Return and refund order information processing, ai decision model training method, and electronic device

Abstract

Embodiments of the present application disclose a return and refund order information processing method, an AI decision model training method and an electronic device, including: after receiving current return and refund order information submitted by a user, determining basic feature information of the current return and refund order; acquiring extended feature information of the current return and refund order according to the basic feature information; inputting the basic feature information and the extended feature information into an AI decision model to determine a plurality of candidate historical return and refund orders with a basic feature similarity to the current return and refund order meeting a condition from a historical return and refund order set, and determine a target historical return and refund order with an extended feature similarity to the current return and refund order meeting a condition, and determine a preferred retention scheme bound to the target historical return and refund order as a recommended retention scheme corresponding to the current return and refund order. Through the embodiments of the present application, the recommended retention scheme can be attractive to the buyer user, so as to improve the adoption rate.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to information processing of return and refund orders, AI decision model training methods, and electronic devices. Background Technology

[0002] In cross-border commodity information service systems, the return and refund process for buyers is typically very lengthy due to the cross-border transportation of goods. For example, if an overseas buyer is dissatisfied with the goods received, they can select "Return and Refund" on the client page, choose the reason for the return and refund, and submit relevant evidence (e.g., photos and text descriptions of the goods). After submitting the "Return and Refund" request, the buyer needs to select a logistics provider on another page and click submit. The system will then generate a refund order. If the merchant has no objection to this refund order, the formal return process can begin. In this process, the buyer needs to contact the logistics provider for pickup and then obtain a logistics receipt from the logistics station. They must then photograph or scan the logistics receipt and upload it to the system interface. The platform receives the logistics receipt and forwards it to its internal logistics department for verification. The goods are typically returned to the platform's warehouse first (in some overseas countries, dedicated city-level warehouses are deployed in major cities; alternatively, warehouses shared by multiple countries / regions can be deployed in a specific area. If a buyer's country / city does not have a dedicated warehouse, the goods can be returned to that shared warehouse). There, quality inspection is conducted, including checking for damage during the return process and the extent of that damage. An inspection report is provided to the merchant. If the merchant needs the goods, they can choose to pick them up from the platform warehouse or have them shipped from the platform warehouse to the merchant's warehouse. Alternatively, if the merchant has an overseas warehouse, the buyer may return the goods directly to that warehouse. The merchant will only issue a refund to the buyer after receiving the returned goods, thus ending the process. If the merchant objects to the refund, the process will first involve negotiation. If negotiation fails, it will proceed to arbitration, where the platform will make a ruling. In most cases, the platform will find the merchant unreasonable, and the return process will continue. Of course, merchants can also appeal the arbitration result. If the arbitration result is indeed problematic, the platform will need to provide compensation to the merchant.

[0003] It is evident that the return and refund process in cross-border scenarios is quite complex, involving customs clearance and other procedures during return shipping. Therefore, buyers may experience a long waiting time from initiating a return and refund request to receiving the refund, resulting in a poor user experience. Furthermore, once the return and refund process begins, regardless of whether the two parties reach an agreement, it may cause losses to the merchant and even the platform. Summary of the Invention

[0004] This application provides information processing for return and refund orders, AI decision model training methods, and electronic devices that can make the recommended retention strategies attractive to buyer users, thereby increasing the adoption rate.

[0005] This application provides the following solution:

[0006] A method for processing return and refund order information includes: After receiving the current return and refund order information submitted by the user, determine the basic feature information of the current return and refund order; Based on the basic feature information, the extended feature information of the current return and refund order is obtained. The extended feature information includes: influencing factors that affect the selection result of retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information of the current return and refund order are input into the AI ​​decision model. The AI ​​decision model is used to determine multiple candidate historical return and refund orders from the set of historical return and refund orders that meet the similarity conditions of the basic features of the current return and refund order, and based on the extended feature information, determine the target historical return and refund order from the multiple candidate historical return and refund orders that meets the similarity conditions of the extended features of the current return and refund order, and determine the preferred retention plan bound to the target historical return and refund order as the recommended retention plan corresponding to the current return and refund order. Provide users with retention information based on the recommended retention strategy.

[0007] The step of obtaining the extended feature information of the current return and refund order based on the basic feature information includes: Based on the product category information associated with the current return and refund order, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the current return and refund order is determined based on the basic feature information of the current return and refund order, and the score information is determined as the extended feature information; the score information includes: the score of the current return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

[0008] The AI ​​decision-making model is trained in the following way: The basic feature information of multiple historical return and refund orders is obtained and input into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders. Based on the adoption rate and the basic feature information of the historical return and refund orders, the after-sales cost rate of multiple retention strategies is used as a condition to determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, the historical return and refund orders are bound with the preferred retention strategy and the extended feature information and then saved. The extended feature information is obtained based on the basic feature information and includes the influencing factors that affect the selection result of the retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information are input as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate recommended retention solutions, and the AI ​​decision-making model is supervised training with the preferred retention solution as the optimization target.

[0009] This also includes: After providing retention information to users based on the recommended retention plan, data tracking is performed, including the actual adoption rate of the retention plan, in order to optimize the AI ​​decision-making model.

[0010] The optimization of the AI ​​decision-making model includes: The prediction model's prediction of user adoption rate is adjusted based on the actual adoption rate, so that the AI ​​decision model can be optimized and trained based on the adjusted prediction results.

[0011] An AI decision-making model training method includes: The basic feature information of multiple historical return and refund orders is obtained and input into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders. Based on the adoption rate and the basic feature information of the historical return and refund orders, the after-sales cost rate of multiple retention strategies is used as a condition to determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, the historical return and refund orders are bound with the preferred retention strategy and the extended feature information and then saved. The extended feature information is obtained based on the basic feature information and includes the influencing factors that affect the selection result of the retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information are input as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate recommended retention solutions, and the AI ​​decision-making model is supervised training with the preferred retention solution as the optimization target.

[0012] The process of determining the optimal retention strategy corresponding to historical return and refund orders includes: Based on the price attribute-related information in the basic feature information of the historical return and refund orders, and the compensation method information corresponding to the multiple retention plans, the after-sales cost rate of the multiple retention plans is calculated. Based on the adoption rate and after-sales cost rate, retention plans with after-sales cost rate within the target range and adoption rate meeting preset conditions are determined as the preferred retention plans for historical return and refund orders.

[0013] The step of obtaining the extended feature information of the historical return and refund order includes: Based on the product category information associated with the historical return and refund orders, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the historical return and refund order is determined based on the basic feature information, and the score information is determined as the extended feature information; the score information includes: the score of the historical return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

[0014] This also includes: Based on the adoption rate and after-sales cost rate of each retention strategy, the optimality of the multiple retention strategies is ranked, and the ranking results are also input into the AI ​​decision-making model for decision-making.

[0015] The basic feature information of the historical return and refund order includes order information, refund data, information of the associated logistics order, information of the first user, and information of the second user, wherein the first user and the second user correspond to the users of the two parties in the transaction.

[0016] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the preceding methods.

[0017] An electronic device, comprising: One or more processors; and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of any of the preceding methods.

[0018] A computer program product includes a computer program / computer executable instructions that, when executed by a processor in an electronic device, implement the steps of any of the preceding methods.

