An e-commerce competitive selling customer service decision closed-loop system based on emotional attribution analysis

Abstract

This invention discloses a closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis, belonging to the field of e-commerce customer service technology. The system includes a data collection unit, an emotion attribution analysis module, a customer service decision support module, and a closed-loop optimization module. The data collection unit collects user behavior data across the entire process. The emotion attribution analysis module correlates user emotions with competitive marketing events and outputs attribution results. The customer service decision support module pushes attribution results and provides decision support information. The closed-loop optimization module tracks processing results and optimizes the attribution model and strategy library. This invention can locate emotion-triggered events in real time, shorten customer service response time, improve user experience, and simultaneously achieve system self-evolution and continuous optimization.

Description

Technical Field

[0001] This invention relates to the field of e-commerce customer service technology, specifically to an e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis. Background Technology

[0002] With the rapid development of e-commerce, product bidding activities on e-commerce platforms are becoming increasingly diverse. When users participate in these activities, situations such as failed bids, unsuccessful purchases, or discounts not taking effect may lead to dissatisfaction or questions, resulting in customer service inquiries. How to quickly pinpoint the root cause of the problem and assist customer service personnel in making accurate responses when users experience emotional fluctuations has become a crucial challenge for e-commerce bidding systems.

[0003] Currently, there are several technical solutions for user emotion recognition in customer service scenarios. For example, patent announcement number CN121234954A proposes a B2C customer service response system based on sentiment analysis. This system obtains the user's input statements in the customer service dialogue, deconstructs semantic conflicts to determine contextual polarity, and then determines the user's final emotional bias. Based on this emotional bias, it matches and adjusts the tone of the customer service response corpus to generate the response output. However, this solution is limited to sentiment analysis and response tone adjustment after the customer service dialogue has occurred, which is a passive response mode. It can only identify the category of user emotions and cannot locate the specific marketing event that triggered the emotion. Another example is patent announcement number CN120894105A, which proposes a consumer behavior analysis method and system for smart retail scenarios based on big data. This system analyzes consumer behavior data and provides personalized behavior prediction and recommendations by constructing a causal reasoning-driven behavior attribution model. This solution focuses on causal attribution analysis of consumer behavior and mainly serves offline application scenarios such as marketing decisions. It does not form a real-time linkage with customer service interaction and also cannot proactively provide specific attribution information when customer service intervenes.

[0004] The aforementioned existing technologies focus on two different technical directions: emotion recognition and behavioral attribution. They fail to organically integrate these two aspects and embed them into the real-time decision-making process of competitive marketing customer service. When users experience negative emotions during competitive marketing activities, traditional customer service systems often only intervene after the user initiates an inquiry. Customer service personnel need to communicate repeatedly with the user to confirm which stage of the competitive marketing process caused the problem, resulting in response delays, a decline in user experience, and difficulty in continuously optimizing customer service strategies based on historical processing results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis. This system uses an emotion attribution matching engine to locate the specific triggering events of user emotions in real time and provides decision-making assistance and closed-loop optimization support for customer service personnel.

[0006] To solve the above-mentioned technical problems, this invention provides the following technical solution: a closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis, the system comprising:

[0007] The data tracking and collection unit deploys standardized data tracking points throughout the entire lifecycle of the marketing campaign to collect user behavior data and contextual parameters at each stage.

[0008] The emotion attribution analysis module is used to monitor users' emotional fluctuations in real time and associate emotions with specific marketing events, outputting attribution analysis results;

[0009] The customer service decision support module is used to push the attribution analysis results to the customer service workbench and provide decision support information;

[0010] The closed-loop optimization module is used to track customer service processing results and optimize the attribution model and strategy library;

[0011] Through the coordinated operation of the data collection unit, the emotion attribution analysis module, the customer service decision support module, and the closed-loop optimization module, this system can form a complete closed loop from data collection to strategy optimization.

