Customer complaint order identification method and device, equipment and medium

Abstract

The invention relates to a customer complaint order recognition method and device, equipment and medium in the technical field of e-commerce, and the method comprises the steps: responding to an order payment event, obtaining a target feature set related to a target order corresponding to the event, and enabling the target feature set to comprise a plurality of model-in features corresponding to the target order in order, buyer and merchant dimensions; identifying a customer complaint risk probability of the target order according to the target feature set by using a customer complaint identification model, and judging whether the customer complaint risk probability exceeds a risk threshold; and when the customer complaint risk probability exceeds a risk threshold value, marking the target order as a high customer complaint risk order, and visually displaying the high customer complaint risk order to a corresponding buyer and a merchant. According to the invention, the customer complaint risk of the order is identified immediately, efficiently and accurately after payment, and the buyer and the merchant are reminded in time when the risk is high.

Description

Technical Field

[0001] This application relates to the field of e-commerce technology, and in particular to a method for identifying customer complaint orders and the corresponding apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] With the rapid development of global e-commerce, cross-border transactions have become an important part of international trade. Against this backdrop, third-party payment platforms, acting as the central hub for fund transfers between cross-border e-commerce platforms and domestic and international buyers and sellers, undertake core functions such as payment settlement, risk management, and transaction security. In practice, when a buyer completes order payment through a third-party payment platform, disputes may arise due to reasons such as lost goods, defective goods, or transactions not conducted by the buyer's consent. Buyers typically file an order complaint directly with the third-party payment platform or the cross-border e-commerce platform. Handling customer complaints not only involves determining liability for a single transaction but also concerns the harmonious development of all three parties involved.

[0003] In traditional technology models, cross-border e-commerce platforms primarily rely on post-event response mechanisms to identify and handle customer complaints. Specifically, platforms often only initiate manual intervention after a buyer actively files a dispute complaint. This involves assigning professionals to connect with the dispute resolution channels and refund processes of third-party payment platforms, while simultaneously coordinating with customer service teams to communicate and negotiate with both the buyer and seller. The technical principle behind this passive response model is that it treats customer complaints as the initial signal to trigger the processing flow, lacking the ability to proactively perceive potential risks. The primary technical problem resulting from this is that platforms cannot provide early warnings or intervention before complaints actually occur; they can only invest resources in remediation after the dispute escalates, leading to a continuous increase in operating costs.

[0004] Furthermore, traditional technologies for assessing customer complaint risk are limited to statistical analysis of superficial indicators, such as calculating historical dispute rates and setting fixed thresholds as standards for merchant entry or penalties. The inherent logic of this approach is to treat customer complaints as a discrete, isolated set of events, failing to establish a real-time, comprehensive, multi-dimensional risk analysis. Therefore, e-commerce platforms struggle to assess risk in a timely manner within the critical window after order payment. This fundamental flaw directly results in a significant delay in risk management. E-commerce platforms cannot identify high-risk orders before irreversible operations such as fund settlement and product delivery, missing the optimal time for handling them, thus making it difficult to effectively suppress customer complaint rates. Moreover, buyers, aware of the risks too late, may file complaints in anger, and merchants are unable to proactively optimize services for high-risk orders, ultimately exacerbating the risk of customer churn. Especially when merchants face account freezes due to customer complaint rates exceeding the red line set by third-party payment platforms, they often choose to leave the e-commerce platform to avoid risk, potentially leading to a vicious cycle of losing valuable merchant resources.

[0005] Therefore, in order to protect the rights and interests of merchants and buyers and maintain the healthy development of cross-border e-commerce platforms, it is necessary to identify and handle customer complaints in a timely and effective manner to reduce losses for all three parties. Summary of the Invention

[0006] The primary objective of this application is to address at least one of the aforementioned problems by providing a customer complaint order identification method and corresponding apparatus, computer equipment, and computer program product.

[0007] To achieve the various objectives of this application, the following technical solution is adopted: A customer complaint order identification method provided for one of the purposes of this application includes the following steps: In response to an order payment event, obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features of the target order in the dimensions of order, buyer, and merchant. A customer complaint identification model is used to identify the probability of customer complaint risk for the target order based on the target feature set, and to determine whether the probability of customer complaint risk exceeds the risk threshold. When the probability of customer complaints exceeds the risk threshold, the target order is marked as a high-risk order for customer complaints, and the high-risk order for customer complaints is displayed visually to the corresponding buyer and merchant.

[0008] On the other hand, a customer complaint order identification device provided to meet one of the purposes of this application includes an event response module, a risk assessment module, and a risk display module. The event response module is used to respond to an order payment event and obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features corresponding to the target order in the dimensions of order, buyer, and merchant. The risk assessment module is used to use a customer complaint identification model to identify the customer complaint risk probability of the target order based on the target feature set and determine whether the customer complaint risk probability exceeds a risk threshold. The risk display module is used to mark the target order as a high-risk customer complaint order when the customer complaint risk probability exceeds the risk threshold and visually display the high-risk customer complaint order to the corresponding buyer and merchant.

[0009] In another aspect, a computer device provided for one of the purposes of this application includes a central processing unit and a memory, the central processing unit being used to invoke and run a computer program stored in the memory to perform the steps of the customer complaint order identification method described in this application.

[0010] In another aspect, a computer program product provided for another purpose of this application includes a computer program / instructions that, when executed by a processor, implement the steps of the method described in any embodiment of this application.

[0011] The technical solution of this application has many advantages, including but not limited to the following aspects: First, by treating all customer complaint orders as prediction targets without distinguishing between specific cause parties or responsible parties, this application effectively avoids modeling and identification errors caused by inaccurate classification of customer complaint causes or responsibilities (such as merchant responsibility or buyer fraud). This ensures the objectivity and generalization ability of the customer complaint identification model, improving the accuracy of model predictions from the source. Second, by constructing a multi-dimensional feature engineering covering orders, buyers, and merchants, it can comprehensively capture various potential factors that may lead to customer complaints. This allows the model to conduct comprehensive risk assessments based on richer and more holistic information, significantly enhancing the sensitivity of risk identification for complex customer complaint orders and reducing risk omissions caused by incomplete feature coverage. Third, by instantly visualizing identified high-risk customer complaint orders to the corresponding buyers and merchants, both parties can be aware of the risk situation at the initial stage of a dispute and may be able to communicate accordingly. This not only improves the transparency of platform services but also helps to encourage merchants to proactively improve their performance or guide buyers to adjust unreasonable expectations, resolving potential conflicts from the source and increasing the possibility and efficiency of self-resolution of disputes. It is evident that the early and proactive identification of potential customer complaints has changed the traditional, reactive model of handling complaints after the fact, thus providing the platform with a valuable window of time for risk management and effectively preventing the escalation of subsequent disputes and the expansion of losses. Attached Figure Description

[0012] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 The network architecture of the e-commerce platform exemplified in this application; Figure 2 This is a flowchart illustrating a typical embodiment of the customer complaint order identification method of this application; Figure 3 This is a schematic block diagram of the customer complaint order identification device of this application; Figure 4 This is a schematic diagram of the structure of a computer device used in this application. Detailed Implementation

[0013] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0014] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0015] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0016] like Figure 1 In the network architecture shown, the e-commerce platform 82 is deployed on the Internet to provide corresponding services to its users. Similarly, the devices 80 of the merchant users and the devices 81 of the consumer users of the e-commerce platform 82 are also connected to the Internet to use the services provided by the e-commerce platform.

[0017] An exemplary e-commerce platform 82 provides supply and demand matching of products and / or services to the general public through the Internet infrastructure. In e-commerce platform 82, products and / or services are provided as commodity information. For the sake of simplicity, the concepts of commodity and product are used in this application to refer to the products and / or services in e-commerce platform 82. Specifically, these may be physical products, digital products, tickets, service subscriptions, other offline services, etc.

[0018] In reality, various entities can access e-commerce platform 82 as users and utilize its online services to participate in the business activities facilitated by the platform. These entities can be natural persons, legal persons, or social organizations. Corresponding to the two types of entities in business activities—merchants and consumers—e-commerce platform 82 has two corresponding categories of users: merchant users and consumer users. Entities involved in the product distribution chain in business activities, including manufacturers, sellers, retailers, and logistics providers, can all use online services on e-commerce platform 82 as merchant users. Similarly, consumers in business activities, including actual or potential consumers, can use online services on e-commerce platform 82 as consumer users. In actual business activities, the same entity can operate as both a merchant user and a consumer user; this should be interpreted flexibly.