[0019] According to the specific embodiments provided in this application, the following technical effects are disclosed: According to the embodiments of this application, after receiving the current return and refund order information submitted by the user, the basic feature information of the current return and refund order can be determined first. Additionally, extended feature information of the current return and refund order can be obtained based on the basic feature information. Then, the basic and extended feature information can be input into an AI decision model to identify multiple candidate historical return and refund orders from the historical return and refund order set that meet the basic feature similarity criteria of the current return and refund order, and to identify the target historical return and refund order that meets the extended feature similarity criteria of the current return and refund order. The preferred retention plan bound to the target historical return and refund order is then determined as the recommended retention plan corresponding to the current return and refund order. Through the above method, the user intent and other information expressed behind the data can be mined, thereby making the retention plan recommended by the model attractive to buyers, improving the adoption rate, and achieving multiple optimizations in user satisfaction, merchant costs, and platform benefits.

[0020] In training the AI ​​decision-making model, more accurate optimal retention strategies can be labeled for historical return and refund orders. Furthermore, by analyzing the influencing factors affecting the selection of retention strategies and the importance of these factors, extended features can be generated for historical return and refund orders. These features can more effectively reveal the specific reasons for selecting the corresponding optimal retention strategy for historical return and refund orders. During model training, the above information can be fully utilized to mine information such as user intent expressed behind the data, thereby making the retention strategies recommended by the model attractive to buyers, improving the adoption rate, and achieving multiple optimizations in user satisfaction, merchant costs, and platform benefits.

[0021] Of course, any product implementing this application does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the system architecture provided in the embodiments of this application; Figure 2 This is a flowchart of the method for processing return and refund order information provided in the embodiments of this application; Figure 3This is a flowchart of the model training method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the interface provided in an embodiment of this application; Figure 5 This is a schematic diagram of the interaction process provided in the embodiments of this application; Figure 6 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0025] In this embodiment of the application, in order to benefit buyers, merchants, and the platform, intervention can be made when a buyer initiates a return and refund request. This can be done by offering compensation to the user to retain them. If the user accepts the retention plan, they may cancel the return and refund, thereby avoiding the subsequent complex and lengthy cross-border return and refund process, and thus reducing the losses of buyers, merchants, and the platform.

[0026] It's worth noting that some existing non-cross-border product information service systems also handle similar user retention strategies. However, the methods used in non-cross-border scenarios are usually simpler, primarily implemented through simple rule configurations. For example, merchants can choose to offer a "partial refund" retention plan for products within a specific price range, setting the refund percentage (e.g., 10% refund for products in the 200-1000 range, etc.). When a buyer initiates a return and refund request for a specific order, the system can display the merchant's configured retention plan to the user. For instance, assuming an order amount of 500 yuan falls within the aforementioned price range, the corresponding retention plan could be a "50 yuan refund," asking the buyer if they accept. If the buyer accepts, the return and refund can be automatically cancelled.

[0027] The above solutions can reduce the rate of returns and refunds to some extent. However, they may not be suitable for cross-border scenarios if applied directly. This is because the solutions implemented through simple rule configuration lack in-depth analysis of the buyer's true intentions and behavioral data. As a result, the retention strategies provided may be inaccurate, leading to unsuccessful retention efforts. This in turn results in unnecessary returns, incurring costs for logistics, quality inspection, and inventory losses, and wasting resources.

[0028] In response to the above situations, this embodiment of the application can provide a more suitable retention plan after identifying the buyer's true intentions, thereby increasing the probability that the user will accept the specific retention plan. To achieve the above objective, the capabilities of AI (Artificial Intelligence) can be utilized. After the buyer submits a return and refund request, real-time AI analysis can be performed to analyze the buyer's true reasons for the refund or their willingness to retain the customer. Then, based on the analysis results, a matching retention plan can be provided to increase the probability that the user will accept the retention plan. Additionally, in an optional manner, the retention strategy provided by AI analysis in this application embodiment can be offered only once. This means there won't be a situation where a user rejects the current retention strategy and another strategy is offered. If the user does not accept the recommended retention strategy, this application embodiment can also provide a "one-stop managed" return and refund process. In other words, if the user does not accept the retention strategy, the return and refund process can be automatically initiated. However, the buyer no longer needs to select a logistics provider; instead, the platform can select a logistics provider. The platform has pre-stored the logistics provider's contact information and provides the logistics provider with the specific buyer's address, platform warehouse or merchant warehouse address, weight, and product title. Then, the logistics provider picks up the goods, the buyer provides the goods to the logistics provider, and the platform's logistics department receives the logistics receipt, logistics route, and other information from the logistics provider. After receiving the returned goods, they can take photos, conduct quality inspections, etc., and then process the refund. This method reduces the operational costs for buyers and also reduces the workload of merchant-side customer service personnel (communication with buyers during the return and refund process).

[0029] In the above approach, since only one retention plan is provided to the buyer, if the buyer does not accept it, the return and refund process can proceed through the "one-stop escrow" system. Therefore, the accuracy of the retention plan is required to be higher, which necessitates the use of AI capabilities.

[0030] Specifically, when using AI to analyze the true reasons for refunds or the willingness to retain buyers, this can be achieved through an AI model. That is, by inputting order information, logistics information, buyer information, and merchant information corresponding to the specific return and refund request into the AI ​​model, the model determines the scores for various retention strategies and recommends the highest-scoring strategy to the buyer. Based on the aforementioned role of the AI ​​model, the specific AI model used in this embodiment can be referred to as an AI decision model.

[0031] The specific AI decision-making model can be selected from existing open-source AI models. However, due to the characteristics of AI models, which perform well in general domains but face severe challenges in adapting to specific professional domains (such as predicting refund reasons or retention intentions in buyer user refund and return scenarios in this application embodiment), it is usually necessary to fine-tune or strengthen the model using relevant training sample data in specific professional domains before it can adapt to the generation requirements of specific professional domains. Otherwise, the model may produce "illusions" or logical inconsistencies during the generation process. In the scenario of this application embodiment, this is specifically manifested as: the identified user intent may not be accurate enough, and consequently, the recommended retention plan is also not accurate enough.

[0032] To train an AI decision-making model, relevant training sample data is needed. For example, historical return and refund records from the system can be obtained, including specific order information, logistics information, buyer information, and merchant information. Furthermore, as training samples, this data usually needs to be labeled, for example, with retention strategies corresponding to specific orders. This labeled information will be used to supervise the training of the AI ​​decision-making model. However, in practical applications, only a small portion of historical return and refund orders may have resulted in customer retention efforts, meaning that most historical return and refund orders do not directly correspond to retention strategy information. Even if retention strategy information is associated, current technologies typically determine this through simple rule configuration. Therefore, regardless of whether the user ultimately accepts the retention strategy, it is unknown whether the specific strategy is the preferred option relative to the user's current reason for the refund or their desire to retain them. In many cases, the answer is no. This makes it difficult to directly obtain high-quality labeled information from historical data. While high-quality labeled data can be obtained through manual annotation, it is costly. Furthermore, for the scenario described in this application, the amount of data to be analyzed is very large (for example, to obtain a buyer's user profile, it is necessary to analyze the user's transaction records, order volume, and refund volume over the past year or even longer, as well as the merchant's user profile, order data, refund data, etc.) before it is possible to analyze the specific order of a specific user and determine the most appropriate retention strategy. Therefore, manual annotation is very difficult and may even be an impossible task.