[0012] Furthermore, the emotion attribution analysis module calculates the association strength score between each candidate marketing event and the user's emotion using an attribution matching algorithm. This attribution matching algorithm includes semantic matching degree calculation, temporal proximity weight calculation, and contextual continuity factor calculation. The formula for calculating the association strength score is as follows: ,in For semantic matching scores, As a weight for time proximity, For continuous factor scores, , , These are the corresponding weight coefficients. By using an attribution matching algorithm to weight and fuse semantic matching degree, temporal proximity, and contextual continuity factors, the system can quantitatively evaluate the correlation strength between competitive marketing events and user emotions, providing a calculable technical basis for attribution localization.

[0013] Furthermore, the time proximity weight The calculation is performed using the exponential decay function, and the formula is as follows: , where is the decay coefficient, and t is the time difference between the time of the candidate event and the time of the emotion, in seconds. By setting an exponential decay function to calculate the time proximity weight, the system gives higher attribution weight to the competing events that are closer to the time of the emotion, thus improving the time sensitivity of attribution positioning.

[0014] Furthermore, the data tracking unit deploys standardized data tracking points throughout the entire lifecycle of the marketing campaign. This lifecycle includes the exposure stage, browsing stage, click stage, participation stage, bidding stage, payment stage, and after-sales stage. The standardized data tracking points for each stage include basic identifier fields, scenario context fields, and attribution parameters. By deploying standardized tracking points at each stage of the marketing campaign, the system can completely record the user's behavior data from exposure to after-sales, providing sufficient data support for sentiment attribution analysis.

[0015] Furthermore, the basic identification fields include user identifier, product identifier, activity identifier, timestamp, page path, and action type. The action type includes seven basic action codes: exposure, click, browsing, participation, bidding, payment, and after-sales service.

[0016] The scenario context fields include activity identifier, strategy identifier, and recommendation channel;

[0017] The attribution parameters include preceding actions, source channels, and exposure positions;

[0018] By incorporating multi-dimensional tracking points such as basic identifier fields, scenario context fields, and attribution parameters, the system can accurately reconstruct the user's operational path and contextual environment during marketing activities, thereby improving the accuracy of attribution analysis.

[0019] Furthermore, the emotion attribution analysis module includes a hybrid architecture of a pre-trained lightweight natural language processing model and a rule engine. The lightweight natural language processing model is fine-tuned on an e-commerce customer service dialogue corpus based on the BERT architecture, classifying user-input text into four emotion levels: positive, neutral, questioning, and dissatisfied, and outputting an emotion confidence score. The rule engine performs emotion determination as a fallback when the confidence score of the lightweight natural language processing model is lower than a preset threshold, using keyword matching. Through the hybrid architecture of the BERT fine-tuned model and the rule engine, the system can perform fallback determination using keyword matching when the model confidence is insufficient, ensuring the stability and robustness of emotion recognition in different scenarios.

[0020] Furthermore, the emotion attribution analysis module outputs the user identifier and timestamp from the emotion data packet, and uses the user identifier and timestamp as search criteria to retrieve all competitive sales event records of the user within a preset session time window from the attribution database maintained by the data collection unit. The search results are sorted according to the timestamp and constructed into an event sequence. Each event record in the event sequence includes an event identifier, event type code, event occurrence timestamp, associated product identifier, associated activity identifier, and event parameters. By retrieving competitive sales event records within a preset time window using the user identifier and timestamp as criteria and constructing an event sequence, the system can quickly obtain the complete event context before and after the moment the emotion occurs, improving the efficiency of attribution retrieval.

[0021] Furthermore, the customer service decision support module maintains a strategy library, which stores corresponding processing strategies using a combination of emotion category and attribution event type as the key. Each processing strategy includes a reassurance script template, action suggestions, and processing priority identifier.

[0022] The customer service decision support module uses the emotion category and the attribution event type with the highest confidence in the attribution results as the key to perform exact matching retrieval in the strategy library. When exact matching fails, it degenerates into fuzzy matching retrieval with the attribution event type as the key.

[0023] Through a two-level retrieval mechanism of precise matching and fuzzy matching in the strategy library, the system can provide accurate strategies when the confidence level is high and general strategies when the confidence level is low, thus ensuring the availability of customer service decision support.