[0019] The infrastructure used to deploy the e-commerce platform 82 mainly includes the backend architecture and frontend devices. The backend architecture runs various online services through a service cluster, including middleware or frontend services for the platform, services for consumers, and services for merchants, to enrich and improve its service functions. The frontend devices mainly cover the terminal devices used by users as clients to access the e-commerce platform 82, including but not limited to various mobile terminals, personal computers, and point-of-sale devices. For example, merchant users can use their terminal device 80 to enter product information for their online stores or use the interfaces opened by the e-commerce platform to generate their product information; consumer users can use their terminal device 81 to access the webpage of the online store implemented by the e-commerce platform 82, trigger the shopping process by clicking the shopping button provided on the webpage, and call various online services provided by the e-commerce platform 82 during the shopping process to achieve the purpose of placing an order.

[0020] In some embodiments, the e-commerce platform 82 may be implemented via a processing facility including a processor and memory, which stores a set of instructions that, when executed, cause the e-commerce platform 82 to perform the e-commerce and support functions as described in this application. The processing facility may be part of a server, client, network infrastructure, mobile computing platform, cloud computing platform, fixed computing platform, or other computing platform, and may provide electronic components, merchant devices, payment gateways, application developers, marketing channels, transportation providers, customer devices, point-of-sale devices, etc., for the e-commerce platform 82.

[0021] E-commerce platform 82 can provide online services such as cloud computing services, Software as a Service (SaaS), Infrastructure as a Service (IaaS), Platform as a Service (PaaS), Desktop as a Service (DaaS), Hosted Software as a Service, Mobile Backend as a Service (MBaaS), and Information Technology Management as a Service (ITMaaS). In some embodiments, the various functional components of e-commerce platform 82 can be implemented to operate on various platforms and operating systems. For example, for an online store, its administrator user enjoys the same or similar functions regardless of whether it is on iOS, Android, HomonyOS, or a web page.

[0022] E-commerce platform 82 enables merchants to create their own independent websites to run their online stores. It provides merchants with corresponding business management engine instances, allowing them to establish, maintain, and operate one or more online stores across these independent websites. The business management engine instance can be used for content management, task automation, and data management for one or more online stores. It can be configured through interfaces or built-in components to support various specific business processes in the online store, supporting business activities. Independent websites are the infrastructure of e-commerce platform 82, which offers cross-border services. Merchants can maintain their online stores relatively independently and centrally based on these independent websites. Independent websites typically have dedicated domain names and storage space, and different independent websites are relatively independent. E-commerce platform 82 can provide standardized or customized technical support for a large number of independent websites, allowing merchants to customize a business management engine instance that suits their needs and use it to maintain one or more online stores.

[0023] Online stores can be configured and maintained in the backend by merchant users logging into their Business Management Engine instance as administrators. Supported by the various online services provided by the e-commerce platform 82's infrastructure, merchant users can configure various functions within their online stores and view various data as administrators. For example, merchant users can manage various aspects of their online stores, such as viewing recent online store activities, updating the online store's product catalog, managing orders, recent visit activity, and total order activity. Merchant users can also view more detailed information about their business and visitors to their online store by obtaining reports or metrics, such as displaying a sales summary of the merchant's overall business, specific sales and engagement data from promotional sales and marketing channels, etc.

[0024] E-commerce platforms 82 can provide communication facilities and associated merchant interfaces for electronic communication and marketing. For example, they can utilize electronic messaging aggregation facilities to collect and analyze communication interactions between merchants, consumers, merchant devices, customer devices, point-of-sale devices, etc., aggregating and analyzing communications to increase the potential for product sales. For instance, a consumer may have product-related questions, which could lead to a dialogue between the consumer and the merchant (or an automated processor-based agent representing the merchant), where the communication facilities handle the interaction and provide the merchant with analysis on how to increase the probability of a sale.

[0025] In some embodiments, applications suitable for installation on terminal devices can be provided to serve the access needs of different users, enabling various users to access the e-commerce platform 82 by running the application on their terminal devices. Examples include the merchant backend module of online stores within the e-commerce platform 82. During the process of conducting business activities through these functions, the e-commerce platform 82 can implement various functions related to business activities as middleware or online services and expose corresponding interfaces. Then, toolkits corresponding to the interface access functions are embedded into the application to achieve functional expansion and task completion. The business management engine can include a series of basic functions and expose these functions to online services and / or applications via APIs. Online services and applications use the corresponding functions by remotely calling the corresponding APIs.

[0026] With the support of various components of the Business Management Engine instance, E-commerce Platform 82 can provide online shopping functionality, enabling merchants to connect with customers in a flexible and transparent manner. Consumers can select items online, create orders, provide delivery addresses in the orders, and complete payment confirmation. Merchants can then review and complete or cancel orders.

[0027] The customer complaint order identification method of this application can be programmed into a computer program product and deployed on a client or server for implementation. For example, in an exemplary application scenario of this application, it can be deployed on the server of an e-commerce customer complaint platform. In this way, the method can be executed by human-computer interaction with the process of the computer program product through a graphical user interface by accessing the interface opened after the computer program product is running.

[0028] Please see Figure 2 The customer complaint order identification method of this application, in its typical embodiment, includes the following steps: Step S1100: Respond to the order payment event and obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features of the target order in the dimensions of order, buyer, and merchant. Responding to order payment events is an event-driven proactive monitoring mechanism. When a buyer on a cross-border e-commerce platform completes a payment transaction through the front-end visual interface of a merchant's independent website, the platform's backend server immediately captures the order payment event triggered by this action. Specifically, after the buyer selects a product, fills in shipping information, chooses a payment method, and confirms payment on the independent website's front-end interface, the buyer's client generates a payment result notification, which constitutes the trigger source for the order payment event. The server listens for payment success notifications in the payment message queue through pre-built event listeners or hook functions, or receives transaction completion messages returned by the payment gateway through the payment callback interface, thereby detecting and capturing the occurrence of order payment events in real time.

[0029] The server further responds to the order payment event, obtaining the payment order generated by this payment as the target order to determine whether it belongs to the high-risk customer complaint category. To detect the target order, the target feature set associated with that order is first obtained. This target feature set employs a three-dimensional feature system design, comprehensively depicting the data sources relied upon for customer complaint analysis from the perspectives of three business entities: order, buyer, and merchant. This includes the input features for each dimension, ensuring that customer complaint risk identification has complete contextual information. The design principle of this system is that customer complaint risk stems from the interaction between the abnormal performance of the order itself, the abnormal behavior patterns of the buyer, and the service quality of the merchant. A single-dimensional feature cannot fully and comprehensively analyze the overall risk. Input features refer to features that, after feature engineering screening, are confirmed to possess inference importance and distribution stability. These features are used by the customer complaint identification model for order identification inference, effectively distinguishing the customer complaint risk of an order.

[0030] Step S1200: Use a customer complaint identification model to identify the customer complaint risk probability of the target order based on the target feature set, and determine whether the customer complaint risk probability exceeds the risk threshold. The customer complaint identification model is a classification model based on machine learning algorithms. This model uses supervised learning to map the relationship between historical orders and customer complaints, enabling it to predict the probability of a customer complaint for a new target order. The customer complaint risk probability is a value between 0 and 1, representing the likelihood of a customer complaint occurring in the future; a higher probability value indicates a greater risk. During the training phase, the model uses a large number of labeled samples. Each sample contains multiple input features from historical orders across order, buyer, and merchant dimensions, and each sample is labeled with a binary label indicating whether a customer complaint has actually occurred for that order. The model adjusts its internal parameters through optimization algorithms to minimize the difference between the predicted probability and the true label. For example, it uses cross-entropy as the loss function and gradient descent for iterative updates.

[0031] Various machine learning algorithms can be used to select the appropriate customer complaint identification model. For example, logistic regression models map linear combination feature values ​​to probabilities using the sigmoid function; this model is computationally efficient and highly interpretable. Decision tree models segment features using a series of if-then rules, directly outputting the proportion of customer complaint samples at leaf nodes as probabilities. Random forest models integrate multiple decision trees, improving prediction stability through voting or averaging. Gradient boosting tree models, such as XGBoost or LightGBM, build strong predictors by progressively correcting the residuals of previous trees. Neural network models learn nonlinear combinations of features using multilayer perceptrons, making them suitable for high-dimensional feature interactions. Each model requires parameter learning using a training set during training, and hyperparameters are adjusted using a validation set to prevent overfitting. Finally, performance metrics such as AUC or F1 score are evaluated using a test set.