[0033] Based on the above, in the embodiments of this application, such as Figure 1 As shown, before specifically training the AI ​​decision-making model, a prediction model and an integer programming optimization algorithm can be provided first to construct the training data for the AI ​​decision-making model. Specifically, the input data for the prediction model can be the basic feature information corresponding to the aforementioned historical return and refund orders (including order information, logistics information, buyer user information, merchant user information; if the historical return and refund orders involve cross-border transactions, it can also include cross-border logistics information, customs clearance status information, overseas warehouse / platform warehouse address information, etc.). Correspondingly, the prediction model can predict the adoption rate (i.e., the probability of being adopted by the corresponding buyer user) of various retention strategies under the above input data conditions, and can also generate an adoption rate-retention strategy ratio curve for each historical return and refund order. Among them, in cross-border scenarios, various retention strategies can include: return to platform warehouse + free door-to-door pickup + red envelope reward, partial return + platform coupon compensation, small amount refund, return to merchant warehouse + merchant coupon, etc.

[0034] After predicting the adoption rates of various retention strategies using a predictive model, a series of processing steps can be performed using an integer programming optimization algorithm. Specifically, the after-sales cost rate corresponding to each of the various retention strategies can be calculated under the same historical return and refund order conditions. The after-sales cost rate can be calculated using a preset formula, for example: After-sales cost rate = (refund amount + quality inspection fee + logistics fee + handling fee + merchant self-pickup fee or shipping fee from platform warehouse to merchant warehouse) / order amount The refund amount can be determined based on the compensation method corresponding to the specific retention plan. For example, for a retention plan involving certain images, the compensation method for users is to refund a certain amount according to a preset percentage, and so on. For each specific retention plan, given the basic characteristics of the same historical return and refund order, the refund amount, quality inspection fee, logistics fee, handling fee, merchant self-pickup fee, or shipping cost from the platform warehouse to the merchant warehouse in the above formula are all fixed. Therefore, they can be directly substituted into the above formula to calculate the after-sales cost rate for each retention plan.

[0035] Given the same historical return and refund order, after obtaining the adoption rate and after-sales cost rate of various retention strategies, the retention strategy with the highest or relatively high adoption rate and an acceptable after-sales cost rate can be selected as the preferred retention strategy under the condition of the input data. This retention strategy can also be used as the labeled data of the above input data for supervised training of the subsequent AI decision model.

[0036] The above describes the annotation of optimal retention strategies corresponding to historical return and refund orders. That is, for a given historical return and refund order, based on specific order information, logistics information, buyer information, and merchant information, an attractive retention strategy for the buyer is determined. However, to enable the AI ​​decision-making model to better perceive the causal relationship between specific historical return and refund orders and optimal retention strategies, and to better reveal the true user intent expressed behind the specific input data, it is necessary not only to know which optimal retention strategy corresponds to a specific historical return and refund order, but also to analyze which factors contributed to the success of that retention strategy and to quantify these influencing factors. Specifically, this application embodiment can pre-store the correspondence between specific product categories, sets of influencing factors, and the weights of each influencing factor. This correspondence can be pre-configured. Different product categories may correspond to different influencing factors, and the weights of each influencing factor may be different. The weight represents the importance of a specific influencing factor in the selection of a specific retention strategy. In this way, for a specific historical return and refund order, the corresponding set of influencing factors and the weight of each factor can be determined first based on its associated product category information. Furthermore, an integer programming algorithm can be used to analyze the score of the historical return and refund order on each influencing factor. Then, the scores of the same historical return and refund order on multiple different influencing factors, along with the weights of each factor, can be used to calculate the total score of the historical return and refund order. Moreover, the specific historical return and refund order can be linked to the previously determined optimal retention strategy and the aforementioned score information (including the total score, scores on multiple different influencing factors, and the weights of each influencing factor).

[0037] Furthermore, the aforementioned integer programming algorithm can also rank different retention strategies for the same historical return and refund order (e.g., based on acceptance rate, after-sales cost rate, etc.). Thus, a specific historical return and refund order, in addition to basic features such as order information, logistics information, buyer user information, and merchant user information, can also possess richer feature information, including scores on multiple influencing factors, the weights of each influencing factor, the total score, the ranking information among different retention strategies, and the optimal retention strategy. This information can serve as an extension or supplement to the basic feature information, and together with the basic feature information, it can be used as input data for training the AI ​​decision-making model. Specifically, during training, the AI ​​decision-making model can determine the preferred retention strategy for a specific historical return and refund order from multiple different options based on the input data, and use the optimal retention strategy bound to the training data for supervision, enabling the AI ​​decision-making model to acquire the ability to determine the optimal retention strategy based on the aforementioned input data.

[0038] After completing the above training, the AI ​​decision-making model can be used for online decision-making. Additionally, during the training process, the following information can be saved: the binding relationships between multiple historical return and refund orders and their respective scores, weights, total scores, and optimal retention strategies on multiple influencing factors. Thus, upon receiving a new return and refund request, the model can first determine the influencing factors and their weights based on the product category information associated with the new return and refund order. Furthermore, it can determine the scores and total score of the new return and refund order on each influencing factor based on order information, logistics information, buyer information, and merchant information. Then, the basic feature information corresponding to the new return and refund order (order information, logistics information, buyer information, merchant information, etc.), as well as the extended feature information related to the aforementioned influencing factor scores, weights, and total scores, can be input into the AI ​​decision-making model. The AI ​​decision-making model can first select multiple historical return and refund orders that are similar to the current new return and refund order based on the aforementioned basic feature information. Then, based on the extended feature information related to the specific score, it selects the historical return and refund order that meets the similarity criteria in terms of extended features (e.g., the highest similarity) from these multiple historical return and refund orders that are similar in basic features. Since the historical return and refund order is bound to a preferred retention plan, the preferred retention plan corresponding to the finally selected historical return and refund order can be determined as the preferred retention plan for the currently received new return and refund order.

[0039] Therefore, since the AI ​​decision-making model can determine the optimal retention plan for return and refund orders, it is possible to retain users based on this optimal plan. Specifically, the optimal retention plan can be provided to the buyer, and the user can be asked whether they accept it. In this embodiment, the specific retention plan is selected by the AI ​​decision-making model from multiple options, and the decision-making process fully utilizes information on factors influencing the selection of the retention plan. Therefore, the probability that the retention plan aligns with the user's actual intentions is increased, leading to a higher acceptance rate of the optimal retention plan and reducing the number of orders actually entering the return and refund process. Of course, if the user does not accept the aforementioned optimal retention plan, they can proceed to the "one-stop managed" return and refund process described earlier.

[0040] The specific implementation schemes provided in the embodiments of this application will be described in detail below.

[0041] Example 1 First, this first embodiment provides an AI decision-making model training method for a specific model training process, see [link to example]. Figure 2 The method may include: S201: Obtain basic feature information of multiple historical return and refund orders, and input the basic feature information into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders.

