A Multi-Agent Dialogue Recommendation System Based on Empirical Rules and Regret Perception

Abstract

This application provides a multi-agent dialogue recommendation system based on empirical rules and regret perception. A candidate generation and perception module generates a set of candidate movies with a base score based on the user's current natural language query and historical behavior data. An empirical rule extraction module extracts reusable empirical rules from historical recommendation interaction data and stores them in an empirical memory. A counterfactual simulation module uses the empirical memory to perform risk assessment on the set of candidate movies with the base score, obtaining risk assessment information for each candidate movie. A multi-agent fusion ranking module uses the risk assessment information to re-rank and score the set of candidate movies with the base score. This invention, through empirical rule extraction, regret perception counterfactual simulation, and multi-agent fusion ranking, accurately identifies users' implicit preferences, avoids the risk of regret after acceptance, reduces deployment costs, and significantly improves recommendation adaptability and ranking quality.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence and recommender system technology, specifically to a multi-agent dialogue recommender system based on empirical rules and regret perception. Background Technology

[0002] Conversational recommendation systems need to understand users' interests, taboos, and viewing scenarios from multi-turn conversations, and then output ranked candidate recommendations based on this understanding. Existing mainstream technologies can be broadly categorized into two types: one is efficient ranking methods centered on dense retrieval and cross-encoders. These methods have strong semantic relevance modeling capabilities but are insufficient at recognizing implicit, fine-grained pragmatic preferences and taboo constraints expressed by users in conversations; the other type directly utilizes large language models for large-scale list-level ranking. While these methods possess strong reasoning capabilities, their deployment costs increase with the size of the candidate pool and the length of metadata, resulting in high engineering deployment costs. Further research indicates that high-frequency failure patterns in conversational recommendations are not entirely irrelevant, but rather superficially semantically related but mismatched in implicit preferences such as rhythm, emotion, sensitive content, and viewing scenarios. This leads to the exclusion of true target items from the recommendation results, causing a decrease in recall and cumulative gain of normalized loss.

[0003] Therefore, how to make the content recommended by the recommendation system more closely match user needs has become a problem that needs to be solved. Summary of the Invention

[0004] The purpose of this application is to address the shortcomings of the existing technology and provide a multi-agent dialogue recommendation system that can effectively identify users' implicit preferences, avoid the risk of users regretting their choices after acceptance, reduce deployment costs, and make recommended content more suitable for users' real needs.

[0005] A multi-agent dialogue recommendation system based on empirical rules and regret perception includes: a candidate generation and perception module, an empirical rule extraction module, a counterfactual simulation module, and a multi-agent fusion ranking module; The candidate generation and perception module is used to search and rank the movie library based on the user's current natural language query and the user's historical behavior data to obtain a set of candidate movies with a base score. The empirical rule extraction module is used to identify user behavior patterns of regret after acceptance from historical recommendation interaction data, and extract reusable empirical rules from them and store them in the empirical memory bank. The counterfactual simulation module is used to extract empirical rules from the experience memory bank to perform risk assessment on each candidate movie in the candidate movie set with a base score, thereby obtaining risk assessment information corresponding to each candidate movie. The risk assessment information includes a regret risk value, a compatibility value, a decision result, and audit information. The regret risk value is used to characterize the degree to which the user is willing to regret their decision. The compatibility value is used to characterize the degree to which the candidate movie matches the user's needs. The decision result is used to characterize whether the candidate movie passes the risk verification. The audit information is traceable information on the counterfactual reasoning process, reasoning basis, parameter source, and judgment logic. The multi-agent fusion ranking module is used to re-rank and score the candidate movie set with a base score based on the risk assessment information corresponding to each candidate movie to determine the final recommendation list and the final score corresponding to each recommendation result.

[0006] Furthermore, the multi-agent dialogue recommendation system based on empirical rules and regret perception, as described above, also includes: a weak signal risk detection module and a user interaction module; The user interaction module is used to display recommended content to the user based on the final recommendation list, and to obtain the user's online reply text to the recommended content and the current dialogue context; The weak signal risk detection module is used to calculate the online risk level and reasoning explanation of the recommended content based on the online reply text, the current dialogue context and the corresponding risk assessment information, and to feed back the online risk level to the multi-agent fusion ranking module. The multi-agent fusion ranking module is also used to dynamically calibrate candidate movies that have not yet been displayed based on the online risk level, and adjust their final scores and ranking order.

[0007] Furthermore, in the multi-agent dialogue recommendation system based on empirical rules and regret perception as described above, the weak signal risk detection module includes: a pragmatic analysis unit, a risk determination unit, a feedback unit, and a feedback accumulation unit; The pragmatic analysis unit is used to identify potential high-risk and low-risk weak signals in the online response text; the high-risk weak signals include: brief and vague expressions of acceptance, hesitant expressions, topic shifting, vague agreement, and continuing to ask for other options; the low-risk weak signals include: making specific comments on recommended movies, expressing clear interest, actively asking for details about the movie, or showing obvious positive emotions; The risk determination unit is used to calculate the online risk level of the recommended content and generate a corresponding reasoning explanation based on the low-risk weak signal, combined with the current dialogue context and the audit information of the current recommendation result; the online risk level is used to characterize whether the user's current response may be a case of superficial acceptance but insufficient real interest; The feedback unit is used to feed back the online risk level to the multi-agent fusion ranking module; The feedback accumulation unit is used to record user response text, online risk level, reasoning explanation, and final interaction result into the system log.

[0008] Furthermore, in the multi-agent dialogue recommendation system based on experience rules and regret perception as described above, the experience rule extraction module includes: a regret sample identification unit, an experience rule generation unit, a rule verification unit, and an experience rule storage unit. The regret sample identification unit is used to identify regret samples from the user's historical recommendation interaction data using preset feedback tags to obtain a regret sample set; the historical recommendation interaction data is obtained from the system log and includes: user historical behavior data, historical dialogue context, recommended movies, user feedback tags, and corresponding movie library metadata; The empirical rule generation unit is used to generate a draft empirical rule based on the regret sample set; the draft empirical rule includes: a pattern key, rule text, slot set, conflict type, evidence, and initial confidence level; the pattern key is used to represent the conditional pattern that triggers the regret behavior, and the rule text is used to represent the recommendation constraint or ranking adjustment strategy that should be adopted under the conditional pattern; The rule verification unit is used to correct the confidence level of the draft empirical rules to obtain the cleaned empirical rules. The empirical rule storage unit is used to store the cleaned empirical rules into the empirical memory bank.

