ​​as an optional implementation, such as product SKU, user ID hash, merchant ID, etc.), Target Encoding is used, with the encoding formula as follows:

[0098] Where count(cat=1) is the number of positive samples in this category, and count(cat) is the total number of samples in this category. α represents the global positive sample rate, and α is the smoothing parameter (default value α=10). Target Encoding effectively prevents overfitting of low-frequency categories by introducing a smoothing parameter, while compressing high-dimensional sparse features into a single value, reducing the dimensionality of the model input.

[0099] Vectorization Methods for Text Features: This invention employs different vectorization strategies for two types of text features: refund reason text and product title. The refund reason text is encoded using a pre-trained Chinese BERT model, with the output vector corresponding to the [CLS] tag serving as the semantic representation of the text. The vector dimension is 768. BERT encoding can capture deep semantic information of the refund reason (such as the semantic similarity between "size too large" and "clothes too big"). The product title uses the TextCNN model to extract n-gram features. TextCNN uses multi-scale convolutional kernels (window sizes of 2, 3, and 4) to capture phrase features of different granularities, which are then concatenated into a fixed-dimensional title semantic vector after max pooling. Compared to BERT, TextCNN has a faster inference speed in short text classification tasks, making it suitable for real-time encoding of short texts such as product titles.

[0100] Encoding method for behavioral sequence features: The user's behavioral sequence over the past 7 days is encoded using Position Encoding and Transformer Encoder. First, the daily behavioral feature vectors (containing 8 dimensions such as number of views, number of add-to-cart actions, number of customer service inquiries, and number of price comparison actions) are concatenated into a 7×8 input matrix. Then, position encoding is added to the input matrix. The position encoding formula is:

[0101] Where pos is the sequence position (0 to 6), i is the dimension index, and d is the encoding dimension.

[0102] Positional encoding enables the model to distinguish behavioral features at different time steps. The encoded matrix is ​​input into a Transformer Encoder (2 layers, 8-head attention, 128 hidden dimensions), and the output is a high-order representation vector of the behavioral sequence.

[0103] The formula for calculating the temporal mutation index: This invention defines the temporal mutation index ΔI(t) to quantify the degree of abnormality in user behavior at time t. The calculation formula is as follows:

[0104] Where x(t) is the behavioral characteristic value at time t (such as the number of views). This represents the average of the same characteristic values ​​over the previous 7 days. The standard deviation of the same feature value over the previous 7 days is represented by ΔI(t), and ε is a smoothing constant to prevent division by zero (default ε = 0.01). When ΔI(t) exceeds a preset threshold (e.g., 3.0, corresponding to approximately 3 times the standard deviation), that time point is marked as a behavioral mutation point. The mutation index comprehensively considers the absolute magnitude of behavioral changes and historical volatility, and has a higher tolerance for users whose behavior itself fluctuates greatly, avoiding misjudgment.

[0105] Dimensionality of the concatenated feature vector: After the feature engineering process described above, the final feature vector is composed of two parts: basic features and multimodal features. The basic features include user behavior features (32 dimensions), order features (28 dimensions), refund reason features (15 dimensions), user profile features (20 dimensions), product lifecycle features (18 dimensions), logistics timeliness features (15 dimensions), social dissemination features (12 dimensions), and various encoded derived features, totaling 386 dimensions. The multimodal features include BERT text vectors (768 dimensions reduced to 64 dimensions), TextCNN title vectors (32 dimensions), multimodal sentiment features (16 dimensions), image quality assessment features (8 dimensions), cross-platform fusion features (12 dimensions), and behavioral sequence Transformer output (10 dimensions), totaling 142 dimensions. After concatenation of the two parts, the final feature vector has a total dimension of 528.

[0106] Step S3 (Probability Evaluation Stage): Input the feature vector into the pre-trained multi-layer machine learning model to calculate the probability of order recovery, the probability of malicious refund identification, the probability of strategy sensitivity, and the probability of refund risk warning.

[0107] The calculation of the order retention probability employs an integrated architecture of a gradient boosting tree model (such as LightGBM) and a deep neural network model. The gradient boosting tree model handles structured features (such as numerical features like order amount and return rate), while the deep neural network model handles high-dimensional sparse features (such as user behavior sequences and refund reason text). The outputs of the two models are integrated through a weighted fusion layer and finally mapped to an order retention probability value in the [0,1] interval using a sigmoid activation function.

[0108] The specific formula for the LightGBM model: LightGBM uses a histogram-based decision tree learning algorithm, and the output of the leaf node of the t-th tree is:

[0109] in The output value of the j-th leaf node in the t-th tree. The feature space region corresponding to the j-th leaf of the t-th tree. The indicator function is 1 if the condition is true, and 0 otherwise. The final model output is a weighted sum of all trees. , Where η is the learning rate (default η=0.05) and T is the total number of trees (default T=500).

[0110] The learning rate η controls the contribution of each tree to the final prediction. A smaller learning rate combined with a larger number of trees can effectively prevent overfitting. LightGBM uses a leaf-wise growth strategy, selecting the leaf node with the largest gain for splitting at each step. Compared with the level-wise strategy, it has a lower training error with the same number of leaf nodes.

[0111] Sigmoid mapping formula: The original output H(x) of the LightGBM model is mapped to the single-win probability through the Sigmoid function. The mapping formula is as follows: , , The output range is [0,1]. The Sigmoid function maps any real value to a probability space, and the output range is [0,1]. Approaching 1, when H(x) approaches negative infinity Approaching 0. The Sigmoid function has the largest slope (derivative is 0.25) at H(x)=0, at which point the model has the strongest ability to distinguish between closing a position and not closing one.

[0112] The activation functions of each layer in a DNN model are as follows: Input layer (528-dimensional feature vector) → Embedding layer (maps sparse class features to dense vectors) → LayerNorm layer (normalizes the Embedding output to stabilize the training process) → BiLSTM layer (bidirectional LSTM, 64 hidden units each for forward and backward directions, capturing sequence dependencies between features) → Attention layer (multi-head attention mechanism, 8 heads, self-attention weighting of the BiLSTM output, focusing on key feature dimensions) → FC layer (fully connected layer, 128 neurons, using the GELU activation function, which has a smoother gradient transition than ReLU, the formula is GELU(x) = x * Φ(x), where Φ(x) is the cumulative distribution function of the standard normal distribution) → Output layer (single neuron, Sigmoid activation). The DNN model prevents overfitting through Dropout (rate 0.3) and L2 regularization (coefficient 1e-4).

[0113] Weighted Fusion Formula: The LightGBM model output P_gbt and the DNN model output P_dnn are weighted and fused to obtain the final order recovery probability. The fusion formula is as follows:

[0114] Where α is the tree model weight (default α=0.6), (1-α) is the DNN model weight (default 0.4), and β is the knowledge graph correction term coefficient (calculated by the GNN module in step S3, with a value range of [-0.15, 0.15]). The weighted fusion weight α is optimized on the validation set using Bayesian optimization, with a search range of [0.3, 0.8]. The knowledge graph correction term, as a fine-tuning factor, only takes effect when a strongly correlated path is detected, avoiding over-correction.