[0042] Specifically, historical return and refund orders can be obtained from relevant databases within the product information service system. Alternatively, basic characteristic information about historical return and refund orders can be obtained through data retrieval from other relevant databases. This basic characteristic information can include order information, refund data, information on associated logistics orders, information on the first user, and information on the second user. The first and second users can correspond to the buyer and seller in the transaction, respectively; for example, the first user could be the buyer, and the second user could be the seller. Specifically, the first user information can include consumer profiles: consumer level, order volume in the past year, refund volume, etc.; and merchant profiles: merchant level, refund volume in the past year, order volume, and the percentage of refunds under different schemes, etc. Order information can include order amount, quantity of goods, product image, order time, and order end time, etc.; refund data can include refund application amount, refund timeframe, refund scheme, logistics fees, quality inspection fees, handling fees, shipping fees, etc.

[0043] The aforementioned basic feature data can be input into the prediction model, which can then predict the user's adoption rate of multiple retention strategies, specifically based on the basic feature information of the historical return and refund orders. The prediction model can be a computationally efficient model with fewer parameters trained using knowledge distillation techniques. That is, through knowledge distillation, the knowledge of a large and complex model (which can be called the teacher model) is transferred to a lightweight model (which can be called the student model). The teacher model outputs a probability distribution on the training data, while the student model not only learns the real labels but also learns to imitate the output of the teacher model, thereby acquiring knowledge with stronger generalization ability, and so on.

[0044] Specifically, the aforementioned predictive model can estimate the adoption rate based on the year-on-year order volume of a specific retention strategy divided by the order volume of the same strategy used in the same industry. For example, for a certain type of return and refund order, the number of return and refund orders using a certain retention strategy, divided by the total number of return and refund orders of that type, yields the adoption rate of that retention strategy for that type of return and refund order. The specific type of return and refund order can also be determined based on basic characteristic information. For example, the same product category and the same degree of damage (such as damage caused during transportation) can correspond to one type of return and refund order, and so on. Regarding the "number of return and refund orders using a certain retention strategy," this can include the percentage of historical return and refund orders that have used that strategy, and it can also estimate the percentage of future return and refund orders of the same type that might use that strategy. Combining these two factors yields the adoption rate of the specific retention strategy under the condition of that type of return and refund order.

[0045] S202: Based on the adoption rate and the basic feature information of the historical return and refund orders, and considering the after-sales cost rate of multiple retention strategies, determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, bind the historical return and refund orders with the preferred retention strategy and the extended feature information, and save them. The extended feature information is obtained based on the basic feature information and includes the influencing factors affecting the retention strategy selection result and the importance information of the influencing factors.

[0046] After predicting the adoption rates of multiple retention strategies using a predictive model, a series of processes can be performed, including calculating the after-sales cost rate of multiple retention strategies, determining the preferred retention strategy for historical return and refund orders, generating extended feature information of historical return and refund orders, and binding historical return and refund orders with the preferred retention strategy and extended feature information, etc.

[0047] In one specific implementation, the above series of processes can be accomplished using integer programming algorithms. Integer programming is a branch of data programming that requires all or some decision variables to take integer values. Essentially, it seeks one (or more) integer solutions under a series of linear or nonlinear constraints, such that the objective function achieves its optimization goal (e.g., maximizing profit or minimizing cost). Unlike linear programming, which allows variables to take any real number, integer programming has a discrete set of points as its solution space. This type of integer programming algorithm is generally used to solve complex optimization problems that require making an "integer number" of decisions or "either / or" choices.

[0048] In this embodiment, since the ultimate goal is to find the optimal retention plan for a given historical return and refund order from among multiple retention plans, it falls under the scenario applicable to the aforementioned integer programming algorithm.

[0049] Specifically, the after-sales cost rate of the retention strategies can be calculated based on the price attribute-related information in the basic feature information of historical return and refund orders, as well as the compensation method information corresponding to each of the multiple retention strategies. For example, as mentioned above: After-sales cost rate = (refund amount + quality inspection fee + logistics fee + handling fee + merchant self-pickup fee or shipping fee from platform warehouse to merchant warehouse) / order amount After calculating the after-sales cost ratio of multiple retention strategies, the retention strategies with after-sales cost ratios within the target range and acceptance rates meeting preset conditions can be determined as the preferred retention strategies for historical return and refund orders, based on the adoption rate and the after-sales cost ratio. For example, the retention strategy with the highest adoption rate under the condition that the after-sales cost ratio is within a certain range (e.g., below a certain threshold) can be determined as the preferred retention strategy for historical return and refund orders.

[0050] Specifically, there are multiple ways to determine the extended feature information. For example, in one approach, multiple influencing factors related to the retention strategy selection and their corresponding weights can be determined based on the product category information associated with historical return and refund orders, as well as the pre-defined mapping relationship between product categories and influencing factor information. Then, the score information of historical return and refund orders is determined based on the basic feature information of these orders, and this score information is defined as the extended feature information of the historical return and refund orders. Specifically, the score information may include: the scores of the historical return and refund orders on the multiple influencing factors, and the total score determined based on the scores and weights of the multiple influencing factors.

[0051] In other words, to better reveal the impact of different user intentions on the selection of retention strategies, some influencing factors can be predefined. Furthermore, the types of influencing factors and their weights (representing their influence or importance on the retention strategy selection results) can vary for different product categories. Specifically, the mapping relationship between the aforementioned product categories and influencing factor information (including the set of influencing factors and the weights corresponding to different influencing factors) can be pre-configured. For a specific historical return and refund order, based on its associated product category information, the mapping relationship information can be used to query which influencing factors correspond to that category and the weight of each factor.

[0052] There can be multiple specific influencing factors, which can be divided into three main categories: objective, subjective, and a combination of both. Specifically: Objective factors can also be divided into order-related and product-related influencing factors, and dispute-related influencing factors: Among the factors influencing orders and products, those related to orders can include: product value (which can be binned based on the percentile of the product price distribution and converted into category characteristics, such as low value, medium value, high value, etc.), product category (which can be directly expressed using category IDs, etc.), product usage time (which can be converted into categories representing the degree of use, such as new, short-term use, long-term use, beyond warranty, etc.), and the matching degree between the return date and the dispute protection period (which can be converted into specific status categories, such as within the protection period, nearing expiration, expired, etc.). Factors related to orders can also include: shipping / taxes (shipping cost / product value, which can calculate relative costs and eliminate the influence of absolute values), and logistics status (which can be expressed using status IDs, such as in transit, signed for, abnormal).

[0053] Dispute-related influencing factors can be further divided into factors related to the dispute situation and the objective facts of the complaint. Factors related to the dispute situation can include the dispute reason code (which can be directly used as the dispute category ID), the requested refund amount (request amount / goods value, a core feature representing the user's expected proportion), and logistics transportation time (which can be compared with the average transportation time of similar products for classification, e.g., fast, normal, slow, severely delayed). Factors related to the objective facts of the complaint can include the document's damage level (using a pre-trained image classification model to judge the damage level of document images, e.g., no damage, minor, moderate, severe) and the document's workmanship / material (using image recognition and text keyword matching to extract specific quality problem tags, e.g., for clothing, categorized as loose threads, stains, incorrect material, incorrect size).

[0054] Objective factors can be further divided into factors related to user expression and factors specific to the context.