[0009] Furthermore, in the multi-agent dialogue recommendation system based on empirical rules and regret perception as described above, the user interaction module is also used to receive the user's current natural language query and obtain the user's historical recommendation interaction data based on the user identifier; The counterfactual simulation module includes: a user constraint parsing unit, an empirical rule retrieval unit, a candidate matching unit, a counterfactual simulation unit, and a rule gating unit; The user constraint parsing unit is used to perform structured processing on the user's current natural language query and historical dialogue context to obtain a user-defined slot set and a natural language constraint description. The defined slot set includes the user's intent, preferred genres, disliked genres, preferred film tone, disliked film tone, preferred film pacing, strongly avoided content, and viewing scenario. The empirical rule retrieval unit is used to perform empirical rule retrieval in the experience memory using the user specification slot set and natural language description as retrieval conditions to obtain a target empirical rule set; the empirical rule retrieval includes: firstly performing a first-stage vector recall to recall previous rules similar to the current user context from the experience memory. The empirical rule is followed; then the second stage of fine-tuning and retrieval is performed to retrieve the previous results. The empirical rules were used to score the relevance of the results, and the top results were obtained. The set of empirical rules is the target empirical rule set; The candidate matching unit is used to match the movie library metadata of the candidate movie with the triggering conditions of each empirical rule in the target rule set to filter out the triggering rule set corresponding to the current candidate movie; wherein, if the slot condition of an empirical rule matches the current user's standard slot set, and the evidence in the empirical rule can be supported in the movie library metadata of the current candidate movie, then it is determined that the empirical rule is triggered by the current candidate movie. The counterfactual simulation unit is used to perform counterfactual reasoning on each candidate movie based on the user-defined slot set, the movie library metadata of the candidate movies, and the trigger rule set corresponding to the candidate movies, so as to obtain the initial risk assessment information corresponding to each candidate movie. The rule gating unit is used to perform evidence constraint processing on the initial risk assessment information to obtain the final risk assessment information; The audit information also records the retrieval process of the target experience rules, the rule matching process, and the trace information related to counterfactual reasoning.

[0010] Furthermore, in the multi-agent dialogue recommendation system based on empirical rules and regret perception as described above, the evidence constraint processing includes: When a triggering rule fails evidence verification or the regret risk value is lower than a preset risk threshold, the corresponding triggering rule will be regarded as reference information and will not be used as a hard blocking basis; when a triggering rule passes evidence verification and the regret risk value is higher than a preset risk threshold, the triggering rule will be regarded as a valid risk rule to constrain the initial risk assessment information. The evidence verification includes matching the evidence with the metadata of the movie library corresponding to the recommended movie to confirm whether the evidence can be supported by the movie synopsis, tags, content hints or attribute fields.

[0011] Furthermore, in the multi-agent dialogue recommendation system based on experience rules and regret perception as described above, the multi-agent fusion ranking module includes: a basic score normalization unit, a fusion decision unit, a ranking output unit, and a result organization unit. The basic score normalization unit is used to perform normalization processing on the basic score of candidate movies to obtain normalized basic scores; The fusion decision unit is used to calculate the calibration score, risk penalty item, and final score of candidate movies based on the normalized base score, compatibility value, and regret risk value; wherein, the risk penalty item is calculated based on the risk value and a preset risk threshold to suppress high-risk candidates; the final score is determined by the base correlation and the risk calibration signal. The sorting output unit is used to re-sort the candidate movies according to their final scores to obtain a final recommendation list, select the top K candidate movies from the final recommendation list as recommendation results, and simultaneously output the audit information corresponding to each recommendation result; the audit information includes: final score, base score, calibration score, risk penalty item, risk value, compatibility value, triggering rule, evidence, and recommendation reason; The result organization unit is used to organize the top K candidate movies into a structured recommendation result, and also simultaneously outputs the risk explanation, compatibility explanation and rule triggering basis for each recommendation result.

[0012] Furthermore, in the multi-agent dialogue recommendation system based on experience rules and regret perception as described above, the experience memory is continuously updated according to the user's online reply text, the online risk level and reasoning explanation corresponding to the recommended content, and each user interaction behavior.

[0013] Furthermore, in the multi-agent dialogue recommendation system based on empirical rules and regret perception as described above, the candidate generation and perception module includes: a dense retrieval unit and a fine ranking unit; The dense retrieval unit is used to encode the user's current natural language query and the user's historical behavior data, and to perform vector retrieval in the movie library to obtain a Top-N candidate set, wherein each candidate movie includes: identifier, basic attributes, metadata and semantic vector; The fine-ranking unit is used to form an input pair between the user's current natural language query and the candidate movie text, sort each candidate movie by rating, and finally generate a set of candidate movies with a base score.

[0014] Furthermore, in the multi-agent dialogue recommendation system based on empirical rules and regret perception as described above, the historical behavior data includes user's historical viewing records, historical ratings, historical acceptance or rejection records, historical preference tags, and historical dialogue summaries; the movie library metadata includes movie identifiers, movie names, genres, plot summaries, actors, directors, tags, film tone, film pacing, sensitive content prompts, and pre-computed semantic vectors.

