An adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity perception

Abstract

The present disclosure is a kind of adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity perception, comprising: structured modeling of user intention and candidate space;Multi-dimensional task complexity calculation;Agent dynamic arrangement based on four-quadrant mapping;Agent collaborative execution and result verification;Slot aggregation based on task characteristics.This embodiment can improve the performance of the system in complex actual scenarios where the candidate pool is limited and the user preferences are diverse, using a structured diversity guarantee mechanism based on slot level role constraints, an agent dynamic arrangement mechanism based on multi-dimensional complexity perception, and a closed-loop anti-illusion system based on candidate constraint graph.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness. Background Technology

[0002] With the maturation of large language model technology, recommender systems are evolving from traditional collaborative filtering to an agent-based generative paradigm. Existing multi-agent recommender frameworks (such as MACRec and MACRS) typically employ centralized task distribution and static collaborative processes. While these systems integrate the expertise of different agents to some extent, they have revealed several technical shortcomings that urgently need to be addressed in practical applications.

[0003] First, existing systems suffer from rigid collaborative topologies, lacking deep task awareness and dynamic orchestration capabilities. Existing "adaptive" techniques are often limited to adjusting weighted parameters of each agent's output, rather than dynamically reconstructing collaborative relationships and data flows between agents at the policy level. This means that regardless of task difficulty, the system must activate all pre-defined agents and execute the complete interaction process. This "one-size-fits-all" static orchestration leads to severe computational resource redundancy and inference latency: for simple tasks, the system allocates excessive computing power; while for tasks with implicitly complex constraints (such as "recommend a non-Marvel movie starring an actor who plays Iron Man"), existing systems often rely on shallow heuristics such as text length or simple binary classifications ("difficult / easy") for judgment, lacking a fine-grained perception of the task's inherent logical complexity, resulting in insufficient inference depth and difficulty in generating accurate planned paths.

[0004] Secondly, there is an irreconcilable structural contradiction between accuracy and diversity. Limited by the "probability maximization" generation mechanism of LLM, single-agent or homogeneous multi-agent systems are prone to falling into the "optimal trap," tending to recommend highly popular top items, resulting in highly homogeneous recommendation results with low coverage and diversity. Existing attempts to improve diversity (such as post-processing rearrangement or simple polling strategies) usually sacrifice recommendation accuracy, lacking an endogenous mechanism that can structurally guarantee distribution diversity within the feature space without sacrificing accuracy.

[0005] Finally, large language models commonly suffer from hallucination and constraint alignment failure in closed-domain recommendations. Unlike open-domain question answering, recommender systems require outputs to strictly correspond to real items in the candidate pool. However, existing LLM-based generative recommender systems often lack strict constraints and backtracking verification mechanisms for the candidate space, leading to the model frequently generating non-existent "fake items" or recommending products that do not meet the user's hard attribute constraints (such as price range or brand restrictions), severely impacting the system's credibility and user experience.

[0006] Therefore, there is an urgent need in this field for a novel multi-agent recommendation method that can precisely perceive the complexity of multi-dimensional tasks, adaptively reconstruct collaborative topologies to match task requirements, and systematically solve the problem of balancing accuracy and diversity while ensuring factual constraints.

[0007] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0008] The purpose of this invention is to provide an adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness, thereby overcoming, to at least to some extent, one or more problems caused by the limitations and defects of related technologies.

[0009] This invention first provides an adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness, comprising: S1. Structured modeling of user intent and candidate space: Receive user queries and historical data, retrieve relevant candidate items from the item library, and construct a candidate constraint graph containing item entities and their attribute relationships; S2. Calculation of multidimensional task complexity: Based on the candidate constraint graph, calculate the first index reflecting the difficulty of matching user intent with system resources and the second index reflecting the dispersion of user interests to obtain the task complexity coordinates. S3. Dynamic orchestration of agents based on four-quadrant mapping: Based on the preset quadrant in which the task complexity coordinates fall, dynamically determine the target agents to be activated and the interaction strategies between them, and generate the orchestration result. S4. Agent Cooperative Execution and Result Verification: Activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and use the connected candidate constraint graph to verify the existence and semantic matching of the generated results. S5. Slot aggregation based on task features: Based on the quadrant where the task complexity coordinates are located, configure differentiated role weights for different positions in the verified recommendation list, and generate the final recommendation list through dynamic weighted aggregation.

[0010] In this invention, in S1, the candidate constraint graph is a heterogeneous graph. The nodes of the candidate constraint graph include item entity nodes and attribute feature nodes. The edges of the candidate constraint graph represent the association between the item and its attributes. The candidate constraint graph is used to construct a data structure to support subsequent restricted decoding and verification.

[0011] In this invention, in step S2, the calculation process of the first indicator includes: calculating the distance between the user intent vector and the overall distribution center of the candidate items in the semantic space, and using this distance as the first indicator; the calculation process of the second indicator includes: analyzing the distribution uniformity of the user's historical behavior on the attribute network of the candidate constraint graph, and using the distribution uniformity as the second indicator; and using the first indicator and the second indicator as coordinates of task complexity.