[0115] The probability of identifying malicious refunds is calculated by an independent risk control model that integrates a rule engine and anomaly detection algorithms. The rule engine sets hard rules based on expert experience (such as more than 10 returns in the past 30 days, multiple accounts associated with the same IP address, etc.), while the anomaly detection algorithms (such as Isolation Forest and Local Anomaly Factor algorithms) identify abnormal samples that deviate from normal user behavior patterns. The risk control model outputs the probability of identifying malicious refunds, which is used to intercept high-risk orders at the front end of the order recovery process.

[0116] The specific implementation of the risk control model: the rule engine outputs binary decisions. ∈ {0, 1}, where =1 indicates a hard rule has been met (such as more than 10 returns in the past 30 days, more than 3 accounts associated with the same device's fingerprint, an abnormal ratio of refund amount to order amount, etc.). The anomaly detection module uses the Isolation Forest algorithm to output anomaly scores. ∈ [0,1], A value closer to 1 indicates that the user's behavior pattern deviates more from the normal group. Isolation forests recursively construct isolation trees by randomly selecting features and split values. Outliers are typically isolated at relatively shallow depths, resulting in shorter path lengths. The final probability of identifying malicious refunds in the risk control model is calculated using a fusion formula: , Where λ is the weighting coefficient for outlier scores (default λ=0.8). This fusion strategy ensures that once a hard rule is hit ( =1), regardless of the anomaly score, it is judged as high risk; when the hard rule is not hit, the anomaly score is scaled by λ as the risk probability. When When the threshold is exceeded (e.g., 0.7), the system will refuse to accept the order and mark it as risk control.

[0117] The strategy sensitivity probability is calculated based on uplift modeling within a causal inference framework. Traditional prediction models only estimate the "probability of retaining the order after a strategy is given," failing to distinguish whether a user retains the order due to the strategy. This invention introduces uplift modeling, dividing users into four categories: "inevitable order retainers" (who will withdraw their refund regardless of whether a strategy is given, accounting for approximately 15%), "inevitable churners" (who will insist on a refund regardless of whether a strategy is given, accounting for approximately 55%), "strategy-sensitive users" (who only withdraw their refund when a strategy is offered, accounting for approximately 20%), and "strategy-averse users" (whose resolve to request a refund is strengthened by the strategy, accounting for approximately 10%). The uplift model employs tree-based methods (such as Causal Forest) or representation learning methods (such as DragonNet) to estimate the Conditional Average Treatment Effect (CATE) for each user. The system only recommends high-cost strategies to "strategy-sensitive users," provides only basic services (such as the exchange process) to "inevitable order retainers," avoiding excessive concessions; and avoids actively retaining orders for "strategy-averse users" to prevent escalation of negative experiences. Uplift modeling reduces the cost of closing deals by approximately 25% to 35% while maintaining the success rate of closing deals.

[0118] The CATE estimation formula for Causal Forest is: Conditional Average Treatment Effect (CAE) is defined as τ(x) = E[Y(1) - Y(0) | X=x], where Y(1) is the potential outcome when a sales recovery strategy is given (whether the sales recovery is successful), Y(0) is the potential outcome when no strategy is given, and X is the user feature vector. Causal Forest estimates the CATE value through the leaf nodes of a heterogeneous causal forest. Within each leaf node, the instrumental variable method or the difference-in-differences method is used to estimate the local treatment effect. The final CATE is the weighted average of the estimates from each leaf node. When τ(x) > (default When τ(x) = 0.1, the user is determined to be a strategy-sensitive user, and a high-cost strategy is recommended; when τ(x) < - If the user is deemed to be against the strategy, the seller should avoid actively trying to salvage the deal.

[0119] Knowledge Graph Enhancement for Order Recovery Probability Correction: This invention constructs a knowledge graph of product-user-refund reason and utilizes graph neural networks (GNNs) to mine deep relationships. The knowledge graph contains three types of nodes: product nodes (attributes include category, brand, price range, material, and applicable scenarios), user nodes (attributes include demographic information, consumption preferences, and credit rating), and refund reason nodes (attributes include reason type, severity, and seasonality). Edge relationships include: user-purchase-product, user-refund-product, product-similar-product, user-similar-user, refund reason-related-product category, etc. Message passing is performed on the knowledge graph using GraphSAGE or GAT graph neural networks to learn high-order interaction representations of user-product-refund reason. The GNN output is used to correct the order recovery probability: if a user's historical refund reason has a high-weighted association path with the current order product in the graph, the order recovery probability is adjusted accordingly (e.g., if a user has previously refunded a brand of clothing due to "size too large," the order recovery probability is increased by 10% to 15% when purchasing the same brand again). The knowledge graph is updated monthly, incorporating new products, new users, and new refund reasons.

[0120] The message passing formula for knowledge graph GNNs: Graph neural networks employ a k-layer message passing mechanism. The hidden state update formula for the k-th layer node v is as follows: , in Let v be the hidden representation of node v at the k-th layer. Let v be the set of neighboring nodes of node v, and AGG be the aggregation function (using mean pooling, i.e., taking the average of the hidden states of all neighboring nodes). The aggregation weight matrix is ​​learnable, and σ is the activation function (using ReLU). In this invention, k is a preset message passing layer, preferably k=2, i.e., performing 2-hop neighbor information aggregation to capture second-order associations (such as the indirect association between user-product-refund reason). Initial node embedding. It is composed of embeddings of node attributes. After two layers of message passing, the final hidden state of the target user node is retrieved. This is mapped to a knowledge graph correction value ΔP_kg via an MLP.

[0121] The refund risk warning probability is based on time-series behavioral sequence features, modeling changes in user behavior patterns before a refund using an LSTM network. The LSTM input is the daily behavioral feature vector for the 7 days prior to the refund (number of browsing visits, number of add-to-cart visits, number of customer service inquiries, number of price comparisons, etc.), and the output is the refund risk probability (between 0 and 1). When the risk probability exceeds a preset threshold (e.g., 0.7), the system triggers a pre-emptive warning intervention in step S5. The LSTM model employs a bidirectional structure (Bi-LSTM) to fully utilize contextual behavioral information, with a hidden layer dimension of 128 and a dropout rate between 0.28 and 0.32, with a dropout rate of 0.3 being optimal. It is trained using backpropagation time (BPTT).