[0015] Beneficial effects: This application achieves efficient retrieval and ranking of user queries and historical behaviors through a candidate generation and perception module, solving the problems of insufficient implicit preference identification and low ranking accuracy in traditional dense retrieval. Through an empirical rule extraction module, reusable rules for accepting and regretting decisions are extracted from historical interactions, addressing the inability of traditional methods to utilize historical failure lessons and the recurrence of implicit mismatches. Through a counterfactual simulation module, the regret risk and compatibility of candidate movies are quantitatively assessed, solving the problem of semantically related but implicitly mismatched preferences and declining core indicators in dialogue recommendation. Through a multi-agent fusion ranking module, the fusion and calibration of basic scores and risk signals are achieved, solving the problems of high ranking costs, difficult engineering deployment, and unreliable recommendation results in traditional large-scale models. Attached Figure Description

[0016] Figure 1 A schematic diagram of the structure of a multi-agent dialogue recommendation system based on empirical rules and regret perception provided in this application; Figure 2 The flowchart of the multi-agent collaborative recommendation method based on empirical rules and regret perception provided in this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions 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, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The dialogue recommendation system provided in this application is used for movie recommendations. Current recommendation systems typically equate user acceptance of a recommendation with successful recommendation, failing to recognize situations where a user accepts the recommendation initially but subsequently regrets it.

[0019] However, in the annotations of public dialogue recommendation datasets, accepting recommendations does not always represent an ideal successful experience. It includes at least four different feedback scenarios: only acceptance with a high rating can be considered a truly successful recommendation; the remaining scenarios—giving a low to medium rating after acceptance, losing interest during viewing, or not completing the purchase—reveal that users regret accepting the recommendation. This phenomenon does not reflect explicit user rejection, but rather more subtle and implicit information about preference mismatch. Existing recommendation technologies generally lack systematic modeling and processing mechanisms for this problem.

[0020] Furthermore, existing recommender systems typically lack the following capabilities: First, they lack a mechanism to extract structured preference rules from historical regret cases, making it difficult to distill historical failures into reusable knowledge; second, they lack a mechanism to externalize empirical rules into a memory bank in a searchable and generalizable manner and participate in online ranking; third, they lack a simulator for counterfactual inference regarding the risk of user acceptance but regret, failing to simultaneously output risk values, compatibility, triggering rules, and their evidence; fourth, they lack a real-time identification mechanism for the difference between polite acceptance and genuine interest in user online feedback, making further dynamic ranking calibration difficult. Therefore, this application provides a multi-agent dialogue recommender system based on empirical rules and regret perception to address the above deficiencies. Figure 1 The schematic diagram of the multi-agent dialogue recommendation system based on empirical rules and regret perception provided in this application is as follows: Figure 1 As shown, the system includes: a candidate generation and perception module 101, an experience rule extraction module 102, a counterfactual simulation module 103, and a multi-agent fusion ranking module 104.

[0021] The candidate generation and perception module 101 is used to search and rank movies in a movie library based on the user's current natural language query and historical behavior data to obtain a set of candidate movies with a base score. The candidate generation and perception module 101 includes a dense retrieval unit and a ranking unit. The dense retrieval unit encodes the user's current natural language query and historical behavior data, and performs vector retrieval in the movie library to obtain a Top-N candidate set. Each candidate movie includes an identifier, basic attributes, metadata, and a semantic vector. The ranking unit forms an input pair between the user's current natural language query and the candidate movie text, sorts each candidate movie by rating, and finally generates a set of candidate movies with a base score.

[0022] Specifically, the candidate generation and perception module, as the system's perception center, receives user input and generates a preliminary candidate set, providing the basic data structure for subsequent empirical rule retrieval and simulation. This module includes the following units and operation steps: (1) Dense retrieval unit Receive user natural language queries and historical behavioral data The query is encoded into a vector representation using an encoding model. A high-dimensional vector search is performed in the movie database to obtain a Top-N candidate set. This includes the identifiers, basic attributes, metadata, and semantic vectors of the candidate movies.

[0023] (2) Fine arrangement unit The user query and candidate movie text are paired as inputs, which are then fed into the Cross-Encoder model for scoring. The candidate set is output in sorted order of scores, and a base score is generated for each candidate for subsequent fusion.

[0024] The experience rule extraction module 102 is used to identify user behavior patterns of regret after acceptance from historical recommendation interaction data, and extract reusable experience rules from them and store them in the experience memory bank.

[0025] The experience rule extraction module serves as the system's experience hub, used to identify user behavior patterns of regret after acceptance from historical recommendation interaction data, extract reusable experience rules, and store them in the experience memory for subsequent counterfactual simulation module to call; it includes: a regret sample identification unit, an experience rule generation unit, a rule verification unit, and an experience rule storage unit.

[0026] The regret sample identification unit is used to identify regret samples from the user's historical recommendation interaction data using preset feedback tags to obtain a regret sample set. The historical recommendation interaction data is obtained from the system log and includes: user historical behavior data, historical dialogue context, recommended movies, user feedback tags, and corresponding movie library metadata. The experience rule generation unit is used to generate a draft experience rule based on the regret sample set. The draft experience rule includes: a mode key, rule text, slot set, conflict type, evidence, and initial confidence level. The mode key is used to represent the conditional pattern that triggers the regret behavior, and the rule text is used to represent the recommendation constraint or ranking adjustment strategy to be taken under the conditional pattern. The rule verification unit is used to correct the confidence level of the draft experience rule to obtain a cleaned experience rule. The experience rule storage unit is used to store the cleaned experience rule into the experience memory bank.

[0027] Specifically, the regret sample identification unit is used to acquire historical recommendation interaction data, which includes: user dialogue context, recommended movies, user feedback tags, and corresponding movie library metadata; it identifies regret samples based on preset feedback tags, wherein moderate rating acceptance, no interest acceptance, and other types of acceptance are identified as regret samples after acceptance, thus obtaining a regret sample set. The user's original needs, recommendation results, feedback results, and corresponding movie library metadata are extracted from the regret sample set as input for subsequent empirical rule generation. The empirical rule generation unit is used to invoke the debriefing agent for each regret sample in the regret sample set, and to perform debriefing analysis based on the user's original needs, recommended movie content, and feedback results; and to generate an empirical rule draft, which includes: condition pattern identifier, rule concise text, slot set, conflict type, evidence, and initial confidence level; wherein, the condition pattern identifier is used to represent the condition pattern that triggers the regret behavior, and the rule concise text is used to represent the recommendation constraint or ranking adjustment strategy that should be taken under the condition pattern.