[0012] In this invention, S3, the process of dynamically determining the activated target agents and their interaction strategies based on the preset quadrant into which the task complexity coordinates fall, includes: When the task complexity coordinate falls into the first quadrant, it indicates that the task matching difficulty is high and the interest dispersion is high. The interaction strategy adopted is to activate all types of intelligent agents and conduct multiple rounds of high-intensity debate and verification. When the task complexity coordinate falls into the second quadrant, it indicates that the task matching difficulty is high but the interest dispersion is low. The interaction strategy adopted is to prioritize activating agents that focus on mining long-tail resources and perform deep constraint verification. When the task complexity coordinate falls into the third quadrant, it indicates that the task matching difficulty is low and the interest dispersion is low. The interaction strategy adopted is to activate only the single expected optimal agent to respond quickly. When the task complexity coordinate falls into the fourth quadrant, it indicates that the task matching difficulty is low but the interest dispersion is high. The interaction strategy adopted is to activate all types of generated agents and conduct medium-to-low intensity debates focusing on diversity.

[0013] In this invention, in S4, the intelligent agent includes: an analytical intelligent agent for parsing user profiles, a trend guide focused on popularity, a novel explorer focused on long-tail resources, a personalized matcher focused on personal historical preferences, and a critical evaluator responsible for connecting candidate constraint graphs to verify the results.

[0014] In this invention, S5, the slot aggregation process based on task characteristics includes: Define each sequential position in the validated recommendation list as a slot; Configure a weight coefficient matrix for each slot, tailored to different agents; Iteratively, for each slot, select the item with the highest product of its base score and the corresponding agent weight from the validated recommendation list and fill it.

[0015] In this invention, the configuration strategy of the weight coefficient matrix is ​​dynamically adjusted according to the quadrant in which the task complexity coordinate falls. The configuration strategy includes: increasing the weight of the mining agent in the front slot in resource-scarce scenarios; making the items of different agents appear alternately at the front of the list in user intent-ambiguous scenarios; and degenerating to global sorting based on basic scores in simple scenarios.

[0016] The present invention further provides an adaptive multi-agent collaborative recommendation system based on multi-dimensional task complexity awareness, the system comprising: The cognitive perception module is used to receive user queries and historical data, retrieve relevant candidate items from the item library, construct a candidate constraint graph containing item entities and their attribute relationships based on the candidate items, and calculate a first index reflecting the difficulty of matching user intent with system resources and a second index reflecting the dispersion of user interests based on the candidate constraint graph to obtain task complexity coordinates. Adaptive orchestration module: used to dynamically determine the activated target agents and their interaction strategies based on the preset quadrant in which the task complexity coordinates fall, and generate orchestration results; The collaborative execution module is used to activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and to perform existence and semantic matching verification on the generated results using the connected candidate constraint graph; based on the quadrant where the task complexity coordinates are located, it configures differentiated role weights for different positions in the verified recommendation list, and generates the final recommendation list through dynamic weighted aggregation.

[0017] The technical solution provided by this invention may include the following beneficial effects: This invention presents an adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness. It constructs a candidate constraint graph (CCG) containing item attributes and semantic relationships, embedding it throughout the entire lifecycle of generation. By embedding the CCG into the decoding and aggregation ends, a closed-loop anti-illusion system of "generation-verification-feedback" is formed, ensuring that the recommendation results are rigorously verified through the completeness of the real candidate pool, reducing or avoiding the illusion problem of generated results. A structured diversity guarantee mechanism based on slot-level role constraints is adopted, significantly improving the diversity and long-tail recall within the recommendation list without significantly sacrificing accuracy. Through continuous multi-dimensional task complexity awareness and adaptive orchestration modules, task difficulty is mapped to a four-quadrant based on intent-resource matching difficulty. Agents are activated on demand according to task difficulty, the number of debate rounds is dynamically adjusted, and slot allocation strategies and aggregation parameters are finely scheduled, thereby reducing inference costs and response latency as needed. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] Figure 1 A flowchart illustrating an exemplary embodiment of the present disclosure is shown. Figure 2 This diagram illustrates the structure of an adaptive multi-agent collaborative recommendation system based on multi-dimensional task complexity awareness in an exemplary embodiment of this disclosure. Figure 3 This diagram illustrates the workflow of the cognitive perception module calculating a multidimensional complexity vector in an exemplary embodiment of this disclosure. Detailed Implementation