[0055] Factors influencing user expression can include: the description of the dispute text (structured features can be extracted first, including emotional words, appeal words, etc., such as text length, whether it contains threatening words, and problem keywords), buyer sentiment, appeals, and risks. Buyer sentiment can be further subdivided into emotional intensity, communication attitude (sentiment analysis models can be used to score or classify user text, such as strongly negative, negative, neutral, positive, or [-1, 1] scores), and current emotional state (fine-grained emotion recognition can be performed on the latest conversation, such as anger, disappointment, calm, and urgency). Regarding appeals and risks, risk levels can be determined using keywords (such as "complaint," "exposure," "law," etc.) and sentence structure, and risk levels can include high risk, medium risk, and low risk.

[0056] The influencing factors of situational specificity can be mainly divided into those related to the product's purpose (e.g., holiday gifts) and those related to the dispute scenario (e.g., late delivery, damage). Factors related to the product's purpose can be determined using keywords (e.g., "gift," "birthday"). Disputes with high time sensitivity usually require faster and higher priority processing; this can be specifically expressed as whether the product is time-sensitive. Factors related to the dispute scenario can be categorized into top-level scenarios by combining the cause of the dispute and user descriptions; this can be expressed using scenario category IDs, etc.

[0057] The mixed subjective and objective influencing factors can include influencing factors related to buyer profiles and influencing factors related to merchant profiles.

[0058] Influencing factors related to buyer profiling can include consumer level, historical purchase frequency / amount (which can be based on user segmentation logic, converting continuous values ​​into user tags, such as low frequency / low value, medium frequency / medium value, high frequency / high value), historical dispute count (which can be used to identify whether a user is a "high-risk" or "sensitive" user, specifically including first dispute, occasional dispute, and high frequency dispute), historical partial withdrawal rate (representing the user's behavioral preferences in historical negotiations, which is important for the strategy selection of AI decision-making models, for example, it can be divided into high acceptance, medium acceptance, low acceptance, and never accepting), and consumer growth potential assessment (which can stratify the predicted growth potential score for consideration of the user's long-term value when making decisions, such as including low potential, medium potential, and high potential).

[0059] Influencing factors related to merchant profiles may include merchant rating / service score (which reflects the merchant's service level and historical performance, such as excellent, good, average, and needs improvement) and historical dispute rate (which reflects the quality stability of the merchant's goods or services, such as low, medium, and high).

[0060] For different product categories, the importance of the various influencing factors mentioned above in the selection of retention strategies varies, and therefore, they can be assigned different weights. For example, for clothing, factors related to product value may have a higher weight, while for electronic devices, factors related to the degree of damage may have a higher weight, and so on.

[0061] After retrieving the specific influencing factors and their weights by analyzing the product category information corresponding to a particular historical return and refund order, the score for each influencing factor can be determined based on the order's performance on those factors. For example, as mentioned earlier, the various values ​​corresponding to specific influencing factors can be quantified. Assuming that influencing factors related to product value correspond to low, medium, and high values, after determining which value the order belongs to based on its associated product value, it can be further converted into a score. For instance, low value could be 1 point, medium value 2 points, high value 3 points, and so on. This method allows obtaining the scores for the same historical return and refund order on multiple different influencing factors. Furthermore, the scores for these multiple influencing factors can be weighted and summed to obtain a total score. For example, suppose there are three influencing factors. A historical return and refund order scores 2 points in influencing factor A with a weight of 20%, scores 3 points in influencing factor B with a weight of 40%, and scores 1 point in influencing factor C with a weight of 20%. Then the total score of the historical return and refund order could be 2×20%+3×40%+1×20%=1.8.

[0062] The above describes the scores of various influencing factors and the total score. This information can be used as extended feature information of historical return and refund orders, thereby expanding the basic feature information to include more semantics related to user intent analysis, which can provide richer information for subsequent AI decision-making models.

[0063] In practice, specific historical return and refund orders can be bound to the corresponding determined preferred retention plan and extended feature information (including scores, weights, and total scores under various influencing factors) and saved. This information will be used at least in the subsequent online decision-making process of the AI ​​decision-making model, which will be described in detail later.

[0064] In addition to obtaining the optimal retention strategy for specific historical return and refund orders, the algorithm can also rank the optimality of multiple retention strategies based on their adoption rates and after-sales cost rates. This ranking information can also be provided to the AI ​​decision-making model during the training process to help the AI ​​decision-making model make decisions.

[0065] S203: Input the basic feature information and the extended feature information as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate a recommended retention plan, and use the preferred retention plan as the optimization target to supervise the training of the AI ​​decision-making model.

[0066] After identifying the extended feature information, the basic feature information corresponding to historical return and refund orders, along with the extended feature information, can be input into the AI ​​decision model as training data. The AI ​​decision model is then guided to generate recommended retention strategies through prompts and other methods. Furthermore, the optimal retention strategies bound to specific historical return and refund orders can be used as optimization targets for supervised training of the AI ​​decision model. In other words, by learning from the optimal retention strategies generated by integer programming algorithms in historical scenarios, a lightweight regression model is trained. This model can predict near-optimal retention strategies for individual return and refund orders in real time, achieving rapid response and supporting high-concurrency online business needs while meeting constraints such as specified adoption rates and after-sales cost rates.

[0067] Specifically, the input features of the AI ​​decision-making model can include: structured basic features of each historical return and refund document (such as user profile, product attributes, order amount, refund reason, historical behavior, etc.), and extended features (including scores, weights, total scores, etc. under multiple different influencing factors). Constraints can be: under the premise that the after-sales cost rate is within a certain range, the adoption rate meets the conditions (e.g., the highest). Supervision signals can be: under the aforementioned constraints, the integer programming algorithm outputs the optimal retention plan for each historical return and refund document (i.e., the label marked on the specific historical return and refund document). Based on the above information, the output information of the AI ​​decision-making model is the recommended retention plan for the specific historical return and refund document.

[0068] Specifically, in the process of training the AI ​​decision-making model, weighted cross-entropy loss can be used. Since the number of acceptance and rejection samples may be unbalanced, a higher weight can be assigned to the class with fewer samples to avoid the model being biased towards predicting the majority class.

[0069] In addition, regarding the learning rate, a linear decay mechanism with warm-up can be used. That is, a small learning rate can be used in the early stage of training (warm-up phase), then linearly increased to a set value, and then linearly decayed to 0 with the number of training steps. This helps the model to be more stable in the early stage of training.

[0070] After completing the above training, the specific AI decision-making model can be deployed in online scenarios. When a user submits a new return and refund order, the AI ​​decision-making model can be used to determine the recommended retention plan and prompt the user. If the user chooses to accept the retention plan, the subsequent formal return and refund process can be avoided, thus preventing losses to the buyer, the merchant, and the platform.