[0024] Furthermore, the closed-loop optimization module evaluates the effectiveness of strategies in the strategy library using the sentiment conversion rate as an indicator. The sentiment conversion rate is configured as follows:

[0025] The emotion conversion rate is equal to the number of customer service conversations in a group of conversations where the user's emotion changes from questioning or dissatisfaction to neutral or positive, divided by the total number of conversations in that group.

[0026] The closed-loop optimization module periodically evaluates each strategy in the strategy library on an evaluation cycle basis.

[0027] Remove strategies with an emotion conversion rate below the first preset elimination threshold from the active strategy library;

[0028] Strategies with an emotion conversion rate higher than the second preset excellent threshold are marked as recommended strategies and displayed in a highlighted manner.

[0029] By regularly evaluating the effectiveness of each strategy in the strategy library using the emotion conversion rate as an indicator, the system can automatically identify and eliminate inefficient strategies and highlight and recommend efficient strategies, thus achieving continuous optimization and iteration of customer service strategies.

[0030] Furthermore, the closed-loop optimization module feeds back the strategy evaluation results to the emotion attribution analysis module, dynamically adjusting the weight coefficients and attribution bias values ​​in the attribution matching algorithm. This dynamic adjustment employs a multi-armed gambling machine algorithm. , , The three weight coefficients are adjusted in small steps, with each adjustment not exceeding the preset coefficient change limit. The emotion and event type combinations that are verified to be strongly correlated in the strategy evaluation are recorded in the high confidence mapping table of the attribution matching engine. In subsequent attribution analysis, a preset bias value is added to the matching score of the combination. By adjusting the attribution weight coefficients in small steps through the multi-armed gambling machine algorithm and combining the bias value update mechanism, the system can continuously optimize the accuracy and stability of the attribution model based on actual feedback.

[0031] Compared with existing technologies, this closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis has the following beneficial effects:

[0032] I. This invention organically integrates user emotion recognition technology with competitive marketing event attribution analysis technology. It deploys standardized data tracking points throughout the entire lifecycle of a competitive marketing activity to collect user behavior data across the entire process. Combining multi-dimensional attribution matching algorithms based on semantic matching, temporal proximity, and contextual continuity, it quantifies and evaluates the correlation between competitive marketing events and user emotions. This allows for real-time location of the specific competitive marketing event that triggered the user's emotional fluctuations, eliminating the need for customer service personnel to repeatedly communicate with users to confirm the root cause of the problem. This significantly shortens customer service response time and effectively improves the user's consultation experience during competitive marketing activities.

[0033] Second, this invention constructs a complete closed-loop system from data collection and analysis to decision support and strategy optimization. It uses a multi-armed gambling machine algorithm to dynamically adjust the weight coefficients of the attribution model and regularly evaluates and iterates the customer service strategy library with the emotion conversion rate as the core indicator. It can automatically identify and eliminate inefficient strategies while recommending efficient strategies, continuously improving the accuracy and stability of emotion attribution. As a result, the system's decision support capability gradually increases over time, effectively reducing the workload and cost of manual operation and maintenance.

[0034] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0036] Figure 1 This is a flowchart illustrating the overall closed-loop process of the system according to the present invention.

[0037] Figure 2 This is a flowchart of the customer service decision support module of the present invention;

[0038] Figure 3 This is a flowchart of the feedback adjustment process for the closed-loop optimization module of the present invention. Detailed Implementation

[0039] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0040] like Figure 1 As shown, this embodiment provides a closed-loop decision-making system for e-commerce competitive marketing customer service based on sentiment attribution analysis. This system is mainly applied to competitive marketing activities on e-commerce platforms, such as flash sales, group buying, auctions, or lotteries. The overall operation flow of the system is as follows:

[0041] First, data collection units deployed across various pages of the e-commerce platform are used to collect user behavior data in real time throughout the entire marketing campaign. This data is then transmitted to the emotion attribution analysis module. When users engage in text-based communication with customer service, this module analyzes the user's dialogue text in real time, identifies their emotional state, and, based on the collected behavioral data, pinpoints specific marketing events that may have triggered the current emotion.