[0122] The specific structure of the Bi-LSTM early warning model is as follows: The input dimension is 32 (8 behavioral features per day × a 4-day sliding window, i.e., behavioral data from day t-4 to day t-1). The 8 behavioral features are: number of product page views, number of shopping cart operations, number of customer service inquiries, number of price comparisons, normalized dwell time on the order details page, App push notification ignore rate, number of views on the refund policy page, and number of views on similar products. The Bi-LSTM hidden layer dimension is 128 (64 forward + 64 backward). The output is mapped to the refund risk probability through a fully connected layer (64 neurons, ReLU activation) and then through a Sigmoid function. ∈[0,1]. The model was trained using the Adam optimizer with an initial learning rate of 0.001, followed by cosine annealing after a 1000-step warm-up. The training data labels were determined by whether the user initiated a refund request within 7 days on day t (1 = initiated, 0 = not initiated).

[0123] The model training samples are divided into positive samples (successful retention, label=1), negative samples (failed retention, label=0), and neutral samples (those not triggered for retention or whose attempts to exploit loopholes were blocked, and are not included in the training set). The actual retention success rate is approximately 15% to 25%, with a positive to negative ratio of approximately 1:4 to 1:6. Imbalanced sample conditions are addressed by adjusting the weights of positive samples (scale_pos_weight) or using SMOTE oversampling technology. The model employs a training strategy combining weekly full updates and daily incremental updates to ensure timely adaptation to market changes.

[0124] Step S4 (Strategy Recommendation Stage): When the probability of malicious refunds exceeds a preset threshold (e.g., 70%), the order is rejected and flagged for risk control, triggering an account association investigation (checking other accounts associated with the same device fingerprint, the same shipping address, and the same payment account). Otherwise, the strategy recommendation process continues.

[0125] The core of the strategy recommendation is the comprehensive benefit evaluation function: Benefit evaluation score = (Probability of closing the deal x User lifetime value coefficient x User acceptance willingness coefficient x Amount of closing the deal) / (Cost of closing the deal + Risk cost + User churn cost).

[0126] The complete formula expansion of the comprehensive benefit evaluation function:

[0127] in The final probability of saving the order is calculated in step S3. User lifetime value coefficient The user acceptance coefficient for strategy s. This is the actual amount paid for the order. The execution cost of strategy s (such as coupon value, price difference compensation, etc.). To identify the probability of malicious refunds, Risk coefficient (default) =0.3), User churn cost (present value of future expected revenue loss estimated based on a user LTV model). Risk item. The potential losses and risk coefficients were quantified when attempts to salvage a deal failed and malicious refunds were encountered. =0.3 indicates that the system takes a cautious approach to risk and avoids excessive order recovery in high-risk scenarios. The long-term cost of user churn due to failed sales recovery is estimated using a user LTV decay model: ,in To measure the user retention rate after a failed attempt to salvage the deal (historically, it was approximately 0.4 to 0.6).

[0128] The User Lifetime Value (LTV) coefficient is calculated based on a user's historical spending, repurchase frequency, referral value (the number and spending of new users tracked via invitation codes), and expected future revenue. Even if the cost of retaining a high-LTV user is high, retaining that user should be prioritized. The User Acceptance Rate coefficient is the historical acceptance rate statistics for each strategy (exchange 78% > return shipping fee 65% > price difference refund 58% > coupon 52% > partial refund 45% > full refund 32%), and this coefficient is continuously updated using historical data. The Risk Cost is the product of the probability of malicious refund detection and the potential loss. The User Churn Cost is the expected future revenue loss due to failed retention efforts, estimated based on the user's LTV model.

[0129] The RFM-SL calculation formula for LTV coefficient: This invention uses an extended RFM-SL model to calculate the user lifetime value coefficient.

[0130] Where R is the number of days since the last purchase (Recency), F is the purchase frequency (Frequency), M is the total monetary value (Monetary), S is the social value (Social, based on the number of new users and their spending based on invitation codes), and L is the loyalty score (Loyalty, based on repurchase rate and product category diversity). The weights for each dimension are as follows: =0.15, =0.25, =0.30, =0.15, =0.15. Each dimension is normalized to the [0,1] interval using Min-Max:

[0131] final ∈ [0, 1], serving as the w_ltv coefficient in the comprehensive benefit evaluation function. The weight allocation reflects the dominance of spending amount (0.30) and purchase frequency (0.25) in LTV evaluation, while also taking into account social value and loyalty.

[0132] Dynamic Pricing and Stackelberg Game Model: This invention innovatively models the decision-making process of merchants and users as a dynamic game. In this game model, the merchant, as the leader, first decides the type and amount of the sales retention strategy, while the consumer, as the follower, decides whether to accept the sales retention based on the merchant's strategy. The merchant's payoff function is: Payoff = Sales Retention Amount - Sales Retention Cost - Reputation Loss Cost; the user's utility function is: Utility = Value of Retaining the Goods + Strategy Compensation Amount - Amount Obtainable from a Refund + Psychological Satisfaction. The Stackelberg game theory solves the subgame perfect Nash equilibrium through backward induction: first, the user's optimal response function (the user's decision given the merchant's strategy) is solved, and then the optimal strategy is solved by substituting it into the merchant's payoff function. The dynamic pricing engine adjusts the strategy amount in real time: when the user hesitates (e.g., the pop-up window stays for more than 3 seconds without clicking), the system automatically reduces the strategy amount to test the user's price sensitivity; when the user explicitly refuses, the system calculates the optimal amount for the second offer based on the game model (usually 1.2 to 1.5 times the initial offer, but not exceeding the merchant's cost limit). Game theory-based dynamic pricing increases the success rate of closing deals by approximately 12% to 18%, while avoiding cost overruns caused by blindly raising prices.

[0133] Mathematical modeling of Stackelberg game: The payoff function of the firm (leader) is , in x is the actual amount paid for the order, and x is the amount of compensation provided by the merchant to salvage the order. The reputation loss coefficient, This is an indicator function for whether a product has quality problems (if quality problems exist, the cost of reputational damage is even higher). The utility function of the user (follower) is: , in To retain users' subjective value assessment of the product, The amount that can be refunded. This represents the increase in psychological satisfaction (influenced by factors such as attitude towards closing the deal and response speed). The Nash equilibrium solution process involves first solving for the user's optimal response.

[0134] That is, the minimum compensation amount that the user accepts to salvage the deal; then the merchant imposes constraints. (Maximizing the revenue function of merchants under the constraint of cost ceiling) ,in As the upper limit of merchant costs, when If the merchant is unable to provide compensation that meets the user's minimum requirements, the order is deemed irrecoverable and will proceed directly to the refund process.