[0020] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0021] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0022] This invention aims to address the technical problems in existing multi-agent recommender systems based on large language models, such as severe homogenization of recommendation results, uneven allocation of inference resources, and frequent model illusions. It proposes an adaptive multi-agent collaborative recommender method based on multi-dimensional task complexity awareness. Please refer to [reference needed]. Figure 1 This method may include: S1-S5, as follows: S1. Structured modeling of user intent and candidate space: Receive user queries and historical data, retrieve relevant candidate items from the item library, and construct a candidate constraint graph containing item entities and their attribute relationships; S2. Calculation of multidimensional task complexity: Based on the candidate constraint graph, calculate the first index reflecting the difficulty of matching user intent with system resources and the second index reflecting the dispersion of user interests to obtain the task complexity coordinates. S3. Dynamic orchestration of agents based on four-quadrant mapping: Based on the preset quadrant in which the task complexity coordinates fall, dynamically determine the target agents to be activated and the interaction strategies between them, and generate the orchestration result. S4. Agent Cooperative Execution and Result Verification: Activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and use the connected candidate constraint graph to verify the existence and semantic matching of the generated results. S5. Slot aggregation based on task features: Based on the quadrant where the task complexity coordinates are located, configure differentiated role weights for different positions in the verified recommendation list, and generate the final recommendation list through dynamic weighted aggregation.

[0023] In this embodiment, a candidate constraint graph (CCG) containing item attributes and semantic relationships is constructed and embedded into the entire lifecycle of the generated product. By embedding the candidate constraint graph into the decoding end (restricted decoding based on Trie) and the aggregation end (Critic credibility), a closed-loop anti-illusion system of "generation-verification-feedback" is formed, ensuring that the recommendation results are strictly verified for completeness by the real candidate pool, reducing or avoiding the illusion problem of generated results. A structured diversity guarantee mechanism based on slot-level role constraints is adopted to significantly improve the diversity and long-tail recall in the recommendation list without significantly sacrificing accuracy. Through a continuous multi-dimensional task complexity perception and adaptive orchestration module, the task difficulty is mapped to a four-quadrant based on intent-resource matching difficulty. The agent is activated on demand according to the task difficulty, the debate rounds are dynamically adjusted, and the slot allocation strategy and aggregation parameters are finely scheduled, thereby reducing inference cost and response latency on demand.

[0024] Using the above embodiments, this invention provides an adaptive multi-agent collaborative recommendation system based on multi-dimensional task complexity awareness. Please refer to [link / reference]. Figure 2 The collaborative recommendation system includes: The cognitive perception module is used to receive user queries and historical data, retrieve relevant candidate items from the item library, construct a candidate constraint graph containing item entities and their attribute relationships based on the candidate items, and calculate a first index reflecting the difficulty of matching user intent with system resources and a second index reflecting the dispersion of user interests based on the candidate constraint graph to obtain task complexity coordinates. Adaptive orchestration module: used to dynamically determine the activated target agents and their interaction strategies based on the preset quadrant in which the task complexity coordinates fall, and generate orchestration results; The collaborative execution module is used to activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and to perform existence and semantic matching verification on the generated results using the connected candidate constraint graph; based on the quadrant where the task complexity coordinates are located, it configures differentiated role weights for different positions in the verified recommendation list, and generates the final recommendation list through dynamic weighted aggregation.

[0025] The specific structure of the system in the above embodiments is described below.

[0026] The detailed structure and functions of each component and its internal units in this collaborative recommendation system are as follows: 1. Cognitive Perception Module Please refer to Figure 3 The cognitive perception module, acting as the system's "perception center," is responsible for structured modeling and complex quantification of user intent and candidate environments. Specifically, the cognitive perception module includes: (1) Candidate Constraint Graph (CCG) building unit: used to connect the item library, extract the item identifier, semantic vector and attribute features, construct a heterogeneous graph containing the relationship between item entities and attributes, and provide the underlying data structure for constrained decoding in subsequent processing.

[0027] (2) Multidimensional Complexity Calculation Unit: Used to construct a two-dimensional task complexity coordinate system. This unit contains two sub-calculators: ① Resource Matching Calculator (X-axis): Calculates the distance between the user intent vector and the nearest neighbor cluster center in the semantic space of candidate items, quantifies the difficulty of resource matching between "what the user wants" and "what the system can provide", and outputs "intent-resource matching difficulty" as the primary indicator. .

[0028] ② Preference Entropy Statistician (Y-axis): Used to statistically analyze the probability distribution of users' historical behavior in the attribute space and calculate Shannon entropy, quantifying the inherent uncertainty and dispersion of user interests, and outputting "preference dispersion" as a secondary indicator. .

[0029] 2. Adaptive Orchestration Module The adaptive orchestration module, acting as the system's "decision-making brain," is responsible for dynamically adjusting the system's multi-agent topology based on task complexity. Specifically, the adaptive orchestration module includes: (1) Four-quadrant mapping unit: preset based on Threshold and The threshold discrimination logic maps the coordinate points output by the cognitive perception module to one of the four quadrants.

[0030] (2) Topology dynamic reconstruction unit: Based on the mapping results, it generates control instructions containing "role activation list" and "interaction strategy configuration", and determines which agents are instantiated in subsequent steps and the debate rounds and intensity between them.