[0042] Subsequently, the customer service decision support module pushes the analyzed attribution results, along with suggested handling strategies, to the customer service staff's workspace. After the customer service staff handles the user's issue based on this information, the closed-loop optimization module tracks the final processing result of the session and uses this result data to regularly update and optimize the emotion recognition model, attribution matching algorithm, and customer service strategy library, thereby achieving a complete closed loop from data collection and analysis support to strategy iteration.

[0043] The specific deployment and implementation method of the data collection unit are as follows:

[0044] In order to fully record the user's behavior during the competitive marketing campaign, this system has deployed standardized data tracking points in seven key stages of the campaign: exposure, browsing, clicking, participation, bidding, payment, and after-sales service.

[0045] Taking a typical flash sale as an example, the following events trigger event tracking: when a product is displayed to a user on the homepage, the exposure stage is triggered; when a user clicks on a product image to enter the details page, the click stage is triggered; when a user browses the event rules and product information, the browsing stage is triggered; when a user clicks the "Buy Now" button, the participation stage is triggered; and when a user submits a bid or order, the corresponding bidding or payment stage is triggered. Each event tracking point reports data containing three types of fields.

[0046] The basic identifier fields describe the fundamental attributes of the action, including: user identifier (to uniquely identify the user); product identifier (to indicate the product associated with the action); activity identifier (to indicate the current promotional activity); timestamp (to accurately record the moment the action occurred, with millisecond-level precision); page path (to record the specific page address where the action occurred); and action type (using seven basic action codes: exposure, click, view, participation, bidding, payment, and after-sales). The scenario context fields reconstruct the user's environment, including: activity identifier (to further clarify the activity context); strategy identifier (to record the recommendation or operational strategy that triggered this event); and recommendation channel (to indicate which entry point or ad placement the user entered the activity from). Attribution parameters track the source chain of the action, including: preceding actions (to record the user's action before the current action); source channel (to mark the traffic source, such as organic search or paid advertising); and exposure position (to record the specific display location of the activity or product on the page).

[0047] All collected data is stored in the attribution database in real time using user ID and timestamp as a combined index, so as to facilitate efficient retrieval by subsequent modules.

[0048] Specifically, the emotion attribution analysis module consists of two parts working collaboratively: an emotion recognition unit and an attribution matching engine. The emotion recognition unit employs a hybrid architecture combining a pre-trained lightweight natural language processing model with a rule engine. The lightweight natural language processing model is based on the BERT architecture and has been fine-tuned on approximately 10,000 e-commerce customer service dialogues labeled with emotion levels. This model can classify user-input text into four emotion levels: positive, neutral, questioning, and dissatisfied, and simultaneously output a confidence score between 0 and 1, with higher scores indicating more reliable classification results.

[0049] For example, if a user enters "This event is a scam, right? I clearly won the promotion but it says I didn't succeed," the model might output a skeptical tone with a confidence level of 0.92. To ensure stability across various scenarios, the system presets a confidence threshold, such as 0.6. When the model's output confidence level falls below this threshold, the system will activate the rule engine for a fallback assessment. The rule engine internally maintains a keyword list; for example, for expressing dissatisfaction, the list includes words like "garbage," "bad review," and "complaint"; for expressing skepticism, it includes words like "why," "unreasonable," and "fake."

[0050] When user input text matches these keywords, the rule engine outputs the corresponding sentiment category based on the keyword with the highest weight, assigning a low confidence level, such as 0.55. This hybrid architecture ensures that the system can still provide usable sentiment recognition results even in marginal situations where the model struggles to make a judgment.

[0051] Once the emotion recognition unit outputs an emotion data packet, the attribution matching engine begins operation. This data packet contains the user identifier, the timestamp of the emotion's occurrence, and the emotion category. Using the user identifier and timestamp as the core search criteria, the engine retrieves all marketing event records for that user within a preset session time window prior to the time the emotion occurred from the attribution database maintained by the data collection unit.