[0135] Reinforcement Learning Strategy Optimization: This invention models the order recovery strategy recommendation as a Markov Decision Process (MDP), maximizing long-term returns through deep reinforcement learning. The MDP state space includes: user feature vector, order feature vector, current order recovery round, historical strategy response records, and time decay factor; the action space consists of discrete strategy type selection (coupons, exchanges, price difference compensation, etc.) and continuous strategy amounts (within the merchant's configuration range); the reward function comprehensively considers immediate rewards (order amount obtained upon successful order recovery) and long-term rewards (user repurchase value, word-of-mouth value, strategy cost). This invention uses the Proximal Policy Optimization (PPO) algorithm to train the policy network. The policy network structure is: state input layer → fully connected layer (512-256-128) → policy head (output action probability distribution) and value head (output state value estimate). PPO training employs an experience replay mechanism, sampling trajectories from historical order recovery interaction data for offline training, while simultaneously collecting new data through online exploration (epsilon-greedy strategy). Reinforcement learning strategy optimization is particularly suitable for multi-round lead recovery scenarios (such as making a second attempt to recover a lead through outbound calls after the initial pop-up failure), avoiding short-sighted decision-making through long-term value estimation. Experiments show that the PPO-optimized strategy improves long-term returns by approximately 22% compared to the greedy strategy over 180 days.

[0136] The core formula of the PPO algorithm: PPO limits the policy update magnitude by pruning the objective function, and the approximate estimate of the policy gradient is: , in The ratio of the probabilities of the new and old strategies is given. `clip` is the estimated advantage function, `clip` is the pruning function, and `ε` is the pruning parameter (default ε = 0.2). The pruning mechanism ensures that each policy update does not deviate too far, improving training stability. The advantage function uses Generalized Advantage Estimation (GAE): , in λ represents the time-series difference error, γ is the discount factor (default γ=0.99), λ is the GAE parameter (default λ=0.95), and l is the number of steps in the multi-step expansion. GAE achieves a balance between bias and variance by adjusting λ: when λ approaches 0, it degenerates into single-step TD estimation (low variance, high bias), and when λ approaches 1, it degenerates into Monte Carlo estimation (high variance, low bias).

[0137] The benefits strategy pool includes at least two of the following strategy types: coupons (suitable for price-sensitive users; coupon amounts are intelligently calculated by AI within the thresholds configured by the merchant, supporting various forms such as full-reduction coupons, discount coupons, and no-threshold coupons), product coupons (suitable for precise discounts on high-value products; coupon amounts can be set by SKU), return shipping fees (suitable for refunds due to logistics / shipping fee disputes; the merchant bears the return shipping costs, and the user keeps the product), price difference compensation (suitable for disputes over promotional price differences / price fluctuations; the price difference is compensated in the form of red envelopes / compensation; a price difference cap must be set to prevent over-compensation), exchanges (suitable for product specifications not matching but the user still needs the product category; zero financial loss to maintain GMV), and gifts / reissues (suitable for missing / incomplete / damaged shipments; cost-effective). Transparent and quick resolution, partial refund (suitable for products with minor issues but still usable; refund percentage is automatically calculated based on the reason for refund and the extent of product damage), full refund without return (suitable for high-priced items and scenarios with a high probability of order recovery; highest cost, requires strict approval), intelligent outbound calling (suitable for high-value orders and scenarios with a high probability of order recovery; automatically triggers AI voice calls for proactive communication), private domain community outreach (sends personalized order recovery messages through merchant communities, official accounts, or WeChat Work; suitable for users with established private domain relationships), member points compensation (suitable for high-level members; compensates members in the form of points to enhance member stickiness), and personalized recommendation of alternative solutions (recommends similar but more suitable alternative products; suitable for accidental purchases / specification mismatch scenarios).

[0138] Step S5 (Execution Phase): Execute the optimal sales recovery strategy through five methods: pop-up interaction, intelligent outbound calling, private domain outreach, manual intervention, or pre-warning intervention.

[0139] The pop-up window automatically appears when a customer applies for a return and refund, showcasing a recovery plan and related benefits (such as a "5 yuan coupon" or "free exchange"). The user can click "Accept" to cancel the refund request, or click "Reject" to continue the original refund process. The pop-up interface uses A / B testing to optimize the copy and visual design, improving user acceptance. Each order only triggers one recovery opportunity.

[0140] The technical implementation of pop-up rendering: The order-recovery pop-up is encapsulated as an independent component based on the Web Component standard. Style isolation is achieved through Shadow DOM, ensuring that the pop-up style is not affected by the host page. The pop-up's text content, benefit display, button styles, etc., are dynamically configured through JSON Schema. The system automatically fills in the Schema fields and renders the pop-up according to the optimal strategy recommended in step S4. The JSON Schema defines configurable items such as text templates (supporting variable interpolation, such as "compensation of {{amount}} yuan coupon"), benefit icons, button text, and pop-up theme colors. The pop-up component supports lazy loading, loading component resources only on demand when the order-recovery process is triggered, keeping the first-screen rendering time within 100ms. The pop-up component also has a built-in event tracking SDK, automatically collecting user interaction behaviors (pop-up display duration, button click events, scrolling behavior, etc.) for subsequent effect analysis.

[0141] Pop-up A / B testing traffic splitting algorithm: The A / B test uses a deterministic hash-based traffic splitting algorithm based on user ID. The splitting formula is: bucket = hash(user_id) mod 100, assigning users to 100 buckets from 0 to 99. Each A / B test experiment occupies a certain number of buckets (e.g., 50 buckets for the experimental group and 50 buckets for the control group). The same user is always assigned to the same version in different experiments, ensuring the repeatability of experimental results. The hash function used is the MurmurHash3 algorithm, which has good uniform distribution and computational efficiency. The traffic splitting results are cached in the client's local storage to prevent the same user from seeing different versions in the same session. The minimum running period for the A / B test is 7 days to ensure that the sample size meets the statistical significance requirements.

[0142] Intelligent outbound calling is automatically triggered for high-value orders (order payment amount exceeding a preset threshold) with a high probability of order recovery (above 60%) and a user lifetime value level higher than a preset level. Before the call, a pre-call recovery plan is sent to the user via SMS or app. During the call, the AI ​​customer service dynamically adjusts its script and plan based on the user's voice feedback (e.g., automatically upgrading to a price difference refund or exchange plan if the user refuses a coupon). Call results are transmitted back to the system in real time for updating user profiles and model training.

[0143] The intelligent outbound calling system utilizes speech synthesis technology: Outbound call scripts are generated into natural speech via a Text-to-Speech (TTS) engine, supporting emotional speech synthesis with adjustable speech rate (0.8x to 1.2x), pitch (fundamental frequency offset ±20%), and timbre (male / female / child's voice). The TTS engine employs an end-to-end neural network speech synthesis model (such as Tacotron2 or FastSpeech2), converting text input into Mel spectrum, which is then processed by a vocoder (such as WaveGlow or HiFi-GAN) to generate waveform audio. Emotional speech synthesis automatically selects the appropriate voice style based on the closing scenario: a gentle and friendly tone (speech rate 0.9x, pitch +5%) is used for hesitant users, while a sincere apologetic tone (speech rate 0.85x, pitch -5%) is used for dissatisfied users. TTS synthesis latency is controlled within 500ms to ensure the natural fluency of outbound conversations.