[0031] 3. Collaborative Execution Module The collaborative execution module, acting as the system's "execution limb," consists of several Large Language Model (LLM-based) agents with specific roles. Upon receiving instructions from the adaptive orchestration module, each agent activates as needed and collaborates according to the strategy. The system's predefined agent role library includes: (1) User analysis agent: As an auxiliary agent, it is responsible for deeply analyzing user profiles and historical behavior sequences, generating natural language descriptions of users' explicit and implicit preferences, and providing personalized context for other recommendation agents.

[0032] (2) Popularity Guide: Recommendation Generation Agent. Its Prompt contains a popularity-first generation strategy, which focuses on retrieving candidate sets from global popularity to ensure the robustness of recommendation results and public acceptance.

[0033] (3) Novel Explorer: Recommendation Generator. It is equipped with a high temperature coefficient and a counter-trend strategy, focusing on discovering long-tail items, hidden gems, or resources with potential surprises, and is responsible for improving the surprise factor of the recommendation system.

[0034] (4) Personalized Matcher: The recommendation generation agent. It strictly aligns with the output of the user analysis agent, focusing on finding items with the highest semantic similarity to the user's historical interests, ensuring the accuracy and personalization of the recommendation results.

[0035] (5) Critical Evaluator: An evaluation and verification agent. It is responsible for reviewing the outputs of the three generative agents mentioned above during multiple rounds of debate. It connects to the candidate constraint graph (CCG) and has the functions of "illusion recognition" and "constraint verification" to eliminate false items that are not in the candidate pool or recommendations that do not meet the hard constraints.

[0036] The adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness of this application will be further illustrated below through specific embodiments.

[0037] S1, the structured modeling of user intent and candidate space, is performed by the cognitive perception module. First, it acquires the user's natural language query and user profile data. Simultaneously, the candidate constraint graph (CCG) construction unit retrieves a set of candidate items related to the query from the item database, extracts the item identifiers, semantic vectors, and attribute features, and constructs a candidate constraint graph (CCG) containing item entity nodes and attribute relationship edges. This candidate constraint graph will serve as the geometric basis for subsequent complexity calculations and the underlying constraint data for restricted decoding.

[0038] The core of this step lies in mapping unstructured natural language and discrete item data into a unified vector space and graph structure. The specific process includes the mathematical modeling of the following three sub-steps: (1) Vectorized representation of user intent and profile The system first receives the user's natural language query. and user profile collection (Includes historical interaction sequences) With explicit preference labels It is mapped to using a pre-trained lightweight semantic representation encoder. The query vector obtained from the 3D semantic space. as follows:

[0039] In a preferred embodiment of the invention, the encoder employs a lightweight converter model based on the Sentence-BERT (SBERT) architecture (e.g., all-minilm-l6-v2 or a similar BERT variant), with the output dimension set as... This type of model, after comparative learning and fine-tuning, can map semantically similar queries and item descriptions to vector spaces with similar geometric distances, thereby ensuring the accuracy of subsequent distance calculations.

[0040] Meanwhile, user intent vector Defined as query vector Image vector Fusion:

[0041] in, To adjust the hyperparameters that weigh the current query intent against historical preferences, This vector will serve as the reference point for calculating the X-axis (intent-resource matching difficulty) in step S2.

[0042] (2) Preliminary retrieval of the candidate item set To construct the candidate space, the system first uses semantic similarity from the full item database. The top-K candidate items are retrieved to form the initial candidate set. :

[0043] in, For items semantic embedding vector, The cosine similarity function is used. This is the preset recall threshold.

[0044] (3) Construction of candidate constraint graph (CCG) To support the anti-hallucination restricted decoding in step S4, the system is based on Construct the heterogeneous graph GCCG. This graph is defined by the following triples:

[0045] in, It is a set of nodes, including two types of nodes: item entity nodes. and attribute feature nodes (Such as the specific director's name, genre tags, price range, etc.).

[0046]

[0047] Let be the set of edges. This is a relational type. Define a mapping function. Return Item The set of attributes. For any item node and attribute nodes ,like Then there exists an edge whose semantic relation is :

[0048] (4) The underlying data foundation of restricted decoding (Trie construction mapping) based on The system further defines the set of legal prefixes. For any attribute (e.g., "science fiction movies"), the corresponding set of valid candidate item IDs is :

[0049] This set It is transformed into a trie structure. Used to calculate the Token mask in step S4:

[0050] S2, the physical space calculation of the multidimensional task complexity vector, is performed by the multidimensional complexity calculation unit in the cognitive perception module. The system calls the candidate constraint graph (CCG) constructed in step S1 as the underlying data. Based on the geometric features of the vector space in the graph, a two-dimensional task complexity coordinate system consisting of two orthogonal dimensions is constructed, and the coordinates of the current task are calculated. : (1) X-axis index (Intent-Resource Matching Difficulty): Metric The primary metric is designed to quantify a user's current search intent. The set of candidate items currently available to the system contained in CCG Distance in semantic space. The resource matching calculator traverses the item entity nodes in CCG and extracts their pre-stored semantic vectors. It calculates the Euclidean distance between the user intent vector and the nearest neighbor cluster centers of these vectors in space, which is used to quantify the difference in resource matching between "what the user wants" and "what the system can provide." Specifically: The system will first extract a set of candidate items from CCG. Clustering There are several semantic clusters, and the centroid vector of each cluster is calculated. X-axis values Defined as the Euclidean distance from the user intent vector to the centroid of the nearest neighbor cluster:

[0051] in, Let be the center vector of the semantic cluster of the k-th candidate item, calculated as the mean of the vectors of all CCG item nodes within that cluster:

[0052] : denotes the Euclidean norm (L2-Norm).

[0053] Physical meaning: The larger the value, the more it indicates that the user's intention deviates from the semantic distribution of mainstream items covered by CCG, that is, "the existing resources of the system are difficult to directly meet the user's needs", and the task belongs to the resource-scarce type.

[0054] (2) Y-axis index (Preference Entropy): Indicator As a second indicator, it aims to quantify the inherent uncertainty and divergence of user interests. The preference entropy statistician calculates the distribution probability of user historical interaction behaviors in the attribute graph space defined by the CCG and calculates its Shannon entropy. Specifically, the system maps user historical interaction items to attribute nodes in the CCG and aggregates attribute frequencies using the "item-attribute" association edges in the CCG. (Y-axis values ​​are also shown.) The calculation is as follows:

[0055] in, M The total number of unique attribute tags involved in the user's historical behavior and existing in the CCG attribute node set; :property The probability (frequency) of occurrence in a user's historical behavior, i.e. ,in For attributes retrieved via CCG edge relations Number of times it appears This represents the total number of historical interactions.

[0056] Physical meaning: The higher the value, the more dispersed the user's wandering path in the CCG attribute network, the more uniform the distribution of interests (increased entropy), the more ambiguous the intent, and the more diverse the recommendation results required.

[0057] S3, based on two-dimensional four-quadrant mapping, is an intelligent agent orchestration module executed by an adaptive orchestration module.

[0058] The system is preset based on Threshold and A two-dimensional four-quadrant problem perception model with threshold division. The four-quadrant mapping unit calculates the threshold division in step S2. Map the coordinates to the corresponding quadrants to determine the list of characters to be activated and the debate interaction strategy: (1) First Quadrant Pattern (High) &high This is an adversarial correction mode, as detailed below: Scenario characteristics: Users have discerning and diverse tastes, while the candidate pool is extremely scarce (the most difficult scenario).

[0059] Activation strategy: Force activation of all three types of recommendation-generating agents (popular, novel, personalized) and critical evaluators.

[0060] Interaction strategy: Initiating a high-intensity, complementary debate ( The key is to explore the long tail through novel explorers while providing critical evaluators with access to CCG for rigorous fact and constraint verification, so as to prevent the illusion of a large model due to resource scarcity.

[0061] (2) Second Quadrant Pattern (High) &Low This refers to the long-tail refined search mode, as detailed below: Scenario characteristics: The user's intent is clear but it is directed towards unpopular / scarce resources.

[0062] Activation strategy: Prioritize activation of "novel explorers" and critical evaluators, and suppress "popular trendsetters".

[0063] Interaction strategy: Configure a high-priority "constraint verification" function to focus on deep semantic mining within limited resources and ensure the existence of recommendation results.

[0064] (3) Third Quadrant Pattern (low) &Low This is the extremely simple and straightforward mode, as detailed below: Scenario characteristics: The user's intent is clear and resources are sufficient (simple scenario).

[0065] Activation strategy: Activate only the single character that is expected to perform best in the current situation (usually "personalized matchmaker" or "popular guide").

[0066] Interaction strategy: Skip all debates and complex verification processes, directly use a single agent and simple verification output results to achieve a fast response with minimal computing power.

[0067] (4) Fourth Quadrant Pattern (low) &high This represents a diverse divergent pattern, specifically as follows: Scenario characteristics: Abundant resources but broad user interests / vague intent.

[0068] Activation strategy: Activate all three types of recommendation-generating agents.

[0069] Interaction strategy: Initiate complementary debates of medium to low intensity. 1) The key is to supplement the recommendations with the perspectives of different roles to maximize the coverage and diversity of the recommendations.

[0070] S4: Cooperative execution role activation and recommendation result generation. The cooperative execution module responds to the instructions in step S3, instantiates the agent as needed, and executes the following process: (1) S4-1: Deep analysis of user profile (prerequisite dependency). If the current strategy requires the intervention of personalized features, the user analysis agent is activated first. This agent deeply analyzes the user's historical behavior sequence, generates natural language descriptions of the user's explicit and implicit preferences, and injects them as contextual cues into the input of the subsequent recommendation agent.