[0052] The length of this time window can be configured according to the characteristics of different promotional activities. For example, for fast-paced flash sales, the window length can be set to 5 minutes; for longer pre-sale activities, it can be set to 30 minutes. The retrieved records are strictly sorted from earliest to latest by timestamp, forming an event sequence. Each event record in the sequence contains an event identifier, event type code, event occurrence timestamp, associated product identifier, associated activity identifier, and event parameters.

[0053] Subsequently, for each candidate event in the event sequence, the attribution matching engine calculates its correlation strength score with the user's current emotion using an attribution matching algorithm. The formula for this algorithm is: .in, The engine assigns a semantic matching score. It encodes the type and key parameters of candidate events, such as a failed bid due to insufficient inventory, and calculates semantic similarity between these parameters and the emotion category output by the emotion recognition unit and the user's original text. A word vector model can be used to convert the text into vectors, and then cosine similarity can be calculated to obtain a semantic matching score between 0 and 1.

[0054] This is the time proximity weight. This weight is calculated using an exponential decay function, as shown in the formula: Here, is the decay coefficient, whose value determines the rate of time decay. It is usually set based on historical data statistical analysis, for example, it can be set to 0.1. t is the difference between the time the candidate event occurs and the time the emotion is generated, in seconds. For example, if an event occurs 10 seconds before the user experiences the emotion, then... The value is approximately 0.37; if it occurs 60 seconds earlier, the value drops to approximately 0.002, indicating that the closer an event is to the moment the emotion is generated, the higher its attribution weight.

[0055] This is the score for the context continuity factor. This factor is used to assess the coherence of the user behavior chain. If there is a logical break between the current candidate event and the previous event in the event sequence, for example, after a user browses product A, a payment event for product B suddenly appears, and there is no corresponding browsing or clicking record in between, then the payment event for product B is considered a break. The score will be lower. Conversely, for events with a complete behavioral chain and logical consistency, The scores were higher.

[0056] Weighting coefficient , , The initial values ​​for the three factors mentioned above, corresponding to their respective importance, can be obtained through offline historical data regression analysis, for example, set to 0.4, 0.4, and 0.2 respectively. By calculating the score for all candidate events, the event with the highest score is the most likely cause of the emotion. The attribution analysis module ultimately outputs an attribution result, which includes the emotion category, the type of attributable event with the highest confidence, and its detailed parameters.

[0057] like Figure 2 As shown, after receiving the attribution results from the emotion attribution analysis module, the customer service decision support module maintains a structured strategy library internally. This strategy library is essentially a key-value database table, where the keys are composite keys combining emotion category and attribution event type. For example, one key could be "dissatisfaction - failed bid," and another could be "questioning - discount not applied." Each key corresponds to a processing strategy, and this strategy data structure contains three parts: a reassurance script template, such as "I'm very sorry for the bad experience; I understand you're disappointed that your bid for [Event Name] failed"; action suggestions, such as "Query the event inventory log to confirm the inventory status and inform the user of the real reason"; and a processing priority indicator, such as "high," "medium," and "low."

[0058] Once the customer service decision support module receives the attribution results, it first performs a precise matching search, that is, it uses the emotion category and attribution event type in the attribution results as a combination key to search for a perfectly matching strategy in the strategy library.

[0059] If an exact match is found, the corresponding strategy is invoked directly. If an exact match fails, for example, if the strategy library does not yet have a strategy configured for the combination of dissatisfaction and failed bid, the system degenerates into a fuzzy matching search, using only the attribution event type, i.e., failed bid, as the key to retrieve a general handling strategy. In this case, the appeasement script template may weaken the specific description of the emotion.

[0060] The retrieved strategies, along with the original user information and sentiment attribution analysis results, will be pushed to a dedicated interface on the customer service workbench. This interface will highlight the processing priority, suggested responses, and action steps to help customer service personnel respond to users quickly and accurately.