[0144] The decision-making logic for dynamic adjustment of outbound call scripts: During outbound calls, the system uses ASR (Automatic Speech Recognition) to transcribe the user's speech in real time and combines it with a sentiment analysis model (a dual-modal sentiment analysis based on speech acoustic features and text semantics) to determine the user's emotional state. When negative emotions are detected (emotional score below the threshold of 0.3), the system automatically switches script templates: from standard closing scripts to reassuring scripts (such as "We are very sorry for the bad experience, we understand your feelings"), and appropriately increases the compensation (automatically increasing by 10% to 20% within the merchant's cost limit). When positive emotions are detected (emotional score above 0.7), the system accelerates the closing plan and directly guides the user to confirm acceptance. The script template library contains more than 20 sets of script templates in four categories: standard, reassuring, urgent, and upgraded, dynamically selected by the NLP engine based on user profiles and real-time emotional states.

[0145] Private domain outreach involves sending personalized sales recovery messages to users through established merchant communities, official accounts, or WeChat Work channels. Message content is customized based on user profiles and refund reasons. For example, messages like "Exclusive community benefits: free gifts with exchanges" might be sent to active community users, while text-and-image sales recovery plans might be sent to followers of the official account. The advantages of private domain outreach include low outreach costs, high user trust, and the ability to subtly recover sales by leveraging the community atmosphere.

[0146] Private Domain Outreach Message Template Engine: This invention employs a message template engine based on Jinja2 template syntax, supporting variable substitution, conditional rendering, and loop rendering. Template variables include: {{user_name}} (user nickname), {{order_id}} (order number), {{product_name}} (product name), {{strategy_detail}} (strategy details, such as compensation amount, exchange plan, etc.), {{refund_reason}} (refund reason), and {{deadline}} (validity period of the order retention plan). Template example: "Dear {{user_name}}, if you confirm to retain {{product_name}} before {{deadline}}, you will receive exclusive benefits for {{strategy_detail}}. If you have any questions, please feel free to contact us." The template engine performs variable validation and sensitive word filtering before sending to ensure the accuracy and compliance of the message content. The template library has more than 30 pre-set message templates based on refund reason type and user segmentation, and merchants can also customize templates.

[0147] Human intervention is suitable for complex refund scenarios (products with significant disputes, high risk of customer complaints, or involving food safety or personal safety), where the refund is transferred to a human customer service representative. The customer service workbench integrates system-recommended strategies, scripts, and user profile summaries to assist customer service representatives in making quick decisions. Results of human-assisted order resolution are also incorporated into the feedback optimization module.

[0148] Pre-emptive warning intervention is an innovative implementation method of this invention, proactively intervening before a user formally initiates a refund request. Based on the refund risk warning probability in step S3 (output by the Bi-LSTM model) and the temporal behavioral sequence features in step S2, the system identifies orders with potential refund risks, especially high-risk orders. Warning trigger conditions include: a refund risk probability greater than 0.7 and the actual payment amount of the order exceeding a preset threshold; or the user frequently browsing the refund policy page, deleting items from their shopping cart, or comparing prices across platforms within 48 hours before the refund request. Warning intervention methods include: proactively pushing caring messages (such as "Hi, we noticed you've been looking at this product recently, how can we help you?"), automatically triggering dedicated customer service follow-up, issuing targeted coupons or gift promises in advance, and optimizing logistics and delivery plans (such as expedited shipping or changing logistics providers). Pre-emptive warning intervention shifts the opportunity to salvage the order from "after refund" to "before refund," fundamentally reducing the refund initiation rate. Experimental data shows that pre-emptive warning intervention reduced the refund initiation rate by approximately 18% to 25%. Furthermore, due to lower psychological defenses before a refund, the success rate of recovering the order with warnings is 30% to 40% higher than recovering the order after a refund. The warning system employs a tiered response mechanism: a risk probability of 0.7 to 0.8 triggers a mild warning (app push notifications); 0.8 to 0.9 triggers a moderate warning (dedicated customer service proactively communicates); and above 0.9 triggers a severe warning (outbound phone calls plus targeted benefits).

[0149] The formula for tiered response decisions based on early warning: Early warning response level = Probability of refund risk The calculation and decision formula is as follows: , in This is a rounding-down function. The `level` mapping represents a three-tiered alert system: when level ∈ {7, 8}, a Level 1 alert (mild alert) is triggered, and the system sends a supportive message via App push notifications or in-app messages, without consuming customer service resources; when level = 9, a Level 2 alert (moderate alert) is triggered, and the system automatically assigns a dedicated customer service representative to proactively contact the user via online chat to understand potential problems and provide solutions; when level = 10 (i.e.,...), a Level 3 alert is triggered. In extreme cases ≥ 1.0, the model is calibrated to [value missing] in practice. When the risk level is ≥ 0.95, a Level 3 alert (severe alert) is triggered. The system automatically initiates an outbound call with a targeted benefit plan (such as exclusive coupons or gift promises), and simultaneously notifies the merchant's operations staff to pay attention. The tiered response mechanism ensures the efficient allocation of alert resources and avoids excessive intervention for low-risk users.

[0150] Step S6 (Feedback Optimization Phase): Receive callback feedback on the order recovery execution results, update the model training samples using the A / B testing framework, and achieve closed-loop optimization. Simultaneously, support multi-vendor joint modeling through a federated learning architecture, sharing model parameters while protecting data privacy.

[0151] Technical Specifications of the Callback Interface: The callback interface for the order recovery execution result adopts a RESTful API design, with the request method being POST and the request body format being JSON. The callback interface defines the following core fields: order_id (unique order identifier), retention_status (order recovery status: success / fail / timeout), retention_amount (order recovery amount when successful, unit: cents), strategy_type (strategy type), strategy_cost (actual order recovery cost, unit: cents), callback_trigger (callback trigger type: user_accept / user_reject / timeout / callback_expired), outbound_status (outbound call status: not_triggered / in_progress / completed / failed), private_domain_status (private domain reach status: not_sent / sent / clicked / replied), and emotion_score (user feedback emotion score, range 0 to 1). The callback interface employs a timeout retry strategy: after the initial callback timeout, an exponential backoff algorithm is used for retries, with retry intervals of 1 second, 2 seconds, and 4 seconds, and a maximum of 3 retries. If all 3 retries fail, the callback message is persisted to a dead-letter queue for manual handling by operations personnel. The callback interface supports idempotency design, using `order_id` + `callback_trigger` as the idempotency key to ensure that duplicate callbacks do not lead to duplicate data processing.