[0071] (2) S4-2: Context constraint generation based on role-based prompting engineering. The activated recommendation generation agent (popular guide / novel explorer / personalized matcher) constructs input prompts based on the structured prompt template of its corresponding role r. Where r is the role identifier of the agent that generates the recommendation. Structured prompts are generated to recommend agents for input to this role. The prompts... The following constraint information is explicitly included: Role definition: Define the behavioral guidelines for the current intelligent agent (e.g., "find niche but highly rated movies" or "match user's historical interests"). User profile (optional, only provided to personalized matchers): Natural language description of user preferences from step S4-1; Candidate Pool Context: The system converts the set of legitimate items retrieved from the candidate constraint graph (CCG) into structured text (such as a JSON list or an Item List with attributes) and embeds it into the context window of the prompt words as a reference retrieval source for the generated search.

[0072] During the generation process, the agent performs standard autoregressive generation (restricted decoding). The model is based on... The system selects and generates a preliminary set of candidate items from the candidate pool context, based on the instructions in the code. At this point, the system mainly relies on the instruction compliance ability and context learning ability of the large language model to ensure that the output is within the candidate pool as much as possible. The final preliminary result will be directly transmitted to subsequent steps for verification by critical evaluators.

[0073] (3) S4-3: Conditional adversarial questioning and illusion correction. Based on the strategy determined in step S3, the critical evaluator performs multidimensional verification of the generated results. Only when the overall credibility score is... The result is only retained if the target is met; otherwise, controlled regeneration is triggered (at most). Second-rate).

[0074] The verification process involves calculations of the following two core dimensions: ① Existence check: After standardizing the generated items, map them to CCG and output a Boolean value. .

[0075] ② Semantic matching verification: Calculate the generated item vector With user intent vector The cosine similarity is used to determine whether the semantic threshold is met. The boolean result of the semantic matching verification is as follows:

[0076] in The parentheses represent Iverson. The value is 1 if the condition inside the parentheses is true, and 0 otherwise.

[0077] ③ Overall credibility score:

[0078] In the formula, This is the weighting coefficient. If... If the value is below 0.8, the verification is considered to have failed, and the failure_reason is fed back to the generator for correction.

[0079] S5, slot-by-slot aggregation based on quadrant features: The collaborative execution module ultimately generates the final list based on the task difficulty quadrant determined by the adaptive orchestration module. Unlike traditional global sorting, this invention introduces "slots" as independent decision-making units and employs a strategy of "first calculating the global quality score, then dynamically weighting by slot" to generate the final list. The specific process is as follows: S5-1: Slot Definition and Initialization. The system first defines the final output recommendation list. Given a sequence containing K ordered positions (e.g., K = 10 when outputting 10 recommendations). Definition: The first s Each slot ( () represents different levels of user attention allocation, corresponding to the [number]th [item] in the future recommendation list. s Items. s = 1 is the primary anchor point and has the greatest impact on click-through rate; as... s As user attention diminishes, the function of slots is gradually shifting from "precise targeting" to "diversified supplementation."

[0080] Initialization: Construct an empty list of length K and a set of items to be selected. (Including all candidate items output by the agents in step S4 that have passed the critical evaluator's verification), the recommended items are defined as candidate instances in the form of tuples. = , indicating "item" By the character (If the same item is recommended by different recommendation agents at the same time, it will be saved as a different instance.)

[0081] S5-2: Calculation of the global base score for candidate items. The system... Each instance in = Perform the following calculations: (1) Role-specific normalization. Since the original confidence distributions of different agents differ, normalization needs to be performed for each role. Internal candidate items The original scores are normalized. For example... Its normalized score depends on the item's role. Relative rank or value in the original output list:

[0082] In the formula, For the role The set of all candidate scores output. To prevent tiny quantities with a denominator of zero.

[0083] (2) Then calculate the candidate items The fusion quality score is calculated using the following formula:

[0084] In the formula, It is the overall credibility score calculated in S4-3.

[0085] S5-3: Construct a "Slot-Role" weight matrix. To achieve the correspondence between slots and tasks, the system generates a weight matrix based on the task difficulty quadrant determined in step S3. weight matrix ( (Number of characters). Matrix elements Defined the first Each slot is for characters Acceptance level (multiplier) of items: Scenario A: Quadrant 1 / 2 (Resource Scarcity / Hard Mode) Strategy: Focus on strong upward movement at the top, and provide a safety net at the bottom.

[0086] Matrix configuration: Preceding slots ( ) assigns high multipliers to "novel explorers" (such as This forces the results output by agents with mining capabilities to be prioritized; subsequent slots gradually return to a balanced state.

[0087] Scenario B: Quadrant 4 (Ambiguous / Divergent Intent) Strategy: Alternate in waves to maximize diversity.

[0088] Matrix configuration:

[0089]

[0090]

[0091] This process continues, resulting in a Round-Robin weight distribution.

[0092] Scenario C: Third Quadrant (Simple Straight-Through Mode) Strategy: Efficiency first.