[0071] The training method of the AI ​​decision-making model provided in this application embodiment has been described above. This method first predicts the user's adoption rate of multiple retention strategies based on the basic feature information of the historical return and refund order. Then, an integer programming algorithm is used to perform a series of processes, including determining the preferred retention strategy corresponding to the historical return and refund order based on the adoption rate and the after-sales cost rate of the multiple retention strategies based on the basic feature information of the historical return and refund order, obtaining the extended feature information of the historical return and refund order, and binding and saving the historical return and refund order with the preferred retention strategy and the extended feature information. The extended feature information is obtained based on the basic feature information and includes influencing factors affecting the retention strategy selection result and the importance information of the influencing factors. Then, the basic feature information and the extended feature information can be used as training data input into the AI ​​decision-making model to guide the AI ​​decision-making model to generate recommended retention strategies, and the AI ​​decision-making model is trained under supervision using the preferred retention strategy as the optimization target. The above methods can be used to more accurately identify optimal retention strategies for historical return and refund orders. Furthermore, by analyzing the influencing factors and their importance in determining the retention strategy selection results, extended features can be generated for historical return and refund orders. These features can more effectively reveal the specific reasons for selecting the corresponding optimal retention strategy for each historical return and refund order. During model training, this information can be fully utilized to mine the user intent and other information expressed behind the data, thereby making the retention strategies recommended by the model most attractive to buyers, improving the adoption rate, and achieving multiple optimizations in user satisfaction, merchant costs, and platform benefits.

[0072] Example 2 The first embodiment described a specific training method for the AI ​​decision-making model. This second embodiment primarily describes the online decision-making process of the AI ​​decision-making model. Preferably, the AI ​​decision-making model can be trained using the scheme described in the first embodiment. Specifically, this second embodiment provides a method for processing return and refund order information, see [link to relevant documentation]. Figure 3 The method may specifically include: S301: After receiving the current return and refund order information submitted by the user, determine the basic feature information of the current return and refund order.

[0073] In practical implementation, a user-facing front-end interface can provide a simple workflow through an entry point for the "return and refund" function, allowing users to initiate return and refund requests. Furthermore, when a user initiates a return and refund request, they can be dynamically guided to upload supporting documentation (such as product images / videos) and fill in remarks. This information, along with the corresponding order information, can be submitted to the server as the information for the current return and refund order. Additionally, after receiving the information for the current return and refund order, the server can also obtain its basic characteristic information, which may include the order information, refund data, associated logistics order information, information of the first user, and information of the second user, as mentioned above.

[0074] S302: Obtain extended feature information of the current return and refund order based on the basic feature information. The extended feature information includes: influencing factors affecting the retention strategy selection result and the importance information of the influencing factors.

[0075] Based on the acquired basic feature information, extended feature information of the current return and refund order can also be obtained. Similar to the training process, the specific extended feature information includes: influencing factors affecting the retention strategy selection result and the importance information of these influencing factors. The importance of the influencing factors can be determined by the score of the specific return and refund order on the corresponding influencing factor and the weight of the influencing factor. For example, in one specific implementation, multiple influencing factors related to the retention strategy selection and their corresponding weights can be determined first based on the product category information associated with the current return and refund order and the pre-defined mapping relationship between product categories and influencing factor information. Then, the score information of the current return and refund order is determined based on its basic feature information, and this score information is defined as the extended feature information. Specifically, the score information can include: the score of the current return and refund order on each of the multiple influencing factors, and the total score determined based on the scores and weights of the multiple influencing factors. In other words, after receiving the current return and refund order information and acquiring the corresponding basic feature information, the above score information can be calculated, and thus this score information can be used as the extended feature information of the current return and refund order.

[0076] S303: Input the basic feature information and the extended feature information of the current return and refund order into the AI ​​decision model. The AI ​​decision model is used to determine multiple candidate historical return and refund orders from the historical return and refund order set that meet the similarity conditions of the basic features of the current return and refund order, and based on the extended feature information, determine the target historical return and refund order from the multiple candidate historical return and refund orders that meets the similarity conditions of the extended features of the current return and refund order, and determine the preferred retention plan bound to the target historical return and refund order as the recommended retention plan corresponding to the current return and refund order.

[0077] After obtaining the basic and extended feature information of the current return and refund order, it can be input into the AI ​​decision-making model. Based on this information, the AI ​​model can then determine a recommended retention strategy. Specifically, multiple candidate historical return and refund orders that meet the basic feature similarity criteria of the current order can be identified from the historical return and refund order set. For example, these could be historical return and refund orders that meet the similarity criteria in dimensions such as product category or buyer profile. Since there may be multiple such similar historical return and refund orders, the target historical return and refund order that meets the extended feature similarity criteria of the current order can be further identified from these multiple candidate historical return and refund orders based on the extended feature information. Because a preferred retention strategy has already been bound to a specific historical return and refund order during training, this preferred retention strategy bound to the target historical return and refund order can be determined as the recommended retention strategy for the current return and refund order.

[0078] It should be noted that the AI ​​decision-making model in this application embodiment can have multimodal information processing capabilities, that is, it can process information of various modalities such as text, images, and videos. For example, specific basic feature information may include user profile data described in text, as well as user-uploaded credential images, etc. The AI ​​decision-making model can process the above-mentioned information of different modalities such as text and images.

[0079] S304: Provide retention information to the user based on the recommended retention plan.

[0080] After determining the recommended retention strategy, you can provide the user with retention information. Additionally, you can provide options for indicating acceptance or rejection. For example, ... Figure 4 As shown, specific retention plans can be offered: a refund only plus an extra bonus. Users can also choose between "accept the retention plan" and "continue to apply for a refund," allowing them to select based on their specific circumstances.

[0081] If the user chooses to "accept the retention plan," compensation will be provided according to the specific compensation method corresponding to that plan, and the current request can be terminated without entering the formal return and refund process. If the user chooses to "continue applying for a refund," this embodiment can also provide a "one-stop managed" return and refund process. As mentioned earlier, the buyer no longer needs to select a logistics provider; instead, the platform can choose the logistics provider. Since the platform has pre-stored the logistics provider's contact information, it can provide the logistics provider with the buyer's address, platform warehouse or merchant warehouse address, weight, product title, etc. Then, the logistics provider picks up the goods, the buyer provides the goods to the logistics provider, and the platform's logistics department receives the logistics receipt, logistics route, and other information from the logistics provider. After receiving the returned goods in the platform warehouse, it can take photos, conduct quality inspections, etc., and then process the refund. This method reduces the operational costs for the buyer and also reduces the workload of the merchant's customer service staff (communication with the buyer during the return and refund process).

[0082] It's important to note that, in practice, the specific retention strategies available can be configured by the system operations staff. A visualization platform can be provided to allow operations personnel to flexibly configure retention strategies, budget thresholds (some compensation methods provided in the retention plan may require payment from the platform, or, if the goods are damaged during the return process in the subsequent one-stop managed return and refund process, the platform may need to compensate the merchant; therefore, the platform's budget in this regard can be configured), blacklists and whitelists (malicious refund users can be blacklisted and no longer pursued for retention), and adoption rate constraint thresholds, achieving "zero-code" strategy iteration.

[0083] In addition, after the system recommends retention strategies, it can also perform full-link data tracking, including tracking key indicators such as refund traffic, actual adoption rate of retention strategies, and GMV (Gross Merchandise Volume) recovery. If anomalies are found, it can automatically issue an alert to ensure that the optimization effect is measurable and the risks are controllable.

[0084] The specific objects monitored can include traffic, effectiveness, user experience, and system health. Traffic metrics include the number of refund requests per minute, exposure of retention strategies, and user churn rate. Effectiveness metrics include retention adoption rate (by product category / user level), average amount recovered, and the percentage of refunds only in the total number of returned items. User experience metrics include decision response time, image recognition failure rate, and the number of refunds associated with user complaints. System health metrics include model inference success rate, API (Application Programming Interface) error rate, and so on.