[0152] Callback fields include: order recovery amount (unit: cents), actual recovery cost (unit: cents), recovery status (success / failure), whether an outbound call was made, outbound call status, private domain outreach status, and user feedback sentiment score (based on NLP sentiment analysis of user comments or voice). Callback triggering timings include: immediately after the user accepts the recovery plan, after the user explicitly refuses or fails to respond within the timeout period, after the intelligent outbound call is completed, after the user clicks a link or replies to a message after private domain outreach, and when the user initiates after-sales service again after the recovery (at which point the system automatically deducts the already refunded portion to avoid duplicate refunds).

[0153] The A / B testing framework randomly assigns different strategies to similar user groups (based on user profile clustering). By comparing the success rate of closing deals, user satisfaction (collected through after-sales reviews), short-term repurchase rate (whether the customer repurchases within 30 days after closing a deal), and long-term repurchase rate (repurchase frequency within 180 days after closing a deal), the strategy recommendation model is continuously optimized. A / B test results are summarized and analyzed weekly. Significantly superior strategies are automatically expanded in application, while poorly performing strategies are automatically taken offline or have their parameters adjusted.

[0154] Statistical testing methods for A / B testing: This invention employs corresponding statistical testing methods for different types of evaluation indicators. Proportional indicators include the success rate of closing deals. For proportional indicators, the chi-square test (χ² test) is used to compare whether the difference between the experimental group and the control group is statistically significant. The chi-square statistic... , in For the frequency of observation, This represents the expected frequency. Continuous indicators include order value and user satisfaction rating. Independent samples t-tests are used to compare the differences between the means of two groups for continuous indicators. The significance level for all tests was set at α = 0.05, meaning that a p-value less than 0.05 was considered statistically significant. To control for false positives from multiple comparisons, Bonferroni correction was used: the corrected significance level was... =α / m, where m is the number of indicators tested simultaneously.

[0155] Minimum Sample Size Calculation Formula: Before initiating an A / B test, the minimum sample size needs to be calculated to ensure sufficient statistical power. The formula is as follows:

[0156] in =1.96 is the quantile of the normal distribution corresponding to the 95% confidence level, p is the basic success rate of closing a deal (historical statistical value, such as 0.20), δ is the minimum detectable effect (default δ=0.02, i.e., a 2% increase in the success rate of closing a deal), and n is the minimum sample size required for each group. Taking p=0.20 and δ=0.02 as an example... , This means that each group requires at least 3073 samples. When the A / B test starts, the system automatically checks whether the current traffic meets the minimum sample size requirement. If it does not, the test period is extended or the traffic allocation ratio is increased.

[0157] Federated Learning Architecture: This invention introduces a federated learning framework to address the poor model performance caused by insufficient data from small and medium-sized businesses. In federated learning mode, each business's local data remains within its domain, sharing only model parameters (gradient or weight updates). The system adopts a horizontal federated learning (sample alignment, feature alignment) architecture. After training their models locally, each business uploads encrypted gradients to a central coordination server. The server aggregates the global model (e.g., the FedAvg algorithm: global weights = weighted average of each business's local weights, where weights represent the proportion of each business's sample size) and then distributes the updates. To protect model parameter privacy, differential privacy (DP) technology is used to add calibrated noise to the gradients (the noise scale is proportional to the gradient sensitivity), or secure multi-party computation (SMPC) is used to achieve encrypted aggregation. Federated learning is particularly suitable for the following scenarios: cold start for newly joined businesses (rapid deployment using shared models from other businesses), seasonal goods businesses (enhancing model generalization ability using data from non-seasonal businesses), and vertical category businesses (learning general order-taking patterns using data from cross-category businesses). Experiments show that federated learning improves the success rate of order recovery for small and medium-sized businesses with fewer than 1,000 samples by approximately 15% to 20%, while ensuring that user privacy data does not leave the business's local environment. The federated learning training cycle involves global aggregation every two weeks and daily local incremental updates.

[0158] FedAvg aggregation formula for federated learning: In each round of federated aggregation, the central coordinating server collects the local model weights of K vendors and aggregates them into global model weights using the federated averaging algorithm (FedAvg). The aggregation formula is as follows:

[0159] in The global model weights for round t+1 are... Let K be the local model weights for merchant k after training round t. Let K be the local training sample size for merchant k. Let K be the total sample size of all merchants, and K be the total number of merchants participating in federated learning. FedAvg uses sample size weighting to ensure that model updates from large merchants have a greater impact, while updates from small merchants can also contribute to the global model. The aggregated global model weights are distributed to each merchant as the initial weights for the next round of local training. The default communication round of federated learning is once every two weeks (T=2 weeks), and each round of aggregation takes about 30 minutes (including gradient transfer, aggregation calculation, and model distribution).

[0160] The noise addition formula for differential privacy: To prevent model parameters from being used to infer user privacy data, this invention introduces a differential privacy protection mechanism in the gradient upload stage of federated learning. The noise addition formula is as follows: , Where g is the original gradient vector. Let σ be the noise following a multivariate normal distribution, and I be the identity matrix. The formula for calculating the noise scale σ is: Where Δf is the gradient sensitivity (i.e., the maximum influence of a single sample on the gradient vector, controlled by the gradient clipping threshold c), c is the clipping threshold (default c=1.0), and ε is the privacy budget (default ε=10; the smaller ε is, the stronger the privacy protection but the greater the loss in model accuracy). Differential privacy ensures that attackers cannot infer whether a specific user participated in training by observing changes in model parameters. When ε=10, the model accuracy loss is approximately 2% to 5%, which is within an acceptable range. The privacy budget consumed in each round of federated aggregation is added to the total budget; federated learning training stops when the total budget is exhausted.

[0161] The embodiments in this specification are only used to illustrate the technical solution of the artificial intelligence order recovery of the present invention and do not constitute a limitation of the present invention. Those skilled in the art can optimize the algorithm, split the modules, adapt the terminal, and combine strategies based on the core logic of the present invention's loss identification, intelligent recommendation for order recovery, and layered outreach intervention. All equivalent substitutions and simple modifications fall within the protection scope of the present invention, and the scope of protection is defined by the claims.

Claims

1. A method for recovering e-commerce orders based on artificial intelligence, characterized in that, Includes the following steps: S1: In response to the detection of a consumer's return and refund request for an order on the e-commerce platform, the order recovery process is triggered; S2: Collect multi-dimensional feature data, which includes at least user behavior features, order features, refund reason features, user profile features, product lifecycle features, logistics timeliness features, and social dissemination features. Standardize and encode the feature data of each dimension and splice them together to form a feature vector of a unified dimension. S3: Input the feature vector into a pre-trained machine learning model, and the machine learning model outputs the probability of winning the sale and the probability of identifying malicious refunds; S4: Determine whether the probability of identifying malicious refunds exceeds a preset threshold; If the limit is exceeded, the attempt to recover the order will be rejected and an account association risk control investigation will be triggered. If the probability of closing the deal is not exceeded, the benefit evaluation score of each candidate strategy is calculated by using a comprehensive benefit evaluation function based on the aforementioned probability of closing the deal and the preset equity strategy pool. The strategy with the highest benefit evaluation score is selected as the optimal strategy for closing the deal. S5: Execute the optimal order recovery strategy through pop-up interaction, intelligent outbound calling, private domain outreach, or manual intervention; S6: Receive the callback feedback of the order recovery execution result, and use the callback feedback as a training sample to update the machine learning model.