[0032] The customer complaint identification model performs forward propagation inference calculations on the target feature set and outputs the customer complaint risk probability. The recommended embodiment uses the gradient boosting tree model LightGBM for customer complaint identification. During the inference calculation, the model traverses each tree sequentially: starting from the root node, it compares the value of the current input feature with the node's preset splitting threshold to determine whether it falls into the left or right subtree, until it reaches a leaf node. This leaf node contains a weight score. The weight scores of all leaf nodes are summed, and then the sum is mapped using the sigmoid function, ultimately outputting a value between 0 and 1 representing the customer complaint risk probability.

[0033] After the customer complaint identification model outputs a probability, the server further compares this probability with a preset risk threshold. The risk threshold is a configurable value, typically determined based on business needs by balancing the selection costs of false positives and false negatives. Those skilled in the art can flexibly configure it according to the instructions disclosed here; for example, a risk threshold of 0.7 means that an order is considered high-risk when the probability exceeds 0.7. During the judgment process, if the customer complaint risk probability is greater than the risk threshold, the subsequent high-risk processing procedure is triggered; otherwise, the target order is considered a relatively low-risk order and requires no further processing. In other embodiments, the risk threshold can also be determined based on model performance, selecting the optimal threshold point that maximizes precision and recall.

[0034] Step S1300: When the customer complaint risk probability exceeds the risk threshold, the target order is marked as a high customer complaint risk order, and the high customer complaint risk order is visualized and displayed to the corresponding buyer and merchant.

[0035] Once the labeling is complete, information about orders with high customer complaint risk will be displayed in real-time to the corresponding buyers and merchants through the front-end user interface. For buyers, the display occurs on the order details page, such as adding a prominent warning icon (e.g., a red exclamation mark) and text prompts (e.g., "This order has a high risk of customer complaints; we recommend handling it with caution") to the order status area, and may also provide a link to a detailed risk explanation. For merchants, the display is integrated into the order management list in the merchant's backend, such as highlighting the risk label next to the order line item, and supporting filtering and sorting of high-risk orders so that merchants can prioritize their handling.

[0036] In other embodiments, the specific form of the visualization can be diverse, including but not limited to embedding dynamic prompt components in web pages, generating push notifications to mobile client applications of buyers or merchants, or sending risk alert summaries via email. For example, a floating banner can be added to the order confirmation page on the buyer's side to display risk warning information; a dedicated panel can be added to the order processing interface on the merchant's side to summarize and display all high-risk orders. This display mechanism ensures that both buyers and sellers can perceive risks in a timely manner and decide on subsequent actions based on the visualized information. For example, buyers can choose to cancel orders or contact the customer service of independent websites or e-commerce platforms, and merchants can strengthen order verification or proactively communicate, thereby reducing the actual customer complaint rate.

[0037] As can be understood from the disclosure of this exemplary embodiment, this application contains many advantages, including but not limited to the following aspects: First, by treating all customer complaint orders as prediction targets without distinguishing between specific cause parties or responsible parties, this application effectively avoids modeling and identification errors caused by inaccurate classification of customer complaint causes or responsibilities (such as merchant responsibility or buyer fraud). This ensures the objectivity and generalization ability of the customer complaint identification model, improving the accuracy of model predictions from the source. Second, by constructing a multi-dimensional feature engineering covering orders, buyers, and merchants, it can comprehensively capture various potential factors that may lead to customer complaints. This allows the model to conduct comprehensive risk assessments based on richer and more holistic information, significantly enhancing the sensitivity of risk identification for complex customer complaint orders and reducing risk omissions caused by incomplete feature coverage. Third, by instantly visualizing identified high-risk customer complaint orders to the corresponding buyers and merchants, both parties can be aware of the risk situation at the initial stage of a dispute and may be able to communicate accordingly. This not only improves the transparency of platform services but also helps to encourage merchants to proactively improve their performance or guide buyers to adjust unreasonable expectations, resolving potential conflicts from the source and increasing the possibility and efficiency of self-resolution of disputes. It is evident that the early and proactive identification of potential customer complaints has changed the traditional, reactive model of handling complaints after the fact, thus providing the platform with a valuable window of time for risk management and effectively preventing the escalation of subsequent disputes and the expansion of losses.

[0038] In a further embodiment, before step S1100, which responds to the order payment event, the following steps are included: Step S1000: Collect multiple historical customer complaint orders and historical non-customer complaint orders within a preset customer complaint duration, determine multiple candidate features corresponding to each order in the dimensions of order, buyer, and merchant, and construct a historical feature set; The preset customer complaint period refers to the time interval between order payment completion and customer complaint initiation. It covers the vast majority of complaint orders. For example, a 60-day period determined based on historical customer complaint data from the platform covers over 90% of complaint orders where complaints were initiated within 60 days of payment. Historical complaint orders refer to orders where complaints were actually initiated within the preset period, such as when a buyer initiated a dispute or compensation claim through a third-party payment platform or e-commerce platform within 60 days of order payment. Historical non-complaint orders refer to orders where no complaints were actually initiated, such as when no complaint was initiated by the corresponding buyer within 60 days of order payment.

[0039] For each order, multiple candidate features are identified at the order, buyer, and merchant dimensions. Candidate features refer to variables extracted and / or derived from the original order, buyer, and merchant data that can be used for supervised training and model identification of customer complaint risks. These include any one or more of numerical, categorical, and statistical features. Constructing a historical feature set means organizing the candidate features for the corresponding order in a structured format to adapt to the model's input format requirements. Specifically, in a further embodiment, this includes: Step S1001: For each order, obtain the basic attribute features of the order, and determine the statistical features and product association features of the order from historical order data based on at least one preset time window; The basic attributes of an order include, but are not limited to: total order amount, order payment discount rate, order shipping fee, order shipping fee as a percentage of the total order amount, total number of items in the order, unique identifier of each item in the order, category of each item in the order, listing duration of each item in the order (time difference between the listing timestamp and the current timestamp), order payment duration (time difference between the payment completion timestamp and the payment start timestamp), order completion duration (time difference between the payment completion timestamp and the order creation timestamp), order payment IP, country and / or city of the order payment IP, and whether the order contains ultra-low-priced, prohibited, or infringing items.

[0040] Statistical characteristics of orders include, but are not limited to: the average total order amount, average order payment duration, and average order completion duration for the order payment date / 3 days / 5 days / 7 days / 15 days prior to the date; the average total order amount, average order payment duration, average order completion duration, total number of customer complaint orders, and their corresponding cumulative total order amount for the same payment IP address within 1 hour / 3 hours / 6 hours prior to order payment / on the same day / 3 days / 5 days / 7 days / 15 days prior to the date; The product association characteristics of an order include, but are not limited to: the total number of orders containing the same product within 1 hour / 3 hours / 6 hours before the order payment, the total number of customer complaint orders, and their corresponding total order amount within the same day / 3 days / 5 days / 7 days / 15 days before the order payment.

[0041] Step S1002: Obtain the basic attribute features and blacklist association features of the buyer corresponding to the order, and determine the buyer's behavioral statistical features from the buyer's historical behavior data based on at least one preset time window; The buyer's basic attributes include, but are not limited to: the buyer's unique identifier (e.g., ID), device fingerprint, mobile phone number, email address, and shipping address.

[0042] The blacklist association features for buyers include, but are not limited to: the buyer's unique identifier (e.g., ID), device fingerprint, mobile phone number, email address, and whether the shipping address is on the blacklist; and the text similarity between the buyer's email address, mobile phone number, and shipping address and the most similar email address, mobile phone number, and shipping address on the blacklist, respectively. Taking email address as an example, the same applies to mobile phone number and shipping address. An open-source pre-trained convergent text encoding model can be used to encode the buyer's email address and each email address on the blacklist, and then the cosine similarity between the text encoding vector of the buyer's email address and the text encoding vector of each email address on the blacklist can be calculated as the email text similarity.

[0043] Buyer behavior statistics include, but are not limited to: the total number of orders generated by the buyer's unique identifier (e.g., ID) / device fingerprint / phone number / email within the last 1 hour / 3 hours / 6 hours / day / last 3 days / last 7 days / last 15 days and their corresponding cumulative order total amount, the total number of customer complaint orders and their corresponding cumulative order total amount, and the ratio obtained by dividing the total number of customer complaint orders by the total number of orders.