[0028] The rule verification unit is used to clean the slot values ​​in the draft empirical rules, deleting empty slots, illegal slots, and unrecognizable slot values, and aligning free expressions into a preset slot system; it performs fact verification on the evidence by matching the evidence with the metadata of the movie library corresponding to the recommended movie to confirm whether the evidence can be supported by the movie synopsis, tags, content hints, or attribute fields; it standardizes and corrects the conflict types by merging free text into preset conflict types; and it performs confidence correction on the draft empirical rules by combining slot integrity, evidence authenticity, and conflict type identifiability to obtain the cleaned final empirical rules.

[0029] The experience rule storage unit is used to store the cleaned final experience rules into the experience memory; the condition pattern identifier and the rule condensation text are vectorized and stored together with the corresponding metadata of the rules; the metadata includes: regret type, conflict type, evidence, rule confidence and source sample identifier; so as to support the experience rule retrieval and fine ranking search in the subsequent online recommendation stage.

[0030] The counterfactual simulation module 103 is used to extract empirical rules from the experience memory bank to perform risk assessment on each candidate movie in the candidate movie set with a base score, thereby obtaining risk assessment information corresponding to each candidate movie; the risk assessment information includes regret risk value, compatibility value, adjudication result, and audit information; the regret risk value is used to characterize the degree to which the user is willing to regret; the compatibility value is used to characterize the degree of matching between the candidate movie and the user's needs; the adjudication result is used to characterize the qualitative result of the user's potential regret; the audit information is traceable information of the counterfactual reasoning process, reasoning basis, parameter source, and judgment logic.

[0031] The counterfactual simulation module, as the reasoning center of the system, is used to perform regret risk assessment and compatibility assessment on candidate movies, and to determine whether there is a risk of superficial acceptance but subsequent regret if the current candidate movie is recommended to the current user. It includes: user constraint parsing unit, empirical rule retrieval unit, candidate matching unit, counterfactual simulation unit, and rule gating unit.

[0032] The system provided in this application includes a user interaction module for receiving a user's current natural language query and obtaining the user's historical recommended interaction data based on the user identifier. A user constraint parsing unit performs structured processing on the user's current natural language query and historical dialogue context to obtain a user-defined slot set and a natural language constraint description. An empirical rule retrieval unit uses the user-defined slot set and the natural language description as retrieval conditions to perform empirical rule retrieval in the experience memory to obtain a target empirical rule set. The empirical rule retrieval includes: firstly, performing a first-stage vector recall to recall previous rules similar to the current user context from the experience memory. The empirical rule is followed; then the second stage of fine-tuning and retrieval is performed to retrieve the previous results. The empirical rules were used to score the relevance of the results, and the top results were obtained. The set of empirical rules is defined as a target empirical rule set. The candidate matching unit is used to match the movie library metadata of the candidate movie with the triggering conditions of each empirical rule in the target rule set to filter out the triggering rule set corresponding to the current candidate movie. If the slot condition of an empirical rule matches the current user's normative slot set, and the evidence in the empirical rule can be supported by the movie library metadata of the current candidate movie, then it is determined that the empirical rule is triggered by the current candidate movie. The counterfactual simulation unit is used to perform counterfactual reasoning on each candidate movie according to the user's normative slot set, the movie library metadata of the candidate movie, and the triggering rule set corresponding to the candidate movie, so as to obtain the initial risk assessment information corresponding to each candidate movie. The rule gating unit is used to perform evidence constraint processing on the initial risk assessment information to obtain the final risk assessment information. The audit information also records the retrieval process of the target empirical rule, the rule matching process, and the trace information related to counterfactual reasoning. The evidence constraint processing includes: when the triggering rule fails evidence verification, or the regret risk value is lower than a preset risk threshold, the corresponding triggering rule is regarded as reference information and not used as a hard blocking basis; when the triggering rule passes evidence verification and the regret risk value is higher than the preset risk threshold, the triggering rule is regarded as a valid risk rule to perform evidence constraint processing on the initial risk assessment information; the evidence verification includes: matching the evidence with the movie library metadata of the corresponding recommended movie to confirm whether the evidence can be supported in the movie introduction, tags, content prompts or attribute fields.

[0033] Specifically, the user constraint parsing unit uses a pattern key aligner to parse the natural language input into structured slots, including user intent, preferred genres, disliked genres, preferred film tone, preferred film pacing, and strong avoidance criteria. If LLM parsing fails or returns a null value, it reverts to the heuristic rule extraction method to ensure system robustness. The slot set is then transformed into a canonical slot set and a natural language description is generated for subsequent experience retrieval and simulator input.

[0034] The empirical rule retrieval unit is used to perform empirical rule retrieval in the empirical memory using a set of canonical slots and natural language descriptions as retrieval conditions. First, it performs a first-stage vector recall, retrieving rules from the empirical memory that are similar to the current user context. 1. Empirical rules; then perform the second stage of fine-tuning and retrieval, scoring the relevance of the recalled empirical rules to obtain the top [number of results]. The empirical rules serve as the target rule set. ; The candidate matching unit is used to read the film library metadata of the top-M candidate films; and to match the film library metadata of each candidate film with the target rule set. The triggering conditions of each empirical rule are matched; rules that are actually effective for the current candidate movie are filtered out, resulting in the set of triggering rules corresponding to the candidate movie; if the slot conditions of a certain empirical rule itself match the current user specification slot set obtained by the user constraint parsing unit, and the evidence in the empirical rule can be supported by the movie library metadata of the current candidate movie, then it is determined that the empirical rule is triggered by the current candidate movie. The slot conditions of the empirical rule are the scene adaptation conditions fixed during the offline generation and verification process of the empirical rule. The current user specification slot set is the user demand feature obtained after structured parsing of the user's current natural language query and historical dialogue context. The evidence support is that the evidence content in the empirical rule can form a corresponding match with the movie library metadata (including movie introduction, tags, content hints and attribute fields) of the current candidate movie.