2. The method according to claim 1, characterized in that, In step S2: The user behavior characteristics include: browsing depth, time to add to cart, number of times to browse products, comparison behavior, collection behavior, customer service inquiry records, order time, payment method, sharing behavior, and live broadcast interaction behavior; The order characteristics include: actual payment amount, product category, product gross profit margin, discount rate, whether it is a limited-time special offer order, whether it is a live broadcast order, the time from order placement to shipment, whether it is the user's first order, order source channel, and number of related orders; The product lifecycle characteristics include: product shelf life, inventory turnover rate, seasonality index, competitor price fluctuation range, and product sentiment score. The logistics timeliness characteristics include: delivery timeliness, transit time, remoteness of the delivery area, number of logistics anomalies, and deviation in the signing time.

3. The method according to claim 1, characterized in that, Step S2 further includes the collection of multimodal sentiment analysis features: the multimodal sentiment analysis features include: the sentiment tags and sentiment intensity scores obtained by encoding the refund message text using the BERT model, the voice sentiment vector constructed after the voice customer service call content is transcribed by ASR and the acoustic features are extracted, and the video conversation sentiment time-series curve generated after the video customer service facial expressions are recognized by the CNN model; The emotional labels are divided into five categories: "anger", "disappointment", "hesitation", "calm" and "satisfaction", and the emotional intensity score range is 0 to 1. The acoustic features include speech rate, volume, pitch variation and pause frequency. The CNN model is trained on the FER2013 dataset and recognizes seven expressions: anger, disgust, fear, happiness, sadness, surprise and neutral.

4. The method according to claim 3, characterized in that, The multimodal sentiment analysis features dynamically weight and fuse the sentiment scores of each modality through an attention mechanism to output comprehensive sentiment state features, which are used to correct the probability of winning the sale.

5. The method according to claim 1, characterized in that, Step S2 also includes collecting product image quality assessment features: The product image quality assessment features include: damage score of returned product images output by the ResNet-50 model, heat map of damaged area location, and image shooting quality assessment results; The damage rating range is from 0 to 1, which is used to cross-validate logistics anomaly records and assist in calculating partial refund amounts.

6. The method according to claim 1, characterized in that, Step S2 also includes the collection of cross-platform user identity fusion features and time-series behavioral sequence features: The cross-platform user identity fusion features include: cross-platform purchase frequency, cross-platform return rate, cross-platform average order value comparison, cross-platform category preference consistency, and cross-platform coupon sensitivity; the cross-platform user identity is identified through device fingerprint technology and behavioral pattern matching. The temporal behavioral sequence features include: the trend of product page views within a preset time period before refund, the behavior of deleting items from the shopping cart, the frequency of comparing prices of similar products, the number of times customer service inquires about refund policies, the change in the dwell time on the order details page, and the ignore rate of App push messages. The temporal behavioral sequence features are statistically calculated using a sliding window to obtain a behavioral mutation index.

7. The method according to claim 1, characterized in that, In step S3: The machine learning model adopts an integrated architecture of gradient boosting tree model and deep neural network model; The gradient boosting tree model processes the structured numerical features in the feature vector and outputs a first predicted probability. The deep neural network model processes the high-dimensional sparse features in the feature vector and outputs a second prediction probability; the first prediction probability and the second prediction probability are weighted and fused, and then mapped to the order recovery probability through the Sigmoid activation function.

8. The method according to claim 7, characterized in that, The structure of the deep neural network model includes: The input layer receives the feature vector; The Embedding layer maps sparse class features into dense vectors; Layer normalization layer, which performs layer normalization on the Embedding output; A bidirectional LSTM layer with 64 hidden units each for the forward and backward directions; A multi-head attention layer with 8 heads performs self-attention weighting on the bidirectional LSTM output; The fully connected layer has 128 neurons and uses the GELU activation function. The output layer consists of a single neuron and uses the Sigmoid activation function to output the second predicted probability.

9. The method according to claim 1, characterized in that, Step S3 further includes: The probability of identifying malicious refunds is calculated by an independent risk control model, which integrates a rule engine and an anomaly detection algorithm. The rule engine outputs binary decision results based on preset hard rules; The anomaly detection algorithm uses the isolated forest algorithm to output anomaly scores; The probability of identifying malicious refunds is calculated using a fusion formula: ， in, The binary decision result of the rule engine. Let λ be the abnormal score, and λ be the abnormal score weighting coefficient.

10. The method according to claim 1, characterized in that, Step S3 further includes calculating the policy sensitivity probability: The strategy sensitivity probability is calculated based on uplift modeling in the causal inference framework. The average conditional treatment effect of each user is estimated by CausalForest, and users are divided into four groups: users who are bound to win orders, users who are bound to churn, strategy-sensitive users, and strategy-averse users.

11. The method according to claim 1, characterized in that, Step S3 also includes knowledge graph-enhanced order recovery probability correction: Construct a knowledge graph that includes product nodes, user nodes, and refund reason nodes; Message passing is performed on the knowledge graph using a graph neural network to learn a high-order interactive representation of user-product-refund reason; When a high-weighted association path in the knowledge graph is detected between a user's historical refund reasons and the current order items, the order recovery probability is adjusted according to the weight of the association path.

12. The method according to claim 11, characterized in that, The message passing mechanism of the graph neural network, the hidden state update formula of the k-th layer node v is: , in Let v be the hidden representation of node v at the k-th layer. Let v be the set of neighboring nodes. σ is a learnable aggregate weight matrix, σ is the activation function, and k is the preset message passing layer.

13. The method according to claim 1, characterized in that, The comprehensive benefit evaluation function in step S4 is: Benefit evaluation score = (Probability of closing the deal × Customer lifetime value coefficient × Customer acceptance willingness coefficient × Deal closing amount) / (Cost of closing the deal + Risk cost + Customer churn cost); The user lifetime value coefficient is calculated based on the user's historical spending amount, repurchase frequency, and referral value; the user acceptance willingness coefficient is the historical acceptance rate statistics of each strategy; the risk cost is the product of the probability of malicious refund identification and the potential loss; and the user churn cost is the loss of expected future revenue due to the failure to recover the order.

14. The method according to claim 1, characterized in that, Step S4 also includes Stackelberg game-based dynamic pricing: The decision-making process of merchants and users is modeled as a dynamic game, with merchants as leaders deciding the type and amount of the sales recovery strategy, and users as followers deciding whether to accept the sales recovery based on the merchant's strategy. Solving the subgame perfect Nash equilibrium using backward induction; The dynamic pricing engine adjusts the strategy amount based on the user's real-time response.