[0044] Step S1003: Obtain the basic attribute features and blacklist association features of the merchant and its store corresponding to the order, and determine the merchant's business statistics features from the merchant's historical business data based on at least one preset time window.

[0045] It's easy to understand that a merchant can operate multiple independent online stores. These stores may have different opening dates and / or different types of business. Furthermore, some merchants may intentionally open multiple independent online stores to mitigate risk, abandoning one store if it experiences operational difficulties and switching to another. Therefore, to comprehensively consider customer complaint risks on the merchant's side, it's necessary to consider various characteristics of the individual and all their stores, including: The basic attributes of each merchant's store include, but are not limited to: the number of days the store has been registered, the total number of products listed, the latest new product launch time (the time difference between the timestamp of the most recent product launch and the current timestamp), and the latest transaction time (the time difference between the most recent order and the current timestamp).

[0046] The basic attributes of a merchant include, but are not limited to: the total number of stores under the merchant's name, the total number of stores in operation, and the total number of closed stores.

[0047] The characteristics associated with a merchant's blacklist include, but are not limited to: the total number of penalized stores under that merchant's name on the blacklist.

[0048] The blacklist association features for each merchant's store include, but are not limited to: whether the store is a penalized store on the blacklist, and the time since the latest penalty (the time difference obtained by subtracting the previous penalty timestamp from the current timestamp).

[0049] The operational statistical characteristics of merchants include, but are not limited to: the total number of orders generated by each store in the most recent 1 day / 3 days / 7 days / 15 days / 30 days and their corresponding cumulative total order amount, the total number of buyers, the total number of products sold, the maximum single order sales amount in a single day and its time elapsed since then; the total number of customer complaint orders generated by each store in the most recent 1 day / 3 days / 7 days / 15 days / 30 days and their corresponding cumulative total order amount; and the total number of customer complaint orders generated by each store in the most recent 7 days / 15 days / 30 days. The ratio of the total number of customer complaint orders to the total number of orders generated; the total number of orders generated by the merchant in the most recent 1 day / 3 days / 7 days / 15 days / 30 days and their corresponding cumulative order total amount; the total number of stores with transactions; the average total number of orders with transactions across all stores; the average cumulative order total amount with transactions across all stores; the maximum total number of orders with transactions across all stores; the maximum cumulative order total amount with transactions across all stores; the minimum total number of orders with transactions across all stores; and the minimum cumulative order total amount with transactions across all stores.

[0050] Step S1010: Take the historical feature set of each historical customer complaint order as the first sample, and mark the supervision label of each first sample to indicate that the corresponding order has initiated a customer complaint. At the same time, take the historical feature set of each historical non-customer complaint order as the second sample, and mark the supervision label of each second sample to indicate that the corresponding order has not initiated a customer complaint. The supervision label for the first sample can be 1 in the binary label, which indicates that a customer complaint has been initiated for the corresponding order; the supervision label for the second sample can be 0 in the binary label, which indicates that a customer complaint has not been initiated for the corresponding order.

[0051] Step S1020: Construct a dataset using all the first samples and their supervisory labels and the second samples and their supervisory labels. Take a portion of the samples in the dataset within the preset recent historical payment period as test samples, divide the remaining samples into training samples and validation samples according to a preset ratio, associate each type of sample with its supervisory label, and construct the test set, training set, and validation set accordingly. The most recent historical payment period is relatively new compared to other payment periods within the historical payment time range. The specific duration of this period can be configured as needed by those skilled in the art. For example, 2 months would correspond to the most recent historical payment period as [November 1, 2024 to December 31, 2024].

[0052] The preset ratio on which the training and validation samples are divided can be configured as needed by those skilled in the art, for example, 7:3. This means that 70% of the remaining samples are divided into training samples and 30% into validation samples. It can be seen that each sub-ratio in the preset ratio is the ratio of the number of samples of the corresponding class to the number of remaining samples.

[0053] Furthermore, after constructing the training set and validation set, for each set, the ratio obtained by dividing the total number of first samples in the set by the total number of all samples is determined. When this ratio is less than a preset ratio (e.g., 1%, which can be set by those skilled in the art as needed), it indicates that the proportion of the first sample in the total is too low, and the proportion distribution of the first sample and the second sample in the set is seriously unbalanced and deviates too much from the actual proportion distribution. Therefore, it is necessary to randomly sample some second samples under different payment periods to remove them, so that the proportion of the first sample in the total increases after removal and matches the actual proportion distribution. For example, the proportion of the first sample in the total is in the range of 3% to 5%, and the proportion of the second sample in the total is in the range of 95% to 97%.

[0054] Step S1030: For each candidate feature, determine the candidate distribution difference value between the test set and the training set. When the distribution difference value reaches the stable distribution threshold, it is determined as an input feature; otherwise, the candidate feature is removed from the test set, training set, and validation set.

[0055] In the feature selection process of the customer complaint order identification method, to ensure the distribution stability of the selected features on both the training and test sets, thereby guaranteeing the model's generalization ability, the Population Stability Index (PSI) is used as a specific quantitative indicator of the difference in candidate distributions. PSI is a statistical indicator that measures the difference between two probability distributions, evaluating feature stability by comparing the frequency distribution differences of feature values ​​in the training and test sets. The calculation of this index is based on binning of feature values, and the binning method varies depending on the type of feature data.

[0056] For numerical candidate features, such as order amount or payment duration, equal-width binning or equal-frequency binning methods are used for discretization. Equal-width binning divides the feature value range into multiple equally wide intervals; for example, order amount from $0 to $1000 is divided into 10 intervals, each with a width of $100. Equal-frequency binning ensures that each bin contains approximately the same number of samples; for example, samples are sorted by order amount and divided into 10 groups, each containing 10% of the samples. For categorical candidate features, such as the country of payment IP or product category, each category is directly used as a bin, without additional discretization.

[0057] When calculating the PSI, the binning boundaries or class groups of features are first determined based on the training set as the baseline distribution. For each bin, the sample frequency of that bin in the training set is calculated as the expected percentage, and the sample frequency of the corresponding bin in the test set is calculated as the actual percentage. The PSI value is calculated using the formula ∑(actual percentage - expected percentage) × ln(actual percentage / expected percentage), where the summation operation covers all bins. For example, assuming a feature is divided into 5 bins in the training and test sets, with bin frequencies of [0.2, 0.3, 0.25, 0.15, 0.1] in the training set and [0.25, 0.28, 0.22, 0.18, 0.07] in the test set, the PSI components of each bin are (0.05 × ln(0.25 / 0.2)), (-0.02 × ln(0.28 / 0.3)), etc., and the sum of these components yields the total PSI value.

[0058] A higher PSI value indicates a greater difference in feature distribution between the training and test sets. This difference may stem from data collection bias, changes in business rules, or sample selection bias, leading to inconsistent feature distributions. For example, if a feature's PSI value exceeds a stable distribution threshold (typically set to 0.25, but can be set differently by those skilled in the art), it means that the feature's distribution on the test set has shifted significantly compared to the training set. This shift prevents the patterns learned by the model during training from being applicable to the test set, meaning they cannot generalize to real-world online recognition scenarios, thus causing overfitting or a decline in generalization performance. Therefore, when a PSI value reaches or exceeds the threshold, the candidate feature is considered unstable and must be removed from the training, test, and validation sets to eliminate the interference of distribution differences on model training. Conversely, features with PSI values ​​below the threshold (e.g., less than 0.1) are considered stable and retained as model features. This mechanism ensures that the model is trained only on features with consistent distributions, improving the model's predictive reliability and interpretability on unknown data.

[0059] In this embodiment, historical customer complaint orders and non-customer complaint orders within a preset customer complaint duration are collected, and a historical feature set covering three dimensions—order, buyer, and merchant—is constructed to ensure the comprehensiveness and representativeness of the training data. This enables the model to learn multi-dimensional and multi-perspective risk patterns, avoiding model bias caused by incomplete feature coverage. More importantly, the PSI index is used to quantify the distribution difference of each candidate feature between the training and test sets, and unstable features (PSI values ​​exceeding a threshold) are removed, effectively preventing model overfitting and ensuring the stability of the input features over time. This allows the model to maintain reliable predictive performance when facing new data in the future, thereby significantly improving the model's generalization ability and recognition accuracy in actual deployment. Furthermore, during this period, the preset customer complaint duration (e.g., 60 days) covers the vast majority of customer complaint orders and ensures the consistency between the training data and online data distribution. In terms of data preprocessing, this reduces the risk of model performance degradation due to changes in business rules or data drift, laying a solid foundation for the long-term stable operation of the model.