[0035] The counterfactual simulation unit is used to perform counterfactual reasoning on each candidate movie, and calculate its initial regret risk value, initial compatibility value, initial decision result, and initial audit information. The regret risk value is used to represent the probability or risk level of regret if the user accepts the candidate movie, and the compatibility value is used to represent the degree of matching between the candidate movie and the current user's explicit and implicit preferences.

[0036] The rule gating unit is used to perform evidence constraint processing on the counterfactual simulation results. When a triggering rule fails evidence verification or the regret risk value is lower than the preset risk threshold, the corresponding triggering rule is regarded as reference information and not used as a hard blocking basis. When a triggering rule passes evidence verification and the regret risk value is higher than the preset risk threshold, the triggering rule is used as a valid risk rule to participate in the subsequent ranking calibration. Finally, the final regret risk value, compatibility value, adjudication result and audit information of the candidate movie are output for the subsequent multi-agent fusion ranking module to call.

[0037] The multi-agent fusion ranking module 104 is used to re-rank and score the candidate movie set with a base score based on the risk assessment information corresponding to each candidate movie to determine the final recommendation list and the final score corresponding to each recommendation result. As the decision-making execution center of the system, the multi-agent fusion ranking module is used to fuse the base score, regret risk, and compatibility signals to output the final ranking result; it includes: a base score normalization unit, a fusion decision unit, a ranking output unit, and a result organization unit.

[0038] The system comprises the following components: a base score normalization unit, which normalizes the base scores of candidate movies to obtain normalized base scores; a fusion decision unit, which calculates the calibration score, risk penalty, and final score of candidate movies based on the normalized base score, compatibility value, and regret risk value; wherein the risk penalty is calculated based on the risk value and a preset risk threshold to suppress high-risk candidates; and the final score is determined by both the base correlation and the risk calibration signal. A sorting output unit re-sorts candidate movies based on their final scores to obtain a final recommendation list, selects the top K candidate movies from the final recommendation list as recommendation results, and simultaneously outputs the audit information corresponding to each recommendation result; the audit information includes: final score, base score, calibration score, risk penalty, risk value, compatibility value, triggering rule, evidence, and recommendation reason; and a result organization unit, which organizes the top K candidate movies into a structured recommendation result and simultaneously outputs the risk explanation, compatibility explanation, and rule triggering basis for each recommendation result.

[0039] Specifically, the multi-agent fusion ranking module, as the decision-making execution center of the system, is used to fuse the base score, regret risk, and compatibility signals to output the final ranking result. The base score normalization unit normalizes the base scores of candidate movies to obtain a normalized base score, eliminating differences in scoring scales between different candidates and providing a unified dimension for subsequent multi-signal fusion. The fusion decision unit calculates the calibration score, risk penalty, and final score of candidate movies based on the normalized base score n, compatibility value, and risk value. The risk penalty is calculated based on the risk value and a preset risk threshold to suppress high-risk candidates; the final score is determined by the base correlation and risk calibration signal and is used for subsequent ranking. The ranking output unit re-ranks the candidate movies according to their final scores to obtain the final candidate list; outputs the top-K candidate movies as recommendations, and simultaneously outputs audit information corresponding to each recommendation; the audit information includes: final score, base score, calibration score, risk penalty, risk value, compatibility value, trigger rule, evidence, and recommendation reason. The results organization unit is used to organize the top-K candidate results output by the sorting output unit into a structured recommendation result; so that the system not only outputs a list of recommended movies, but also synchronously outputs the risk explanation, compatibility explanation and rule triggering basis for each recommendation result; so as to facilitate subsequent user interaction, log recording and weak signal risk detection module calls.

[0040] Furthermore, the system provided in this application also includes a weak signal risk detection module. The user interaction module is used to display recommended content to the user based on the final recommendation list and to obtain the user's online reply text and the current dialogue context regarding the recommended content; the weak signal risk detection module is used to calculate the online risk level and reasoning explanation of the recommended content based on the online reply text, the current dialogue context, and the corresponding risk assessment information, and to feed back the online risk level to the multi-agent fusion ranking module; the multi-agent fusion ranking module is also used to dynamically calibrate candidate movies that have not yet been displayed based on the online risk level, adjusting their final scores and ranking order.

[0041] The weak signal risk detection module includes: a pragmatic analysis unit, a risk determination unit, a feedback unit, and a feedback accumulation unit. The pragmatic analysis unit is used to identify potential high-risk and low-risk weak signals in the online response text. High-risk weak signals include: brief and vague expressions of acceptance, hesitant expressions, topic shifts, vague agreement, and continuing to ask for other options. Low-risk weak signals include: making specific evaluations of recommended movies, expressing clear interest, actively asking for details about the movie, or showing obvious positive emotions. The risk determination unit is used to calculate the online risk level of the recommended content and generate corresponding reasoning explanations based on the low-risk weak signals, combined with the current dialogue context and audit information of the current recommendation results. The online risk level is used to characterize whether the user's current response may be a case of superficial acceptance but insufficient genuine interest. The feedback unit is used to feed back the online risk level to the multi-agent fusion ranking module. The feedback accumulation unit is used to record the user's response text, online risk level, reasoning explanation, and final interaction result in the system log.