15. The method according to claim 14, characterized in that, The mathematical modeling of the Stackelberg game: The payoff function of the firm as the leader is: ， in x is the actual amount paid for the order, and x is the amount of compensation provided by the merchant to salvage the order. The reputation loss coefficient, This is an indicator function for whether a product has quality problems. The utility function of a user as a follower is: ， in To retain users' subjective value assessment of the product, The amount that can be refunded. Increase in psychological satisfaction; Nash equilibrium solution process: First, we need to find the user's optimal response. This refers to the minimum compensation amount a user is willing to accept for salvaging the deal. Then the merchant is under control. Maximizing the revenue function of merchants under the constraint of cost ceiling ,in As the upper limit of merchant costs, when If the merchant is unable to provide compensation that meets the user's minimum requirements, the order is deemed irrecoverable and will proceed directly to the refund process.

16. The method according to claim 1, characterized in that, Step S4 further includes deep reinforcement learning strategy optimization: The recommended strategy for closing deals is modeled as a Markov decision process. The state space of the Markov decision process includes user feature vector, order feature vector, current order recovery round, historical strategy response record and time decay factor; The action space includes discrete strategy type selection and continuous strategy amount; The reward function takes into account both immediate and long-term rewards. A proximal policy optimization algorithm is used to train the policy network to maximize long-term cumulative returns.

17. The method according to claim 1, characterized in that, The equity strategy pool in step S4 includes at least two of the following strategy types: Coupons, product vouchers, return shipping fees, price difference refunds, exchanges, gifts or replacements, partial refunds, full refunds without returns, intelligent outbound calling, private community outreach, member points compensation, and personalized recommendation alternatives.

18. The method according to claim 2, characterized in that, The triggering conditions for intelligent outbound calling in step S5 are as follows: when the order recovery probability is greater than a preset threshold, the actual payment amount of the order is greater than a preset amount threshold, and the user's lifetime value level is higher than a preset level, intelligent voice outbound calling is automatically triggered, and an order recovery plan is pushed to the user before the outbound call.

19. The method according to claim 6, characterized in that, The machine learning model in step S3 also outputs a refund risk warning probability: The refund risk warning probability is output based on the sequence modeling of the temporal behavioral sequence features using a bidirectional LSTM network. Step S5 also includes pre-emptive warning intervention: when the refund risk warning probability exceeds a preset threshold, the warning intervention is proactively triggered before the user formally initiates a refund application; The early warning and intervention methods include: proactively pushing care messages, automatically triggering dedicated customer service follow-ups, issuing targeted coupons or gifts in advance, and optimizing logistics and delivery plans; The early warning intervention adopts a tiered response mechanism, triggering different levels of early warning measures based on the numerical range of the refund risk warning probability.

20. The method according to claim 1, characterized in that, In step S2: Numerical features are normalized using the Z-score standardization method, and missing values ​​are filled using the median imputation strategy. In categorical features, those with a cardinality less than a preset value are encoded using One-Hot encoding, while those with a cardinality greater than or equal to a preset value are encoded using Target Encoding. The refund reason text is encoded into a semantic vector using a pre-trained Chinese BERT model; Product titles are used to extract n-gram features using the TextCNN model; User behavior sequences are encoded using Position Encoding and Transformer Encoder.

21. The method according to claim 1, characterized in that, In step S2, the feature acquisition adopts a dual-channel architecture of real-time stream processing and batch processing. The real-time channel extracts real-time features from real-time data sources based on a stream processing engine, with processing latency controlled within 200ms; The batch processing channel pre-calculates user profile features and historical statistical features periodically based on the offline computing engine, and stores them in the feature storage service for real-time querying; The dual-channel features are merged in the feature splicing layer to form the feature vector.

22. The method according to claim 1, characterized in that, In step S6: The callback feedback is received via a RESTful API through a callback interface, and the request body is in JSON format. The callback interface adopts a timeout retry strategy: after the first callback times out, the exponential backoff algorithm is used for retry, with retry intervals of 1 second, 2 seconds, and 4 seconds, and a maximum number of retries of 3. The callback interface uses the combination of the order's unique identifier and the callback trigger type as an idempotent key.

23. The method according to claim 1, characterized in that, Step S6 also includes an A / B testing framework: The A / B testing framework randomly assigns different strategies to similar user groups; By comparing the success rate of order recovery, user satisfaction, short-term repurchase rate and long-term repurchase rate of each group, the strategy recommendation model is continuously optimized. The proportional indicators were statistically significant using the chi-square test, and these proportional indicators included the success rate of order recovery. The continuous indicators, including order amount and user satisfaction score, were tested for statistical significance using an independent samples t-test.

24. The method according to claim 1, characterized in that, Step S6 further includes federated learning optimization: Each participating merchant trains its local model on the premise that its local data does not leave the domain; Each merchant uploads encrypted gradients to the central coordination server; The central coordination server aggregates the local model weights of each merchant through a federated averaging algorithm to generate a global model. The aggregation formula for the federated average algorithm is: ， in The global model weights for round t+1 are... Let K be the local model weights for merchant k after training round t. Let K be the local training sample size for merchant k. K represents the total sample size of all merchants, and K represents the total number of merchants participating in federated learning. Differential privacy techniques are used to add calibration noise to the gradient.

25. An artificial intelligence-based e-commerce order recovery system, characterized in that, include: The trigger module is used to respond to the detection of a user's return and refund request for an order on the e-commerce platform and trigger the order recovery process. The feature acquisition module is used to collect multi-dimensional feature data in response to the detection of a return and refund application event. The multi-dimensional feature data includes at least user behavior features, order features, refund reason features, user profile features, product lifecycle features, logistics timeliness features, and social dissemination features. The feature data of each dimension is standardized and encoded and then concatenated to form a feature vector. The probability evaluation module is used to input the feature vector into a pre-trained machine learning model to calculate the probability of winning the sale and the probability of identifying malicious refunds. The strategy recommendation module is used to determine whether the probability of identifying malicious refunds exceeds a preset threshold. If the probability exceeds the limit, the attempt to recover the order will be rejected and an account association risk control investigation will be triggered; if the probability does not exceed the limit, the benefit evaluation score of each candidate strategy will be calculated through a comprehensive benefit evaluation function based on the recovery probability and the preset benefit strategy pool, and the strategy with the highest benefit evaluation score will be selected as the optimal recovery strategy. The execution module is used to execute the optimal order-saving strategy through pop-up interaction, intelligent outbound calling, private domain outreach, or manual intervention. The feedback optimization module is used to receive callback feedback of the order recovery execution result and use the callback feedback as training samples to update the machine learning model.

26. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 24.

27. A computer program product, comprising a computer program or computer instructions, characterized in that, When the computer program or computer instructions are executed by a processor, the method of any one of claims 1 to 24 is implemented.

28. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 24.

Images