[0060] In a further embodiment, step S1030, for each candidate feature, determines the candidate distribution difference value between the test set and the training set. When the distribution difference value reaches a stable distribution threshold, it is determined as an input feature; otherwise, after removing the candidate feature from the test set, training set, and validation set, the following steps are included: Step S1040: Train the customer complaint identification model to a convergent state using a test set, training set, and validation set, and obtain the importance index of each input feature of the model, as well as the test performance index of the model. In implementing the customer complaint order identification method, the customer complaint identification model first needs to be trained to a convergent state. At this point, the model learns the mapping relationship between orders and customer complaint risks on the training set, and adjusts the hyperparameters on the validation set to prevent overfitting. After the model converges, the importance index of each input feature is obtained. This index quantifies the contribution of the feature to the model's prediction results. For example, in a model based on gradient boosting tree algorithms (such as LightGBM), feature importance can be obtained by calculating the sum of information gains brought by the feature when splitting at a tree node. At the same time, the model's performance index on the test set is calculated, such as the area under the curve (AUC) or the Kolmogorov-Smirnov statistic (KS), to evaluate the model's discriminative ability.

[0061] Step S1041: In the training set, validation set, and test set, remove the input feature with the lowest importance index in each sample. Based on the removed training set and validation set, retrain the customer complaint identification model to a convergent state, and determine the current performance index of the model based on the removed test set. Next, the feature with the lowest importance metric in each sample is removed from the training, validation, and test sets. For example, if the feature importance list shows that the importance score of "City of Order Payment IP" is much lower than other features (such as "Order Amount" or "Buyer's Historical Customer Complaint Rate"), then this feature is removed from all samples. Based on the removed training and validation sets, the customer complaint identification model is retrained to convergence, and the performance metric for this round is calculated using the removed test set. This process aims to simplify the model's input feature structure, reduce the computational overhead of redundant features, and maintain model performance.

[0062] Step S1042: When the difference between the test performance index and the current round performance index does not exceed the tolerance value, return to the previous step to continue execution; otherwise, obtain the customer complaint identification model from the previous round and use it to identify the customer complaint risk of the order based on the feature set composed of the remaining input features corresponding to the order.

[0063] Finally, the difference between the initial test performance metric and the current round performance metric is compared to see if it does not exceed a preset tolerance value. The tolerance value is a configurable threshold, such as an AUC difference not exceeding 0.01, used to measure the impact of feature removal on model performance. If the difference is within the tolerance range, it indicates that the model performance has not significantly degraded, and within a certain allowable error range, the process returns to the previous step to continue removing the least important features in the current round for the next iteration. If the difference exceeds the tolerance value, the iteration stops, and the customer complaint identification model trained in the previous round is used. This model identifies customer complaint risks based on the feature set composed of the remaining input features. This ensures that the model maintains prediction accuracy while optimizing the feature set and improving computational efficiency. For example, in specific implementations, the tolerance value can be flexibly set according to business needs, such as setting a tolerance of 0.005 for high-precision scenarios and 0.02 for general scenarios.

[0064] In this embodiment, the feature set is gradually simplified by calculating the importance index (e.g., based on information gain) for each input feature and iteratively removing the least important features. This reduces the computational complexity and resource consumption during model inference, lowers data processing and storage overhead, and thus improves the response speed of online risk identification, making it particularly suitable for e-commerce platform scenarios with high concurrency and real-time requirements. The model is also more lightweight while maintaining performance, which is beneficial for deployment in resource-constrained environments.

[0065] In a further embodiment, step S1300, when the customer complaint risk probability exceeds the risk threshold, marks the target order as a high customer complaint risk order and visualizes the high customer complaint risk order to the corresponding buyer and merchant, includes the following steps: Step S1400: Obtain the customer complaint risk probability and target feature set of each target order identified in two preset monitoring periods; In customer complaint order identification methods, continuous monitoring after model deployment is crucial for ensuring its long-term effectiveness. This monitoring mechanism is based on two preset monitoring periods, such as adjacent time periods on a weekly or monthly basis. For example, the first monitoring period is from November 1st to November 30th, 2026, and the second monitoring period is from December 1st to December 31st, 2026. The server obtains the customer complaint risk probability and its corresponding target feature set for all target orders processed by the customer complaint identification model within these two periods. The target feature set includes input features from the order dimension, buyer dimension, and merchant dimension. These features have been selected during the model training phase to possess stability and discriminative power.

[0066] Step S1410: Based on the target feature sets of the target orders in the two monitoring periods, determine the difference in the distribution of each input feature in the target feature set between the two monitoring periods. Next, the population distribution difference value is calculated, which quantifies the degree of distribution change of each population feature between two monitoring periods. Specifically, the Population Stability Index (PSI) is used as the calculation indicator for the difference value. PSI assesses stability by comparing the frequency distribution differences of feature values ​​between two periods. For numerical features, such as order amounts, binning is first performed, for example, using equal-width bins to divide the amount into ranges such as $0-100 and $100-200. For categorical features, such as the country of the payment IP, each country is directly used as a bin. During calculation, the feature distribution in the first monitoring period is used as the baseline (expected distribution), and the distribution in the second monitoring period is used as the comparison object (actual distribution). The PSI value calculation formula is: ∑(actual percentage - expected percentage) × ln(actual percentage / expected percentage), where the summation covers all bins. For example, if a feature is distributed as follows: bin A accounts for 30% and bin B accounts for 70% in the first time period, and then changes to bin A accounting for 25% and bin B accounting for 75% in the second time period, then PSI = (0.25-0.3)×ln(0.25 / 0.3) + (0.75-0.7)×ln(0.75 / 0.7) ≈0.006. A higher PSI value indicates a greater change in the feature distribution; a threshold such as 0.25 is typically set, and when the PSI exceeds this value, the feature distribution is considered unstable.

[0067] Step S1420: Based on the customer complaint risk probability of the target orders in the two monitoring periods, determine the difference in the inference distribution of the recognition probability of the customer complaint recognition model between the two monitoring periods. Simultaneously, the inference distribution difference (PSI) value is calculated, which measures the overall distribution change of the model's output probability. The PSI method is also used, but the object is the customer complaint risk probability value. The probability value range, such as 0 to 1, is divided into multiple bins (e.g., 10 equal-width bins), and the proportion of orders with probability values ​​falling into each bin is counted in both monitoring periods. The PSI value is calculated to assess the shift in the probability distribution. For example, if orders in the bin with a probability of 0.8-1.0 account for 5% in the first period and rise to 10% in the second period, the PSI value will increase, indicating a change in the model's predictive behavior.

[0068] Step S1430: When the difference value of the input distribution or the difference value of the inference distribution indicates that the model performance is not up to standard, obtain the customer complaint identification model that has been retrained to the convergence state, and perform hot update on the substandard customer complaint identification model.

[0069] When the difference between the input and inference distributions exceeds a preset threshold, the performance of the characterization model is deemed substandard. Excessive distribution discrepancies may stem from changes in the business environment, such as changes in the transaction market environment. In this case, a model update process is triggered: the latest historical order data is retrieved from the data warehouse, feature engineering and model training are re-executed, resulting in a converged new customer complaint identification model. During hot updates, the server maintains dual model instances: an online service model and a model awaiting update. Seamless switching is achieved through version control middleware: once the new model's performance meets offline verification standards, the routing configuration is updated to redirect a portion of traffic (e.g., 10%) to the new model for A / B testing. If the new model performs stably for 24 consecutive hours, a full switch is completed. The entire process achieves a user-unnoticed model update, ensuring the continuity and accuracy of the customer complaint identification service.

[0070] In the recommended embodiment, to achieve predictive model updates, a prediction mechanism based on performance degradation trends needs to be established. The core of this mechanism lies in quantifying the decay acceleration of the model stability index and calculating the critical time point accordingly. The specific implementation process is as follows: First, continuously record the input distribution difference (feature PSI) and inference distribution difference (probability PSI) for each monitoring period (e.g., daily or weekly). For each stability index, store at least three periods of observations in a continuous time series format, such as a weekly PSI value sequence for four consecutive weeks [PSI1, PSI2, PSI3, PSI4]. Based on these discrete observation points, calculate the decay acceleration using numerical differentiation. Specifically, use the central difference formula to calculate the instantaneous rate of change: for the acceleration a_t at time point t, the formula is a_t = (PSI{t+1} - 2*PSI_t + PSI{t-1}) / Δt 2Where Δt is the monitoring period interval. For example, if the weekly PSI value sequence is [0.08, 0.15, 0.22, 0.30], then the acceleration a3 in the third week is (0.30 - 2 * 0.22 + 0.15) / 1 2 = 0.01.