[0093] Matrix configuration: Generate an all-1 matrix or slightly weight the "personalized matchers", degenerating into a matrix based on... Global sorting.

[0094] S5-4: Slot-by-slot dynamic filling loop. The system enters a greedy iterative process, starting from... Iterate up to K, finding the optimal solution for each slot: (1) Calculate the dynamic priority of the current slot: for the set of candidate items Each remaining item Calculate its final priority for the current slot:

[0095] (2) Optimal filling and state update: First select The highest item As the current slot The filler item. (The rest of the text appears to be a list of keywords or tags and doesn't translate directly.) Fill in the list Then from the set of candidate items Remove from .

[0096] (3) Fallback Mechanism: If, under the current slot, the filtering of the weight matrix causes all candidates to fail... If all values ​​are below the preset safety threshold, the system will temporarily ignore the weight matrix. W Select directly middle Fill the list with the tallest items to ensure there are no empty spaces.

[0097] S5-5: Output the result. When the loop ends ( ) or set When exhausted, the system outputs the final recommendation list. It also includes a source role tag for each item, completing a full adaptive recommendation process.

[0098] In summary, the beneficial effects of this application are as follows: 1. Achieved on-demand allocation of computing resources and high-efficiency agent scheduling: This invention breaks through the static and rigid collaborative topology of existing systems. By constructing a two-dimensional task complexity coordinate system composed of "intent-resource matching difficulty" and "preference entropy," it achieves refined perception of the complexity of user queries. Based on a four-quadrant mapping mechanism, the system can adaptively switch between multiple strategies such as "extremely simple and easy mode" and "adversarial correction mode" according to the actual difficulty of the task. This dynamic orchestration mechanism avoids excessive consumption of computing power for simple tasks while ensuring sufficient inference depth for complex tasks. Experiments show that this method effectively reduces unnecessary agent activation and invalid interaction rounds while maintaining high recommendation quality, significantly reducing the overall inference cost and response latency of the system, and achieving optimal allocation of computing resources.

[0099] 2. A systematic balance between recommendation accuracy and diversity is achieved: Addressing the issue of homogenized recommendation results caused by "probability maximization" in large language models, this invention designs a differentiated role matrix (popularity guide, novelty explorer, preference matcher) and a "slot-by-slot aggregation based on quadrant features" strategy. Unlike traditional global weighted ranking, this invention introduces "slots" as independent decision-making units and achieves fine-grained scheduling of different recommendation roles through a dynamic weight matrix. For example, in scenarios with scarce resources or ambiguous intent, the system can forcibly increase the weight of "novelty explorers" in specific slots. This mechanism intrinsically ensures the diversity of recommendation list distribution and long-tail recall from the feature space without significantly sacrificing recommendation accuracy, effectively balancing users' needs for precise matching and the desire for surprising exploration.

[0100] 3. This invention suppresses generative illusions at the algorithmic level, ensuring the factual accuracy and constraint alignment of recommendation results: Addressing the pain point of large models easily generating factual illusions in closed-domain recommendations, this invention constructs a candidate constraint graph (CCG) containing item attributes and semantic relationships, deeply embedding it into the entire generation process. On one hand, in the task perception stage, the CCG provides a geometric benchmark for task complexity calculation; on the other hand, in the execution stage, the system utilizes the CCG to implement multi-dimensional verification (existence verification and semantic matching verification) based on a "critical evaluator." This closed-loop mechanism of "generation-verification-feedback" ensures that the output results strictly correspond to real-world candidate items and accurately meet the user's hard attribute constraints, significantly improving the system's credibility and user experience.

[0101] 4. Enhanced deep analysis capability for complex constraints and implicit conflicts: Unlike existing technologies that rely solely on shallow features such as text length to determine task difficulty, the "cognitive perception module" proposed in this invention can perform deep modeling from two dimensions: semantic space distance (resource matching degree) and historical behavior distribution dispersion (preference entropy). This means that the system can not only understand "what the user said," but also predict potential risks in recommendation tasks by quantifying "the gap between what the user wants and what the system can provide" and "the uncertainty of the user's interests." This deep perception capability enables the system to plan more accurate agent collaboration paths when processing queries containing implicit complex constraints, significantly improving recommendation performance in complex scenarios.

[0102] It should be noted that although several modules of the system for executing actions are mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into multiple modules for embodiment. Components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0103] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0104] It should be noted that the installation of image acquisition and personal identification equipment in public places involved in this application is necessary for maintaining public safety, complies with relevant national regulations, and is accompanied by prominent warning signs. The collected personal images and identification information can only be used for the purpose of maintaining public safety and not for other purposes; or the images, personal identification data, etc. in this application are all legally and compliantly obtained or collected with the individual's separate consent.