[0061] The closed-loop optimization module is key to the system's self-evolution. This module continuously tracks the final processing result of each customer service session. One of its core functions is to evaluate the effectiveness of strategies in the strategy library using the emotion conversion rate as an indicator. The calculation steps for the emotion conversion rate are as follows: First, select an evaluation period, such as weekly. For a specific strategy in the strategy library, such as questioning-bid failure, extract all customer service session records processed using this strategy within the evaluation period. Count the number of sessions in which the user's initial emotion was questioning or dissatisfaction, and which, after customer service processing, the emotion changed to neutral or positive by the end of the session. The emotion conversion rate is the number of sessions that changed emotion divided by the total number of sessions in that group.

[0062] For example, if a question-and-bid failure strategy was used in 100 sessions in a week, and user sentiment turned neutral or positive at the end of 70 sessions, then the sentiment conversion rate for that strategy that week was 70%. The system will perform this evaluation process periodically. When the sentiment conversion rate of a strategy falls below the first preset elimination threshold for two consecutive evaluation periods, such as below 30%, the strategy will be automatically removed from the active strategy library and moved to a backup library for manual review.

[0063] Conversely, when a strategy's emotional conversion rate exceeds the second preset excellent threshold for two consecutive evaluation periods, such as exceeding 80%, the system will mark it as a recommended strategy and display it in a highlighted color or special icon in the strategy management interface of the customer service decision support module, indicating to the operations staff that this strategy is effective.

[0064] On the other hand, such as Figure 3 As shown, the closed-loop optimization module also feeds back the results of the strategy evaluation to the sentiment attribution analysis module to dynamically adjust the accuracy and stability of the attribution matching algorithm. The specific adjustment methods fall into two categories. The first category involves adjusting the weight coefficients in the attribution matching algorithm. , , The system employs multi-armed gambling machine algorithms, such as the Thompson sampling algorithm, to dynamically adjust these coefficients.

[0065] At the end of each evaluation period, the closed-loop optimization module provides a reward signal to the multi-armed gambling machine algorithm based on the accuracy of the attribution results within that period. The algorithm then fine-tunes the three weight coefficients accordingly, with each adjustment limited to a small range, such as no more than 0.05, to avoid drastic fluctuations in system performance.

[0066] The second category involves establishing and updating a high-confidence mapping table. When the closed-loop optimization module repeatedly verifies a strong correlation between a certain emotion and an event type through strategy evaluation—for example, if dissatisfaction and payment timeout events consistently show high conversion rates in multiple evaluations—the system records this combination (dissatisfaction-payment timeout) in the high-confidence mapping table of the attribution matching engine. In subsequent attribution analysis, when this emotion-event type combination is encountered again, the attribution matching engine adds a preset bias value, such as 0.15 points, to the original association strength score, making it easier to identify this combination as the final attribution result. Through this continuous feedback adjustment mechanism, the attribution accuracy and customer service processing efficiency of the entire system can gradually improve with runtime.

[0067] In summary, the system described in this embodiment, through the coordinated efforts of a data collection unit, an emotion attribution analysis module, a customer service decision support module, and a closed-loop optimization module, constructs a complete technical solution encompassing user behavior and emotion monitoring, root cause analysis, and assisted problem-solving and strategy self-optimization. Upon receiving user inquiries, the system proactively analyzes their historical behavioral patterns and quantifies the correlation between emotions and events, providing customer service personnel with accurate decision support information, rather than passively waiting for inquiries. Simultaneously, the system continuously validates and iterates its models and strategies using actual processing results, allowing its decision support capabilities to be progressively optimized in practical applications.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis, characterized in that, The system consists of: The data tracking and collection unit deploys standardized data tracking points throughout the entire lifecycle of the marketing campaign to collect user behavior data and contextual parameters at each stage. The emotion attribution analysis module is used to monitor users' emotional fluctuations in real time and associate emotions with specific marketing events, outputting attribution analysis results; The customer service decision support module is used to push the attribution analysis results to the customer service workbench and provide decision support information; The closed-loop optimization module is used to track customer service processing results and optimize the attribution model and strategy library.

2. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 1, characterized in that, The emotion attribution analysis module calculates the correlation strength score between each candidate marketing event and the user's emotion using an attribution matching algorithm. This algorithm includes semantic matching degree calculation, temporal proximity weight calculation, and contextual continuity factor calculation. The formula for calculating the correlation strength score is as follows: ,in For semantic matching scores, As a weight for time proximity, For continuous factor scores, , , These are the corresponding weighting coefficients.

3. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 2, characterized in that, The time proximity weight The calculation is performed using the exponential decay function, and the formula is as follows: , where is the attenuation coefficient, and t is the time difference between the occurrence of the candidate event and the generation of the emotion, in seconds.

4. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 1, characterized in that, The data tracking unit deploys standardized data tracking points throughout the entire lifecycle of the marketing campaign. The entire lifecycle includes the exposure stage, browsing stage, click stage, participation stage, bidding stage, payment stage, and after-sales stage. The standardized data tracking points for each stage include basic identifier fields, scenario context fields, and attribution parameters.

5. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 4, characterized in that, The basic identification fields include user identifier, product identifier, activity identifier, timestamp, page path, and action type. The action type includes seven basic action codes: exposure, click, view, participation, bidding, payment, and after-sales service. The scenario context fields include activity identifier, strategy identifier, and recommendation channel; The attribution parameters include preceding actions, source channels, and exposure positions.

6. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 1, characterized in that, The emotion attribution analysis module includes a hybrid architecture of a pre-trained lightweight natural language processing model and a rule engine. The lightweight natural language processing model is based on the BERT architecture and fine-tuned on an e-commerce customer service dialogue corpus. It classifies the user input text into four emotion levels: positive, neutral, questioning, and dissatisfied, and outputs an emotion confidence score. The rule engine performs emotion determination as a fallback when the confidence score of the lightweight natural language processing model is lower than a preset threshold, using keyword matching.

7. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 1, characterized in that, The emotion attribution analysis module outputs the user identifier and timestamp from the emotion data packet, and uses the user identifier and timestamp as search criteria to retrieve all competitive event records of the user within a preset session time window from the attribution database maintained by the data collection unit. The search results are sorted according to the timestamp and constructed into an event sequence. Each event record in the event sequence includes an event identifier, event type code, event occurrence timestamp, associated product identifier, associated activity identifier, and event parameters.

8. The e-commerce competitive marketing customer service decision-making closed-loop system based on emotion attribution analysis according to claim 1, characterized in that, The customer service decision support module maintains a strategy library. The strategy library stores the corresponding processing strategies using a combination of emotion category and attribution event type as the key. Each processing strategy includes a reassurance script template, action suggestions, and processing priority identifier. The customer service decision support module uses the emotion category and the attribution event type with the highest confidence in the attribution results as the key to perform exact matching retrieval in the strategy library. When exact matching fails, it degenerates into fuzzy matching retrieval using the attribution event type as the key.

9. A closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis according to claim 1, characterized in that, The closed-loop optimization module uses the sentiment conversion rate as an indicator to evaluate the effectiveness of strategies in the strategy library. The sentiment conversion rate is configured in the following steps: The emotion conversion rate is equal to the number of customer service conversations in a group of conversations where the user's emotion changes from questioning or dissatisfaction to neutral or positive, divided by the total number of conversations in that group. The closed-loop optimization module periodically evaluates each strategy in the strategy library on an evaluation cycle basis. Remove strategies with an emotion conversion rate below the first preset elimination threshold from the active strategy library; Strategies with an emotion conversion rate higher than the second preset excellent threshold are marked as recommended strategies and displayed in a highlighted manner.

10. A closed-loop decision-making system for e-commerce competitive marketing customer service based on emotion attribution analysis according to claim 1, characterized in that, The closed-loop optimization module feeds back the strategy evaluation results to the emotion attribution analysis module, dynamically adjusting the weight coefficients and attribution bias values ​​in the attribution matching algorithm. This dynamic adjustment employs a multi-armed gambling machine algorithm. , , The three weight coefficients are adjusted in small increments, with each adjustment not exceeding the preset coefficient change limit. The emotion and event type combinations that are verified to be strongly correlated in the strategy evaluation are recorded in the high confidence mapping table of the attribution matching engine, and a preset bias value is added to the matching score of the combination in subsequent attribution analysis.

Images