[0071] Secondly, the time to reach the preset threshold is predicted based on the current decay rate and acceleration. A linear decay model or an exponential decay model is established for trajectory prediction. Taking the linear model as an example: Let the PSI value at the current time t0 be y0, the current instantaneous velocity v0 be calculated using the first-order difference (v0 = (y0 - y_{t-1}) / Δt), and the acceleration a0 be obtained using the central difference method mentioned above. Then, the time Δt_pred to reach the threshold Y_threshold satisfies the formula: Y_threshold = y0 + v0Δt_pred + 0.5a0*Δt_pred 2 Solving this quadratic equation yields the predicted time. For example, if the current week's PSI is 0.25 and last week's was 0.20, the acceleration is 0.01 / week. 2 If the threshold is set to 0.25, the predicted time to reach the threshold is 0 weeks; if the threshold is 0.30, the predicted time to reach the threshold is approximately 0.915 weeks.

[0072] When the predicted time is less than the preset buffer period (e.g., one week, three days, etc., which can be configured as needed by those skilled in the art), the data preparation and training process is automatically triggered. This process first defines the data time window for retraining: using the current time as the endpoint, a dataset with a time span equal to the original training set period is selected (e.g., data from the past 12 months). Next, after filtering out unstable feature data from this dataset, the training set, validation set, and test set are divided using the same training method as before to retrain a converged customer complaint recognition model. The model's discriminative feature set is then selected for use in hot updates.

[0073] In this embodiment, by periodically calculating the input feature distribution difference value (feature PSI) and the model output probability distribution difference value (inference PSI), data distribution drift and model performance degradation can be detected in real time. Once the difference value indicates that the model has degraded to the point where it cannot be used for online inference, a hot update is automatically triggered by calling the retrained model based on the inference prediction at that time point. This ensures dynamic adaptation to real-time changes in the business environment, avoids the lag of traditional passive maintenance, and guarantees the continuous accuracy and reliability of the recognition service.

[0074] In a further embodiment, step S1300, when the customer complaint risk probability exceeds the risk threshold, marks the target order as a high customer complaint risk order and visualizes the high customer complaint risk order to the corresponding buyer and merchant, includes the following steps: Step S1500: Perform payment risk detection on the high customer complaint risk orders to determine the corresponding payment risk probability that represents a non-buyer-initiated transaction; Payment risk detection is performed on orders with high customer complaint risk by analyzing payment-related characteristics to determine the probability of non-buyer-initiated transactions, i.e., payment risk probability. Payment risk detection is based on a machine learning model that uses supervised learning of historical payment data. In one embodiment, the input features of this model include, but are not limited to, the geographic location anomaly of the payment IP, the unfamiliarity of the payment device fingerprint, the deviation of payment time from buyer habits, the comparison of payment amount with historical transactions, and the risk score of the payment channel. For example, when the country of the payment IP is inconsistent with the buyer's commonly used login country, the feature value indicates high risk; when the device fingerprint appears for the first time or matches a blacklisted device, the risk increases. The payment risk probability output is a value between 0 and 1, with higher values ​​indicating a greater likelihood of involuntary transactions, such as fraudulent transactions or account theft. In specific implementations, the payment risk model can use logistic regression or gradient boosting trees. The input features are standardized, and the model is trained using labeled fraudulent transaction samples and non-fraudulent samples, optimizing the cross-entropy loss function.

[0075] Step S1510: Based on the logistics and delivery information of the high customer complaint risk order, determine whether the buyer corresponding to the order has received the goods in the order; Based on the logistics and delivery information of orders with high customer complaint risk, the system determines whether the buyer has received the goods. Logistics and delivery information is obtained from an integrated logistics API, including the courier company tracking number, delivery status update timestamp, and delivery status indicator. The server periodically polls the logistics API or receives webhook pushes to parse the delivery status, such as "shipped," "in transit," or "delivered." When the logistics record shows a status of "delivered" and the recipient information matches the buyer, it is determined that the buyer has received the goods; if the status is "delivery failed" or there are no updates for a long time, it is determined that the goods have not been received. For example, the logistics API returns a delivery confirmation including the delivery time and the recipient's name; the server compares the buyer's name and address to verify consistency.

[0076] Step S1520: When the payment risk probability does not exceed a preset threshold and the buyer receives the goods, the receiving matching degree between the buyer's real-shot image of the goods and the corresponding product image and text information, as well as the receiving integrity degree of the goods in the buyer's real-shot image, are pushed to the buyer and merchant corresponding to the high customer complaint risk order. At this point, the matching degree between the buyer's actual photos of the product and the product's text and image information, as well as the product's condition upon receipt as shown in the buyer's actual photos, are calculated. Specifically, a neural network model can be used to compare the buyer's uploaded photos of the received product with the product's text and image information to assess whether the product description matches the actual item, quantifying the matching degree. Specifically, the neural network model here is a dual-tower model. One tower takes the actual photo as input and performs image encoding to output a corresponding image-encoded vector. The other tower takes the product's text and image information as input and performs image-text fusion encoding to output a corresponding image-text encoded vector. During the image-text fusion encoding process, the product text information in the text and image information is text-encoded to obtain a corresponding text-encoded vector, and the product image information is image-encoded to obtain a corresponding image-encoded vector. These two vectors are in the same semantic space, and then the text-encoded vector and the image-encoded vector are fused together to output an image-text encoded vector. Finally, the cosine similarity between the image-encoded vector and the image-text fusion vector is calculated as the matching degree.

[0077] The system can use an object detection model to identify whether goods in real-life photos are damaged, such as scratches or deformation, and output a score indicating the percentage of goods in good condition as the receipt condition score. The matching degree and condition score calculation results are pushed to the application interfaces of both the buyer and the merchant via a message queue. For example, the order details page may display "Goods matching degree 85%, condition score 90%" for both parties to refer to and confirm the goods' status. Preset thresholds can be configured by those skilled in the art according to business needs. It is understood that if the receipt matching degree and / or receipt condition score are too low and fail to meet the standards, the responsibility for any customer complaint should fall on the merchant; conversely, if the receipt matching degree and receipt condition score are high enough to be marked as good, the responsibility for any customer complaint should fall on the buyer.

[0078] Those skilled in the art can flexibly adapt the neural network model and object detection model as disclosed herein, but this step will not be described in detail.

[0079] Step S1530: When the payment risk probability exceeds a preset threshold or the buyer does not receive the goods, construct a corresponding customer complaint analysis report and push it to the buyer and merchant corresponding to the high customer complaint risk order.

[0080] At this point, a customer complaint analysis report is automatically generated. Furthermore, it can be understood that if the responsibility for this customer complaint does not lie with the buyer and merchant, but with the logistics provider or a transaction not initiated by the buyer, the report primarily aggregates data such as payment risk detection results and logistics status, using a template engine to populate the risk summary. The report content can include the payment risk level, the reason for the logistics anomaly, and suggested actions, such as "Payment IP is abnormal; it is recommended that the merchant contact the buyer for verification." For cases of non-receipt, the report includes a logistics tracking link. The report is pushed to both the buyer and seller via email or in-app notification. In the recommended implementation, interactive elements such as jump links, buttons, and windows are added to the notification information for direct contact with customer service.

[0081] In this embodiment, by assessing the probability of involuntary transactions through payment risk detection and confirming the delivery status of goods in conjunction with logistics information, the potential root causes of customer complaints can be distinguished (such as buyer fraud, merchant responsibility, logistics issues, and transactions not initiated by the buyer). When the payment risk is low and the goods have been delivered, the matching degree and integrity of the goods are further calculated, providing objective evidence for both the buyer and seller and helping to clarify the attribution of responsibility. By visually pushing the matching degree, integrity, or customer complaint analysis report to both the buyer and seller, this embodiment enhances information transparency, enabling both parties to make data-driven decisions. For example, a high matching degree and integrity can prompt the buyer to confirm the status of the goods, reducing unreasonable customer complaints; while when the payment risk is high or the goods have not been received, an automatic report is generated to facilitate rapid action. This promotes early communication, reduces the escalation of disputes from the source, and improves the efficiency of platform dispute resolution.