[0085] The aforementioned tracking data can be displayed on a real-time computing dashboard and can also be stored. This stored tracking data can be used to further optimize the AI ​​decision-making model. For example, the actual adoption rate of recommended retention solutions by users can influence the prediction model's prediction of the adoption rate of various retention solutions, and continue to train the AI ​​decision-making model, thereby forming a closed loop of "perception-decision-execution-optimization".

[0086] For example, in a specific implementation scheme, the above closed loop can be as follows: Figure 5As shown, when a buyer clicks the "Request Return and Refund" button on the user interface, the intelligent retention process can be initiated. In this process, the system first listens for refund messages through the API gateway at the access layer. It also guides the buyer to provide necessary information (such as the reason for the refund, product images, and notes) to more accurately understand their needs. After the information is submitted, the system can aggregate contextual data in real time and provide it to the intelligent decision-making center. The intelligent decision-making center can also obtain configuration data from the operations configuration backend, including specific optional retention plans and business rules such as budget control. Additionally, it can obtain extended feature information related to the factors influencing the selection of retention plans and the importance of these factors. This information can be transmitted to the intelligent decision-making center, where the AI ​​decision model comprehensively analyzes and generates personalized retention plans (such as discounts, exchange suggestions, or refunds only), providing a better retention solution while respecting the user's wishes. The retention plan can then be delivered to the user through channels such as an app (mobile application) or a PC page. If the user accepts the retention plan, the retention strategy can be executed at the execution layer, triggering the strategy's public actions. Additionally, a data and AI engine layer can be provided for analyzing work order data, including basic feature information of current return and refund orders. The real-time monitoring system can collect execution plans, adoption logs, rendering logs, etc., and this data can be fed back into the database for optimizing and training the AI ​​decision-making model. This optimization training may include adjusting the prediction results of the forecasting model used in the aforementioned training process based on the actual adoption rate, so as to retrain the AI ​​decision-making model using the adjusted prediction results.

[0087] According to Embodiment 2 of this application, after receiving the current return and refund order information submitted by the user, the basic feature information of the current return and refund order can be determined first. Additionally, extended feature information of the current return and refund order can be obtained based on the basic feature information. Then, the basic and extended feature information can be input into the AI ​​decision model to identify multiple candidate historical return and refund orders from the historical return and refund order set that meet the basic feature similarity criteria of the current return and refund order, and to identify the target historical return and refund order that meets the extended feature similarity criteria of the current return and refund order. The preferred retention plan bound to the target historical return and refund order is then determined as the recommended retention plan corresponding to the current return and refund order. Through the above method, the user intent and other information expressed behind the data can be mined, thereby making the retention plan recommended by the model attractive to buyers, improving the adoption rate, and achieving multiple optimizations in user satisfaction, merchant costs, and platform benefits.

[0088] It should be noted that the embodiments of this application may involve the use of user data. In practical applications, user-specific personal data may be used in the scheme described herein within the scope permitted by applicable laws and regulations, provided that it complies with the applicable laws and regulations of the country (e.g., with the user's explicit consent, with the user being properly notified, etc.).

[0089] Corresponding to the aforementioned Embodiment 2, this application also provides an apparatus for processing return and refund order information, which may include: The basic feature information receiving unit is used to determine the basic feature information of the current return and refund order after receiving the current return and refund order information submitted by the user; An extended feature information acquisition unit is used to acquire extended feature information of the current return and refund order based on the basic feature information. The extended feature information includes: influencing factors affecting the retention strategy selection result and the importance information of the influencing factors. The decision unit is used to input the basic feature information and the extended feature information of the current return and refund order into the AI ​​decision model. The AI ​​decision model is used to determine multiple candidate historical return and refund orders from the set of historical return and refund orders that meet the similarity conditions of the basic features of the current return and refund order, and to determine the target historical return and refund order from the multiple candidate historical return and refund orders that meets the similarity conditions of the extended features of the current return and refund order based on the extended feature information. The preferred retention plan bound to the target historical return and refund order is determined as the recommended retention plan corresponding to the current return and refund order. The retention information providing unit is used to provide retention information to the user based on the recommended retention plan.

[0090] Specifically, the extended feature information acquisition unit can be used for: Based on the product category information associated with the current return and refund order, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the current return and refund order is determined based on the basic feature information of the current return and refund order, and the score information is determined as the extended feature information; the score information includes: the score of the current return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

[0091] The AI ​​decision-making model is trained in the following way: The basic feature information of multiple historical return and refund orders is obtained and input into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders. Based on the adoption rate and the basic feature information of the historical return and refund orders, the after-sales cost rate of multiple retention strategies is used as a condition to determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, the historical return and refund orders are bound with the preferred retention strategy and the extended feature information and then saved. The extended feature information is obtained based on the basic feature information and includes the influencing factors that affect the selection result of the retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information are input as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate recommended retention solutions, and the AI ​​decision-making model is supervised training with the preferred retention solution as the optimization target.

[0092] Additionally, the device may also include: The data tracking unit is used to track data after providing retention information to users based on the recommended retention plan, including the actual adoption rate of the retention plan, in order to optimize the AI ​​decision-making model.

[0093] Specifically, optimizing the AI ​​decision-making model may include adjusting the prediction model's prediction of the user adoption rate based on the actual adoption rate.

[0094] Corresponding to Embodiment 1, this application also provides an AI decision-making model training device, which may include: The basic feature information acquisition unit is used to acquire basic feature information of multiple historical return and refund orders and input the basic feature information into the prediction model. The prediction model is used to predict the user's adoption rate of multiple retention solutions under the condition of the basic feature information of the historical return and refund orders. The preferred retention strategy acquisition unit is used to determine the preferred retention strategy corresponding to the historical return and refund order based on the adoption rate and the after-sales cost rate of multiple retention strategies, using the basic feature information of the historical return and refund order as conditions. After acquiring the extended feature information of the historical return and refund order, the unit binds the historical return and refund order with the preferred retention strategy and the extended feature information and saves them. The extended feature information is obtained based on the basic feature information and includes the influencing factors affecting the retention strategy selection result and the importance information of the influencing factors. The training data input unit is used to input the basic feature information and the extended feature information as training data into the artificial intelligence (AI) decision model, so as to guide the AI ​​decision model to generate a recommended retention plan, and to conduct supervised training of the AI ​​decision model with the preferred retention plan as the optimization target.

[0095] Specifically, the preferred retention scheme acquisition unit can be used for: Based on the price attribute-related information in the basic feature information of the historical return and refund orders, and the compensation method information corresponding to the multiple retention plans, the after-sales cost rate of the multiple retention plans is calculated. Based on the adoption rate and after-sales cost rate, retention plans with after-sales cost rate within the target range and adoption rate meeting preset conditions are determined as the preferred retention plans for historical return and refund orders.

[0096] Specifically, when obtaining the extended feature information of the historical return and refund order, it may include: Based on the product category information associated with the historical return and refund orders, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the historical return and refund order is determined based on the basic feature information, and the score information is determined as the extended feature information; the score information includes: the score of the historical return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

[0097] Additionally, the device may also include: The sorting unit is used to sort the optimality of multiple retention strategies based on their respective adoption rates and after-sales cost rates, so that the sorting results can also be input into the AI ​​decision-making model for decision-making.