[0042] Specifically, the weak signal risk detection module is used to analyze user responses in real time during the online interaction phase, identify the difference between polite acceptance and genuine interest, and feed the analysis results back to the multi-agent fusion ranking module for dynamic adjustment of subsequent recommendation results. It includes: a pragmatic analysis unit, a risk determination unit, a feedback unit, and a feedback accumulation unit. The pragmatic analysis unit receives the user's response text to the current recommendation result and a summary of the corresponding dialogue context; it analyzes the pragmatic features in the response text to identify potential high-risk and low-risk weak signals. High-risk weak signals include: brief and ambiguous expressions of acceptance, hesitant statements, topic shifts, vague agreement, and continuing to request other options; low-risk weak signals include: making specific evaluations of recommended movies, expressing clear interest, actively asking about movie details, or showing obvious positive emotions. The risk determination unit calculates the online risk level and generates corresponding reasoning explanations based on the weak signal features identified by the pragmatic analysis unit, combined with the current dialogue context and audit information of the current recommendation result. The online risk level characterizes whether the user's current response may be a case of superficial acceptance but insufficient genuine interest. The feedback unit feeds the online risk level back to the multi-agent fusion ranking module; dynamically calibrates candidate movies that have not yet been displayed based on the risk level, adjusting their final scores or ranking order; when a high-risk weak signal is detected, it increases the risk penalty weight of candidates similar to the current high-risk candidate, or elevates the ranking position of more stable candidates; when a low-risk weak signal is detected, it maintains the current ranking result or further enhances the recommendation priority of similar candidates. The feedback accumulation unit records the user's reply text, online risk level, reasoning explanation, and final interaction result in the system log; and uses it as candidate training samples for the subsequent experience rule extraction module, participating in the continuous updating of experience rules; thus forming a closed-loop optimization mechanism of online recognition—offline accumulation—re-call.

[0043] Furthermore, the experience memory is continuously updated based on the user's online reply text, the online risk level and reasoning explanation corresponding to the recommended content, and each user interaction behavior.

[0044] The multi-agent collaborative recommendation method based on experience rules and regret perception described in this invention includes an offline experience rule construction process and an online dialogue recommendation process during system operation. The offline experience rule construction process is used to form a reusable experience memory, and the online dialogue recommendation process is used to generate recommendation results based on the user's current natural language input, and to further accumulate user feedback into historical behavior data that can be used subsequently. Figure 2 The flowchart of the multi-agent collaborative recommendation method based on empirical rules and regret perception provided in this application is as follows: Figure 2 As shown, the method includes the following steps: Step 1: Build an experience memory bank.

[0045] Historical recommendation interaction data is acquired, including user's historical dialogue context, historical recommended movies, user feedback tags, subsequent user ratings or interest feedback, and corresponding movie library metadata. Based on user feedback tags, samples of user acceptance followed by regret are identified and input into the empirical rule extraction module. The empirical rule extraction module performs a retrospective analysis on these samples, generates empirical rules, and performs slot cleaning, evidence verification, conflict type correction, and confidence level correction on these rules. The corrected empirical rules are stored in an empirical memory bank as a source of retrieveable empirical rules for subsequent online recommendation stages.

[0046] Step 2: Obtain recommended input for the current conversation During the online recommendation phase, the user's current natural language query is received. and obtain user historical behavior data. Current dialogue context And film library metadata; including historical behavior data. This includes user viewing history, historical ratings, historical accept or reject records, historical preference tags, and historical conversation summaries; the movie library metadata includes movie identifiers, movie names, genres, plot summaries, actors, directors, tags, film tone, film pacing, sensitive content prompts, and pre-computed semantic vectors.

[0047] Step 3: Generate a candidate movie set and base scores The candidate generation and perception module will generate the user's current query With historical behavioral data The data is fused into the retrieval input and then input into the encoding model to obtain the query vector. Using the query vector High-dimensional vector retrieval is performed in the film database to obtain an initial set of candidate films. ; will display the user's current query With the initial candidate movie set Each candidate movie text is constructed as an input pair and fed into a ranking model for scoring, yielding a base score for each candidate movie. Based on these base scores, the candidate movies are initially ranked to obtain a set of candidate movies with base scores. .

[0048] Step S4: Parse user constraints and generate slot information The user's current query The data is parsed into structured slots, resulting in a slot set. This slot set is then normalized to obtain a standardized slot set. Natural language constraint descriptions are generated based on the standardized slot set. The standardized slot set includes user intent, preferred genres, disliked genres, preferred film tone, disliked film tone, preferred film pacing, strongly avoided content, and viewing scenarios.

[0049] Step 5: Retrieve empirical rules that match the current user context Using the specified slot set and natural language constraint description as search criteria, empirical rule retrieval is performed in the empirical memory.

[0050] Step 6: Generate a set of candidate triggering rules Needle belt basic score candidate movie collection For each candidate film, read its corresponding film library metadata; then combine the candidate film library metadata with the target empirical rule set. The triggering conditions of each empirical rule are matched; if the slot condition of an empirical rule matches the current user's standard slot set, and the evidence in the empirical rule can be supported by the movie library metadata of the candidate movie, then the candidate movie is determined to trigger the empirical rule; thus, the triggering rule set corresponding to each candidate movie is obtained.

[0051] Step 7: Perform counterfactual simulation and generate risk calibration information For each candidate movie, perform counterfactual reasoning to determine whether recommending the candidate movie to the current user could lead to the user superficially accepting it but subsequently regretting it; calculate the regret risk value, compatibility value, decision result, and audit information for each candidate movie.

[0052] Step 8: Combine the base score, risk value, and compatibility value to generate the final ranking. The base scores are normalized to obtain normalized base scores; based on the normalized base scores, compatibility values, and regret risk values, risk penalty items, calibration scores, and final scores are calculated; candidate movies are reordered based on the final scores to obtain the final recommendation list. The final recommended list. The final score information, together with the audit information, is presented to the user as a recommendation explanation output and is subsequently used by the weak signal risk detection module to determine whether the user feedback is consistent with the recommended risk.

[0053] Step 9: Output recommendation results and audit information The system outputs the final recommendation list to the user. Simultaneously, it outputs the audit information corresponding to each candidate movie, where the recommendation results are used for user interaction.

[0054] Step 10: Detect weak signals in user online feedback Pragmatic analysis is performed on the online reply text r to identify whether the user has high-risk weak signals such as polite acceptance, hesitant acceptance, topic shifting, vague agreement, or continuing to ask for other options, or low-risk weak signals such as clear interest, active follow-up questions, and positive evaluation; Based on the weak signal identification results, the current dialogue context, and the audit information of the recommendation results, the online risk level is calculated, and a risk assessment explanation is generated.