[0082] In a further embodiment, step S1300, when the customer complaint risk probability exceeds the risk threshold, marks the target order as a high customer complaint risk order and visualizes the high customer complaint risk order to the corresponding buyer and merchant, includes the following steps: Step S1600: Evaluate the buyer's shopping credit score based on the blacklist association characteristics and historical behavior characteristics of the buyer corresponding to the high customer complaint risk order; Blacklist association features include the buyer's unique identifier, device fingerprint, mobile phone number, email address, and shipping address, and their similarity to records in the platform's blacklist database. For example, a text encoding model can be used to calculate the cosine similarity between the buyer's email address and blacklisted email addresses. Historical behavioral features refer to explicit or implicit violations in the buyer's past shopping behavior. Explicit behaviors include historical customer complaints, return / exchange frequency, and order cancellation rate; implicit behaviors include abnormal shopping times, frequent changes of payment devices, and frequent changes of shipping addresses. When evaluating shopping credit scores, a machine learning model is used. The blacklist association features and historical behavioral features are input, and a score from 0 to 100 is output. It can be understood that the more confidently these features are in determining that the corresponding buyer has violated shopping rules, the lower the corresponding shopping credit score; conversely, the lower the features, the higher the shopping credit score. Those skilled in the art can flexibly adapt the methods disclosed herein to achieve data collection and train the machine learning model to a convergent state; these steps are not detailed here.

[0083] Step S1610: Evaluate the seller's business credit score based on the blacklist association characteristics and historical business characteristics of the sellers corresponding to the high customer complaint risk orders; The blacklist association features include the total number of stores under a merchant's name that have been penalized, whether a store is on the blacklist, and the time since the most recent penalty. Historical operating characteristics refer to explicit or implicit violations in a merchant's business behavior. Explicit violations include historical customer complaint rates, store closure frequency, and product infringement records; implicit violations include abnormal sales fluctuations, frequency of new store openings, and logistics delay rates. When assessing the business credit score, a machine learning model is also used. The blacklist association features and historical operating characteristics are input, and a score from 0 to 100 is output. It can be understood that the more confidently these features are in determining that the corresponding merchant is operating illegally, the lower the corresponding business credit score; conversely, the lower the features, the higher the business credit score. Those skilled in the art can flexibly adapt the methods disclosed herein to achieve data collection and train the machine learning model to a convergent state; these steps are not detailed here.

[0084] Step S1620: When the shopping credit score exceeds the first credibility threshold and the business credit score does not exceed the second credibility threshold, initiate the customer complaint initiation process for the buyer corresponding to the high customer complaint risk order.

[0085] When a buyer's shopping credit score exceeds the first credibility threshold and the seller's business credit score does not exceed the second credibility threshold, the customer complaint initiation process is automatically initiated for the buyer corresponding to the high-risk order. The first and second credibility thresholds are preset values ​​that can be set by those skilled in the art according to business needs. For example, a first credibility threshold of 90 indicates good buyer credit, and a second credibility threshold of 60 indicates poor seller credit. The initiation process includes generating a customer complaint form with pre-filled order information, redirecting to the customer complaint submission interface, or directly triggering a customer service ticket. For example, if the buyer's score is 85 (higher than 80) and the seller's score is 50 (lower than 60), the buyer is deemed trustworthy while the seller is considered high-risk, thus automatically opening the customer complaint channel and reducing the buyer's operational steps. After the process is initiated, the buyer can directly and quickly submit a complaint, and the seller is simultaneously notified to handle it, improving dispute resolution efficiency.

[0086] In this embodiment, by evaluating the buyer's shopping credit score and the merchant's business credit score, and setting a credibility threshold, a customer complaint channel is automatically opened for the buyer when the buyer's credit is high and the merchant's credit is low. This allows for faster rights protection for the buyer, speeds up dispute resolution, improves the buyer's shopping experience, and helps reduce risk handling costs.

[0087] Please see Figure 3 This invention provides a customer complaint order identification device to meet one of the purposes of this application. It is a functional embodiment of the customer complaint order identification method of this application. On another aspect, this customer complaint order identification device, also to meet one of the purposes of this application, includes an event response module 1100, a risk assessment module 1200, and a risk display module 1300. The event response module 1100 is used to respond to an order payment event and obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features corresponding to the target order in the dimensions of order, buyer, and merchant. The risk assessment module 1200 is used to identify the customer complaint risk probability of the target order based on the target feature set using a customer complaint identification model, and determine whether the customer complaint risk probability exceeds a risk threshold. The risk display module 1300 is used to mark the target order as a high-risk customer complaint order when the customer complaint risk probability exceeds the risk threshold, and to visually display the high-risk customer complaint order to the corresponding buyer and merchant.

[0088] In a further embodiment, before the event response module 1100, the module includes: a feature set construction submodule, used to collect multiple historical customer complaint orders and historical non-customer complaint orders within a preset customer complaint duration, determine multiple candidate features corresponding to each order in the order, buyer, and merchant dimensions, and construct a historical feature set; a sample construction submodule, used to take the historical feature set of each historical customer complaint order as a first sample, and label each first sample with a supervision label to indicate that the corresponding order has initiated a customer complaint, and simultaneously take the historical feature set of each historical non-customer complaint order as a second sample, and label each second sample with a supervision label to indicate that the corresponding order has not initiated a customer complaint; set partitioning The first submodule is used to construct a dataset using all first samples and their supervisory labels, and second samples and their supervisory labels. A portion of the samples in the dataset within a preset recent historical payment period are used as test samples, and the remaining samples are divided into training samples and validation samples according to a preset ratio. Each type of sample is associated with its supervisory label to construct a test set, training set, and validation set. The feature screening submodule is used to determine the candidate distribution difference value between the test set and the training set for each candidate feature. When the distribution difference value reaches the stable distribution threshold, it is determined to be an input feature; otherwise, the candidate feature is removed from the test set, training set, and validation set.

[0089] In a further embodiment, after the feature screening submodule, the system includes: an index determination submodule, used to train the customer complaint identification model to a convergent state using a test set, training set, and validation set, to obtain the importance index of each input feature of the model, and the test performance index of the model; a feature fine screening submodule, used to remove the input feature with the lowest importance index in each sample from the training set, validation set, and test set, retrain the customer complaint identification model to a convergent state based on the removed training set and validation set, and determine the current performance index of the model based on the removed test set; and a model positioning submodule, used to return to the previous step and continue execution when the difference between the test performance index and the current performance index does not exceed the tolerance value; otherwise, to obtain the customer complaint identification model from the previous round and use it to identify the customer complaint risk of the order based on the feature set composed of the remaining input features corresponding to the order.

[0090] In a further embodiment, after the risk display module 1300, the module includes: a data acquisition submodule, used to acquire the customer complaint risk probability and target feature set of each target order identified in two preset monitoring periods; a first monitoring submodule, used to determine the difference value of the input distribution of each input feature in the target feature set between the two monitoring periods based on the target feature set of each target order in the two monitoring periods; a second monitoring submodule, used to determine the difference value of the inference distribution of the recognition probability of the customer complaint identification model between the two monitoring periods based on the customer complaint risk probability of each target order in the two monitoring periods; and a model hot update submodule, used to acquire a customer complaint identification model that has been retrained to a convergent state and perform a hot update on the substandard customer complaint identification model when the difference value of the input distribution or the difference value of the inference distribution indicates that the model performance is substandard.

[0091] In a further embodiment, the feature set construction submodule includes: an order dimension feature submodule, used to obtain the basic attribute features of each order, and determine the statistical features and product association features of the order from historical order data based on at least one preset time window; a buyer dimension feature submodule, used to obtain the basic attribute features and blacklist association features of the buyer corresponding to the order, and determine the buyer's behavioral statistical features from the buyer's historical behavior data based on at least one preset time window; and a seller dimension feature submodule, used to obtain the basic attribute features and blacklist association features of the merchant and its store corresponding to the order, and determine the merchant's operational statistical features from the merchant's historical operating data based on at least one preset time window.

[0092] In a further embodiment, the risk display module 1300 includes: a payment detection submodule, which performs payment risk detection on the high customer complaint risk order to determine the probability of payment risk corresponding to a non-buyer-initiated transaction; a receipt determination submodule, which determines whether the buyer corresponding to the order has received the goods in the order based on the logistics and delivery information of the high customer complaint risk order; a buyer feedback submodule, which pushes the receipt matching degree between the buyer's real-shot image of the goods and the corresponding product image and text information, as well as the receipt integrity degree of the goods in the buyer's real-shot image, to the buyer and merchant corresponding to the high customer complaint risk order when the payment risk probability does not exceed a preset threshold and the buyer has received the goods; and an analysis report push submodule, which constructs a corresponding customer complaint analysis report and pushes it to the buyer and merchant corresponding to the high customer complaint risk order when the payment risk probability exceeds a preset threshold or the buyer has not received the goods.