Claims

1. An adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness, characterized in that, include: S1. Structured modeling of user intent and candidate space: Receive user queries and historical data, retrieve relevant candidate items from the item library, and construct a candidate constraint graph containing item entities and their attribute relationships; S2. Calculation of multidimensional task complexity: Based on the candidate constraint graph, calculate the first index reflecting the difficulty of matching user intent with system resources and the second index reflecting the dispersion of user interests to obtain the task complexity coordinates. S3. Dynamic orchestration of agents based on four-quadrant mapping: Based on the preset quadrant in which the task complexity coordinates fall, dynamically determine the target agents to be activated and the interaction strategies between them, and generate the orchestration result. S4. Agent Cooperative Execution and Result Verification: Activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and use the connected candidate constraint graph to verify the existence and semantic matching of the generated results. S5. Slot aggregation based on task features: Based on the quadrant where the task complexity coordinates are located, configure differentiated role weights for different positions in the verified recommendation list, and generate the final recommendation list through dynamic weighted aggregation.

2. The adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness as described in claim 1, characterized in that, In S1, the candidate constraint graph is a heterogeneous graph. The nodes of the candidate constraint graph include item entity nodes and attribute feature nodes. The edges of the candidate constraint graph represent the association between the item and its attributes. The candidate constraint graph is used to construct a data structure to support subsequent restricted decoding and verification.

3. The adaptive multi-agent collaborative recommendation method based on multi-dimensional task complexity awareness as described in claim 1, characterized in that, In S2, the calculation process of the first indicator includes: calculating the distance between the user intent vector and the overall distribution center of the candidate items in the semantic space, and using the distance as the first indicator; the calculation process of the second indicator includes: analyzing the distribution uniformity of the user's historical behavior on the attribute network of the candidate constraint graph, and using the distribution uniformity as the second indicator; and using the first indicator and the second indicator as coordinates of task complexity.

4. The adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness according to claim 1, characterized in that, In S3, the process of dynamically determining the activated target agents and their interaction strategies based on the preset quadrant in which the task complexity coordinates fall includes: When the task complexity coordinate falls into the first quadrant, it indicates that the task matching difficulty is high and the interest dispersion is high. The interaction strategy adopted is to activate all types of intelligent agents and conduct multiple rounds of high-intensity debate and verification. When the task complexity coordinate falls into the second quadrant, it indicates that the task matching difficulty is high but the interest dispersion is low. The interaction strategy adopted is to prioritize activating agents that focus on mining long-tail resources and perform deep constraint verification. When the task complexity coordinate falls into the third quadrant, it indicates that the task matching difficulty is low and the interest dispersion is low. The interaction strategy adopted is to activate only the single expected optimal agent to respond quickly. When the task complexity coordinate falls into the fourth quadrant, it indicates that the task matching difficulty is low but the interest dispersion is high. The interaction strategy adopted is to activate all types of generated agents and conduct medium-to-low intensity debates focusing on diversity.

5. The adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness according to claim 1, characterized in that, In S4, the agents include: an analytical agent for parsing user profiles, a trendsetter focused on popularity, a novelty explorer focused on long-tail resources, a personalization matcher focused on individual historical preferences, and a critical evaluator responsible for connecting candidate constraint graphs to verify the results.

6. The adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness according to claim 1, characterized in that, In S5, the slot aggregation process based on task features includes: Define each sequential position in the validated recommendation list as a slot; Configure a weight coefficient matrix for each slot, tailored to different agents; Iteratively, for each slot, select the item with the highest product of its base score and the corresponding agent weight from the validated recommendation list and fill it.

7. The adaptive multi-agent cooperative recommendation method based on multi-dimensional task complexity awareness according to claim 6, characterized in that, The configuration strategy of the weight coefficient matrix is ​​dynamically adjusted according to the quadrant in which the task complexity coordinate falls. The configuration strategy includes: increasing the weight of the mining agent in the front slot in the scenario of scarce resources; making the items of different agents appear alternately at the front of the list in the scenario of ambiguous user intent; and degenerating to global sorting by basic score in the simple scenario.

8. An adaptive multi-agent collaborative recommendation system based on multi-dimensional task complexity awareness, characterized in that, The system includes: The cognitive perception module is used to receive user queries and historical data, retrieve relevant candidate items from the item library, construct a candidate constraint graph containing item entities and their attribute relationships based on the candidate items, and calculate a first index reflecting the difficulty of matching user intent with system resources and a second index reflecting the dispersion of user interests based on the candidate constraint graph to obtain task complexity coordinates. Adaptive orchestration module: used to dynamically determine the activated target agents and their interaction strategies based on the preset quadrant in which the task complexity coordinates fall, and generate orchestration results; The collaborative execution module is used to activate the corresponding agents to collaboratively generate a recommendation list according to the orchestration results, and to perform existence and semantic matching verification on the generated results using the connected candidate constraint graph; based on the quadrant where the task complexity coordinates are located, it configures differentiated role weights for different positions in the verified recommendation list, and generates the final recommendation list through dynamic weighted aggregation.

Images