[0055] Step 11: Dynamically calibrate subsequent sorting based on weak user feedback The multi-agent fusion ranking module receives online risk levels. If the online risk level indicates that the user has polite acceptance or insufficient genuine interest in the current recommendation results, the risk penalty weight of candidates similar to the current high-risk candidates is increased, or the ranking position of more stable candidates is improved. If the online risk level indicates that the user has a clear interest in the current recommendation results, the current ranking result is maintained, or the compatibility weight of candidates similar to the current positive candidates is increased. The final score of the candidate movies that have not yet been displayed is recalculated based on the dynamically calibrated weights, and the subsequent recommendation ranking is updated.

[0056] Step 12: Accumulate user feedback and form a closed-loop update. The system writes the user's online reply text r, online risk level, risk assessment explanation, current recommendation result, final interaction result, and corresponding audit information into historical behavior data u; If subsequent user feedback shows that they have given low ratings to the accepted recommendations, lost interest, or failed to complete the purchase, then the interaction record will be marked as a new candidate regret sample. The updated historical behavior data u' is used as the historical behavior input in step 2 during the next round of online recommendation; the interaction records marked as regret samples can also enter the offline experience rule construction process in step 1 to update the experience memory bank.

[0057] This application has the following advantages over the prior art: First, it breaks through the traditional view that accepting a recommendation equates to its success, effectively identifying implicit regret behaviors such as low ratings, loss of interest, and incomplete viewing after accepting a recommendation. By specifically designing an empirical rule extraction module, it mines implicit features of user preference mismatches from historical recommendation interaction data and precipitates them as reusable empirical rules, thus overcoming the shortcomings of existing technologies in systematically modeling implicit user preferences, fine-grained taboo constraints, and post-event regret behaviors.

[0058] Secondly, a knowledge accumulation mechanism was constructed for offline extraction, storage, and online retrieval of experience rules. This mechanism solidifies fragmented dialogue scenarios, viewing preferences, sensitive taboos, and reasons for regret into standardized experience rules and stores them in an experience memory bank, enabling continuous iteration and reuse of rules. It can accurately match personalized user constraints without relying on large-scale, real-time, high-intensity inference models, significantly reducing online inference overhead and engineering deployment costs.

[0059] Third, a counterfactual simulation mechanism is introduced to conduct a preliminary risk assessment of candidate movies. Based on empirical rules in the experience memory bank, counterfactual reasoning is performed on each candidate movie to quantify regret risk values, compatibility values, adjudication results, and audit information. This mechanism can predict potential user regret before recommendation output, accurately identify implicit preference mismatches such as pacing, emotion, sensitive content, and viewing scenarios, and avoid ineffective recommendations that are semantically relevant but provide an unpleasant actual experience.

[0060] Fourth, a multi-agent fusion ranking strategy combining basic scores and risk assessment information is adopted. This involves a secondary re-ranking and scoring process that integrates the initial search ranking base scores of candidate movies with multi-dimensional assessment information such as regret risk and preference matching. This effectively avoids the problem of high-quality target recommendations being squeezed out by semantically similar but implicitly mismatched candidates, significantly improving the recall rate and normalized loss cumulative gain of the recommendation results, and enhancing the overall experience of dialogue recommendation.

[0061] Fifth, by setting up a complete record of the counterfactual reasoning process, reasoning basis, parameter sources, and judgment logic for audit information, the entire process of each risk assessment and recommendation decision is traceable and reviewable, which facilitates subsequent iteration of experience rules, optimization of recommendation strategies, and behavioral attribution analysis, ensuring the long-term maintainability and continuous optimization capabilities of the system.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-agent dialogue recommendation system based on empirical rules and regret perception, characterized in that, include: Candidate generation and perception module, empirical rule extraction module, counterfactual simulation module, multi-agent fusion and ranking module; The candidate generation and perception module is used to search and rank the movie library based on the user's current natural language query and the user's historical behavior data to obtain a set of candidate movies with a base score. The empirical rule extraction module is used to identify user behavior patterns of regret after acceptance from historical recommendation interaction data, and extract reusable empirical rules from them and store them in the empirical memory bank. The counterfactual simulation module is used to extract empirical rules from the experience memory bank to perform risk assessment on each candidate movie in the candidate movie set with a base score, thereby obtaining risk assessment information corresponding to each candidate movie. The risk assessment information includes regret risk value, compatibility value, ruling result, and audit information; The regret risk value is used to characterize the degree to which a user is willing to regret their decision; the compatibility value is used to characterize the degree to which the candidate movie matches the user's needs. The decision result is used to characterize whether the candidate film passes the risk verification; the audit information is traceable information on the counterfactual reasoning process, reasoning basis, parameter source and judgment logic; The multi-agent fusion ranking module is used to re-rank and score the candidate movie set with a base score based on the risk assessment information corresponding to each candidate movie to determine the final recommendation list and the final score corresponding to each recommendation result.

2. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 1, characterized in that, The system also includes: a weak signal risk detection module and a user interaction module; The user interaction module is used to display recommended content to the user based on the final recommendation list, and to obtain the user's online reply text to the recommended content and the current dialogue context; The weak signal risk detection module is used to calculate the online risk level and reasoning explanation of the recommended content based on the online reply text, the current dialogue context and the corresponding risk assessment information, and to feed back the online risk level to the multi-agent fusion ranking module. The multi-agent fusion ranking module is also used to dynamically calibrate candidate movies that have not yet been displayed based on the online risk level, and adjust their final scores and ranking order.

3. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 2, characterized in that, The weak signal risk detection module includes: a pragmatic analysis unit, a risk determination unit, a feedback unit, and a feedback accumulation unit; The pragmatic analysis unit is used to identify potential high-risk and low-risk weak signals in the online response text; the high-risk weak signals include: brief and vague expressions of acceptance, hesitant expressions, topic shifting, vague agreement, and continuing to ask for other options; the low-risk weak signals include: making specific comments on recommended movies, expressing clear interest, actively asking for details about the movie, or showing obvious positive emotions; The risk determination unit is used to calculate the online risk level of the recommended content and generate a corresponding reasoning explanation based on the low-risk weak signal, combined with the current dialogue context and the audit information of the current recommendation result; the online risk level is used to characterize whether the user's current response may be a case of superficial acceptance but insufficient real interest; The feedback unit is used to feed back the online risk level to the multi-agent fusion ranking module; The feedback accumulation unit is used to record user response text, online risk level, reasoning explanation, and final interaction result into the system log.

4. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 3, characterized in that, The empirical rule extraction module includes: a regret sample identification unit, an empirical rule generation unit, a rule verification unit, and an empirical rule storage unit; The regret sample identification unit is used to identify regret samples from the user's historical recommendation interaction data using preset feedback tags to obtain a regret sample set; the historical recommendation interaction data is obtained from the system log and includes: user historical behavior data, historical dialogue context, recommended movies, user feedback tags, and corresponding movie library metadata; The empirical rule generation unit is used to generate a draft empirical rule based on the regret sample set; the draft empirical rule includes: a pattern key, rule text, slot set, conflict type, evidence, and initial confidence level; the pattern key is used to represent the conditional pattern that triggers the regret behavior, and the rule text is used to represent the recommendation constraint or ranking adjustment strategy that should be adopted under the conditional pattern; The rule verification unit is used to correct the confidence level of the draft empirical rules to obtain the cleaned empirical rules. The empirical rule storage unit is used to store the cleaned empirical rules into the empirical memory bank.

5. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 3, characterized in that, The user interaction module is also used to receive the user's current natural language query and obtain the user's historical recommendation interaction data based on the user identifier; The counterfactual simulation module includes: a user constraint parsing unit, an empirical rule retrieval unit, a candidate matching unit, a counterfactual simulation unit, and a rule gating unit; The user constraint parsing unit is used to perform structured processing on the user's current natural language query and historical dialogue context to obtain a user-defined slot set and a natural language constraint description. The defined slot set includes the user's intent, preferred genres, disliked genres, preferred film tone, disliked film tone, preferred film pacing, strongly avoided content, and viewing scenario. The empirical rule retrieval unit is used to perform empirical rule retrieval in the experience memory using the user specification slot set and natural language description as retrieval conditions to obtain a target empirical rule set; the empirical rule retrieval includes: firstly performing a first-stage vector recall to recall previous rules similar to the current user context from the experience memory. The empirical rule is followed; then the second stage of fine-tuning and retrieval is performed to retrieve the previous results. The empirical rules were used to score the relevance of the results, and the top results were obtained. The set of empirical rules is the target empirical rule set; The candidate matching unit is used to match the movie library metadata of the candidate movie with the triggering conditions of each empirical rule in the target rule set to filter out the triggering rule set corresponding to the current candidate movie; wherein, if the slot condition of an empirical rule matches the current user's standard slot set, and the evidence in the empirical rule can be supported in the movie library metadata of the current candidate movie, then it is determined that the empirical rule is triggered by the current candidate movie. The counterfactual simulation unit is used to perform counterfactual reasoning on each candidate movie based on the user-defined slot set, the movie library metadata of the candidate movies, and the trigger rule set corresponding to the candidate movies, so as to obtain the initial risk assessment information corresponding to each candidate movie. The rule gating unit is used to perform evidence constraint processing on the initial risk assessment information to obtain the final risk assessment information; The audit information also records the retrieval process of the target experience rules, the rule matching process, and the trace information related to counterfactual reasoning.

6. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 5, characterized in that, The evidence constraint processing includes: When a triggering rule fails evidence verification or the regret risk value is lower than a preset risk threshold, the corresponding triggering rule will be regarded as reference information and will not be used as a hard blocking basis; when a triggering rule passes evidence verification and the regret risk value is higher than a preset risk threshold, the triggering rule will be regarded as a valid risk rule to constrain the initial risk assessment information. The evidence verification includes matching the evidence with the metadata of the movie library corresponding to the recommended movie to confirm whether the evidence can be supported by the movie synopsis, tags, content hints or attribute fields.

7. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 3, characterized in that, The multi-agent fusion ranking module includes: a basic score normalization unit, a fusion decision unit, a ranking output unit, and a result organization unit; The basic score normalization unit is used to perform normalization processing on the basic score of candidate movies to obtain normalized basic scores; The fusion decision unit is used to calculate the calibration score, risk penalty item, and final score of candidate movies based on the normalized base score, compatibility value, and regret risk value; wherein, the risk penalty item is calculated based on the risk value and a preset risk threshold to suppress high-risk candidates; the final score is determined by the base correlation and the risk calibration signal. The sorting output unit is used to re-sort the candidate movies according to their final scores to obtain a final recommendation list, select the top K candidate movies from the final recommendation list as recommendation results, and simultaneously output the audit information corresponding to each recommendation result; the audit information includes: final score, base score, calibration score, risk penalty item, risk value, compatibility value, triggering rule, evidence, and recommendation reason; The result organization unit is used to organize the top K candidate movies into a structured recommendation result, and also simultaneously outputs the risk explanation, compatibility explanation and rule triggering basis for each recommendation result.

8. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 4, characterized in that, The experience memory is continuously updated based on the user's online reply text, the online risk level and reasoning explanation corresponding to the recommended content, and each user interaction behavior.

9. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 5, characterized in that, The candidate generation and perception module includes: a dense retrieval unit and a fine ranking unit; The dense retrieval unit is used to encode the user's current natural language query and the user's historical behavior data, and to perform vector retrieval in the movie library to obtain a Top-N candidate set, wherein each candidate movie includes: identifier, basic attributes, metadata and semantic vector; The fine-ranking unit is used to form an input pair between the user's current natural language query and the candidate movie text, sort each candidate movie by rating, and finally generate a set of candidate movies with a base score.

10. The multi-agent dialogue recommendation system based on empirical rules and regret perception according to claim 4, wherein the historical behavior data includes user historical viewing records, historical ratings, historical acceptance or rejection records, historical preference tags, and historical dialogue summaries; and the movie library metadata includes movie identifiers, movie names, genres, plot summaries, actors, directors, tags, film tone, film pacing, sensitive content prompts, and pre-computed semantic vectors.

Images