[0093] In a further embodiment, after the risk display module 1300, the module includes: a first scoring submodule, used to evaluate the buyer's shopping credit score based on the blacklist association characteristics and historical behavior characteristics of the buyer corresponding to the high customer complaint risk order; a second scoring submodule, used to evaluate the seller's business credit score based on the blacklist association characteristics and historical business characteristics of the seller corresponding to the high customer complaint risk order; and a customer complaint initiation submodule, used to initiate a customer complaint initiation process for the buyer corresponding to the high customer complaint risk order when the shopping credit score exceeds a first credibility threshold and the business credit score does not exceed a second credibility threshold.

[0094] To address the aforementioned technical problems, embodiments of this application also provide computer equipment. For example... Figure 4 The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When the computer-readable instructions are executed by the processor, the processor can implement a customer complaint order identification method. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the customer complaint order identification method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0095] In this embodiment, the processor is used to execute... Figure 3 The system contains the specific functions of each module and its sub-modules. The memory stores the program code and various data required to execute these modules or sub-modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules / sub-modules in the customer complaint order identification device of this application. The server can call the server's program code and data to execute the functions of all sub-modules.

[0096] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the customer complaint order identification method of any embodiment of this application.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0098] In summary, this application enables the immediate, efficient, and accurate identification of customer complaint risks after payment, and timely alerts to buyers and merchants when risks are high.

[0099] Those skilled in the art will understand that the steps, measures, and solutions in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted. Furthermore, other steps, measures, and solutions in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. Furthermore, steps, measures, and solutions in the prior art that are similar to those in the open-source operations, methods, and processes of this application can also be alternated, modified, rearranged, decomposed, combined, or deleted.

[0100] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for identifying customer complaint orders, characterized in that, Includes the following steps: In response to an order payment event, obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features of the target order in the dimensions of order, buyer, and merchant. A customer complaint identification model is used to identify the probability of customer complaint risk for the target order based on the target feature set, and to determine whether the probability of customer complaint risk exceeds the risk threshold. When the probability of customer complaints exceeds the risk threshold, the target order is marked as a high-risk order for customer complaints, and the high-risk order for customer complaints is displayed visually to the corresponding buyer and merchant.

2. The customer complaint order identification method according to claim 1, characterized in that, Before responding to an order payment event, the following steps are included: Collect multiple historical customer complaint orders and historical non-customer complaint orders within a preset customer complaint duration, determine multiple candidate features corresponding to each order in the dimensions of order, buyer, and merchant, and construct a historical feature set; The historical feature set of each historical customer complaint order is used as the first sample, and the supervision label of each first sample is marked to indicate that the corresponding order has initiated a customer complaint. At the same time, the historical feature set of each historical non-customer complaint order is used as the second sample, and the supervision label of each second sample is marked to indicate that the corresponding order has not initiated a customer complaint. A dataset is constructed using all the first samples and their supervisory labels and the second samples and their supervisory labels. A portion of the samples in the dataset within the preset recent historical payment period are used as test samples. The remaining samples are divided into training samples and validation samples according to a preset ratio. Each type of sample is associated with its supervisory label, and test set, training set, and validation set are constructed accordingly. For each candidate feature, the candidate distribution difference value between the test set and the training set is determined. When the distribution difference value reaches the stable distribution threshold, it is determined as the feature to be included in the model. Otherwise, remove the candidate feature from the test set, training set, and validation set.

3. The customer complaint order identification method according to claim 2, characterized in that, For each candidate feature, determine the candidate distribution difference value between the test set and the training set. When the distribution difference value reaches the stable distribution threshold, it is determined as a feature to be included in the model; otherwise, after removing the candidate feature from the test set, training set, and validation set, the following steps are included: The customer complaint identification model is trained to a convergent state using a test set, a training set, and a validation set. The importance index of each input feature of the model and the test performance index of the model are obtained. In the training set, validation set, and test set, the input feature with the lowest importance index in each sample is removed. Based on the removed training set and validation set, the customer complaint identification model is retrained to a convergent state, and the performance index of the model in this round is determined based on the removed test set. When the difference between the test performance index and the current round performance index does not exceed the tolerance value, return to the previous step to continue execution; Otherwise, the customer complaint identification model from the previous round is used to identify the customer complaint risk of the order based on the feature set formed by the remaining input features corresponding to the order.

4. The customer complaint order identification method according to claim 1, characterized in that, When the probability of customer complaints exceeds the risk threshold, the target order is marked as a high-risk order for customer complaints. After visually displaying the high-risk order for customer complaints to the corresponding buyer and merchant, the following steps are included: Obtain the customer complaint risk probability and target feature set for each target order identified within two preset monitoring periods; Based on the target feature sets of the target orders in the two monitoring periods, determine the difference in the distribution of each input feature in the target feature set between the two monitoring periods. Based on the customer complaint risk probability of the target orders in each of the two monitoring periods, determine the difference in the inference distribution of the recognition probability of the customer complaint identification model between the two monitoring periods; When the difference value of the input distribution or the difference value of the inference distribution indicates that the model performance is not up to standard, a customer complaint identification model that has been retrained to the convergence state is obtained, and the substandard customer complaint identification model is hot-updated.

5. The customer complaint order identification method according to claim 1, characterized in that, Determining multiple candidate features for each order across the order, buyer, and merchant dimensions includes the following steps: For each order, the basic attribute features of the order are obtained, and the statistical features and product association features of the order are determined from historical order data based on at least one preset time window; Obtain the basic attribute characteristics and blacklist association characteristics of the buyer corresponding to the order, and determine the buyer's behavioral statistical characteristics from the buyer's historical behavior data based on at least one preset time window; The basic attribute features and blacklist association features of the merchant and its store corresponding to the order are obtained, and the merchant's business statistics features are determined from the merchant's historical business data based on at least one preset time window.

6. The customer complaint order identification method according to claim 1, characterized in that, When the probability of customer complaints exceeds the risk threshold, the target order is marked as a high-risk order for customer complaints. After visually displaying the high-risk order for customer complaints to the corresponding buyer and merchant, the following steps are included: For the orders with high customer complaint risk, payment risk detection is performed to determine the corresponding probability of payment risk, which represents a transaction not initiated by the buyer. Based on the logistics and delivery information of the high customer complaint risk orders, determine whether the buyer corresponding to the order has received the goods in the order; When the payment risk probability does not exceed a preset threshold and the buyer receives the goods, the matching degree between the buyer's real-shot image of the goods and the corresponding product image and text information, as well as the condition of the goods in the buyer's real-shot image, are pushed to the buyer and merchant corresponding to the high customer complaint risk order. When the payment risk probability exceeds a preset threshold or the buyer fails to receive the goods, a corresponding customer complaint analysis report is generated and pushed to the buyer and merchant corresponding to the high customer complaint risk order.

7. The customer complaint order identification method according to claim 1, characterized in that, When the probability of customer complaints exceeds the risk threshold, the target order is marked as a high-risk order for customer complaints. After visually displaying the high-risk order for customer complaints to the corresponding buyer and merchant, the following steps are included: The shopping credit score of the buyer is assessed based on the blacklist association characteristics and historical behavior characteristics of the buyer corresponding to the high customer complaint risk order. The seller's business credit score is assessed based on the blacklist association characteristics and historical business characteristics of the sellers corresponding to the high customer complaint risk orders. When the shopping credit score exceeds the first credibility threshold and the business credit score does not exceed the second credibility threshold, the customer complaint initiation process is initiated for the buyer corresponding to the high customer complaint risk order.

8. A customer complaint order identification device, characterized in that, include: The event response module is used to respond to order payment events and obtain the target feature set associated with the target order corresponding to the event. The target feature set includes multiple input features of the target order in the dimensions of order, buyer, and merchant. The risk assessment module is used to identify the probability of customer complaints for the target order based on the target feature set using a customer complaint identification model, and to determine whether the probability of customer complaints exceeds the risk threshold. The risk display module is used to mark the target order as a high customer complaint risk order when the customer complaint risk probability exceeds the risk threshold, and to visually display the high customer complaint risk order to the corresponding buyer and merchant.

9. A computer device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images