[0098] The basic feature information of the historical return and refund order includes order information, refund data, information of the associated logistics order, information of the first user, and information of the second user, wherein the first user and the second user correspond to the users of the two parties in the transaction.

[0099] In addition, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in any of the foregoing method embodiments.

[0100] And an electronic device, comprising: One or more processors; and A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of the method described in any of the foregoing method embodiments.

[0101] A computer program product includes a computer program / computer executable instructions that, when executed by a processor in an electronic device, implement the steps of the method described in the foregoing method embodiments.

[0102] in, Figure 6 An exemplary architecture of an electronic device is shown, which may include a processor 610, a video display adapter 611, a disk drive 612, an input / output interface 613, a network interface 614, and a memory 620. The processor 610, video display adapter 611, disk drive 612, input / output interface 613, network interface 614, and memory 620 can communicate with each other via a communication bus 630.

[0103] The processor 610 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to achieve the technical solution provided in this application.

[0104] The memory 620 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 620 can store the operating system 621 for controlling the operation of the electronic device 600, and the basic input / output system (BIOS) for controlling the low-level operations of the electronic device 600. Additionally, it can store a web browser 623, a data storage management system 624, and a return and refund order information processing system 625, etc. The aforementioned return and refund order information processing system 625 can be the application program that specifically implements the aforementioned steps in this embodiment. In summary, when implementing the technical solution provided in this application through software or firmware, the relevant program code is stored in the memory 620 and executed by the processor 610.

[0105] Input / output interface 613 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0106] Network interface 614 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0107] Bus 630 includes a pathway for transmitting information between various components of the device, such as processor 610, video display adapter 611, disk drive 612, input / output interface 613, network interface 614, and memory 620.

[0108] It should be noted that although the above-described device only shows the processor 610, video display adapter 611, disk drive 612, input / output interface 613, network interface 614, memory 620, bus 630, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the solution of this application, and does not necessarily include all the components shown in the figures.

[0109] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0110] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0111] The foregoing has provided a detailed description of the return and refund order information processing, AI decision model training method, and electronic device provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are merely for the purpose of helping to understand the methods and core ideas of this application; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for processing return and refund order information, characterized in that, include: After receiving the current return and refund order information submitted by the user, determine the basic feature information of the current return and refund order; Based on the basic feature information, the extended feature information of the current return and refund order is obtained. The extended feature information includes: influencing factors that affect the selection result of retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information of the current return and refund order are input into the AI ​​decision model. The AI ​​decision model is used to determine multiple candidate historical return and refund orders from the set of historical return and refund orders that meet the similarity conditions of the basic features of the current return and refund order, and based on the extended feature information, determine the target historical return and refund order from the multiple candidate historical return and refund orders that meets the similarity conditions of the extended features of the current return and refund order, and determine the preferred retention plan bound to the target historical return and refund order as the recommended retention plan corresponding to the current return and refund order. Provide users with retention information based on the recommended retention strategy.

2. The method according to claim 1, characterized in that, The step of obtaining the extended feature information of the current return and refund order based on the basic feature information includes: Based on the product category information associated with the current return and refund order, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the current return and refund order is determined based on the basic feature information of the current return and refund order, and the score information is determined as the extended feature information; the score information includes: the score of the current return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

3. The method according to claim 1 or 2, characterized in that, The AI ​​decision-making model was trained in the following way: The basic feature information of multiple historical return and refund orders is obtained and input into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders. Based on the adoption rate and the basic feature information of the historical return and refund orders, the after-sales cost rate of multiple retention strategies is used as a condition to determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, the historical return and refund orders are bound with the preferred retention strategy and the extended feature information and then saved. The extended feature information is obtained based on the basic feature information and includes the influencing factors that affect the selection result of the retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information are input as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate recommended retention solutions, and the AI ​​decision-making model is supervised training with the preferred retention solution as the optimization target.

4. The method according to claim 3, characterized in that, Also includes: After providing retention information to users based on the recommended retention plan, data tracking is performed, including the actual adoption rate of the retention plan, in order to optimize the AI ​​decision-making model.

5. The method according to claim 4, characterized in that, The optimization of the AI ​​decision-making model includes: The prediction model's prediction of user adoption rate is adjusted based on the actual adoption rate, so that the AI ​​decision model can be optimized and trained based on the adjusted prediction results.

6. A method for training an AI decision-making model, characterized in that, include: The basic feature information of multiple historical return and refund orders is obtained and input into a prediction model. The prediction model is used to predict the user's adoption rate of multiple retention strategies under the condition of the basic feature information of the historical return and refund orders. Based on the adoption rate and the basic feature information of the historical return and refund orders, the after-sales cost rate of multiple retention strategies is used as a condition to determine the preferred retention strategy corresponding to the historical return and refund orders. After obtaining the extended feature information of the historical return and refund orders, the historical return and refund orders are bound with the preferred retention strategy and the extended feature information and then saved. The extended feature information is obtained based on the basic feature information and includes the influencing factors that affect the selection result of the retention strategy and the importance information of the influencing factors. The basic feature information and the extended feature information are input as training data into the artificial intelligence (AI) decision-making model to guide the AI ​​decision-making model to generate recommended retention solutions, and the AI ​​decision-making model is supervised training with the preferred retention solution as the optimization target.

7. The method according to claim 6, characterized in that, The process of determining the optimal retention strategy corresponding to historical return and refund orders includes: Based on the price attribute-related information in the basic feature information of the historical return and refund orders, and the compensation method information corresponding to the multiple retention plans, the after-sales cost rate of the multiple retention plans is calculated. Based on the adoption rate and after-sales cost rate, retention plans with after-sales cost rate within the target range and adoption rate meeting preset conditions are determined as the preferred retention plans for historical return and refund orders.

8. The method according to claim 6, characterized in that, The process of obtaining the extended feature information of the historical return and refund order includes: Based on the product category information associated with the historical return and refund orders, and the pre-set mapping relationship between product categories and influencing factor information, multiple influencing factors and their corresponding weights related to the selection of retention strategies are determined. The score information of the historical return and refund order is determined based on the basic feature information, and the score information is determined as the extended feature information; the score information includes: the score of the historical return and refund order on the multiple influencing factors, and the total score determined based on the scores and weight information of the multiple influencing factors.

9. The method according to claim 6, characterized in that, Also includes: Based on the adoption rate and after-sales cost rate of each retention strategy, the optimality of the multiple retention strategies is ranked, and the ranking results are also input into the AI ​​decision-making model for decision-making.

10. The method according to claim 6, characterized in that, The basic feature information of the historical return and refund order includes order information, refund data, information of the associated logistics order, information of the first user, and information of the second user, where the first user and the second user correspond to the users of the two parties in the transaction.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1 to 10.

12. An electronic device, characterized in that, include: One or more processors; as well as A memory associated with the one or more processors, the memory being used to store program instructions that, when read and executed by the one or more processors, perform the steps of the method according to any one of claims 1 to 10.

13. A computer program product comprising a computer program / computer executable instructions, characterized in that, When the computer program / computer-executable instructions are executed by a processor in an electronic device, they implement the steps of the method according to any one of claims 1 to 10.

Images