A lightweight cross-domain recommendation method and system based on user alignment agent driving

Abstract

The application discloses a kind of lightweight cross-domain recommendation method and system based on user alignment Agent drive.It first obtains the historical behavior data of user in multiple fields, fuses text, image and other multi-modal content, generates fine-grained interest prototype through cross-domain semantic encoder, and constructs personalized Agent to simulate user intent.Then, in the multi-field collaborative environment, the behavior strategy of Agent is optimized using reinforcement learning and hybrid reward mechanism, and the general preference and domain-specific preference are modeled through hierarchical strategy network, and knowledge fusion is realized through gating mechanism.Subsequently, combined with preference distillation technology, transferable representation is extracted from Agent behavior, and lightweight cross-domain knowledge graph is constructed.Finally, by using behavior trajectory compression and cross-domain preference mapping, efficient and low-consumption personalized recommendation is realized.This method effectively alleviates the problems of cross-domain data sparsity and model complexity, improves recommendation accuracy and system response efficiency, and is suitable for real-time recommendation services in multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and in particular to a lightweight cross-domain recommendation method and system based on user alignment agent-driven approach. Background Technology

[0002] Recommender systems, as a core technology for information filtering, are widely used in video streaming, e-commerce, news, and other fields. Traditional methods, such as collaborative filtering, matrix factorization, and neural networks, achieve personalized recommendations by mining user-item interaction data. However, with the development of multi-domain integration, problems such as data sparsity in a single domain, fragmented user behavior, and cold start are becoming increasingly serious. Especially in cross-domain recommendations, user behavior data is distributed across multiple platforms, making it difficult to capture cross-domain signals, leading to a decline in recommendation effectiveness.

[0003] In recent years, transfer learning, shared representations, and knowledge distillation have been used to alleviate the problems of cross-domain data heterogeneity and sparsity. However, existing methods face two major challenges: first, real-time recommendation requires high-frequency data interaction and model updates, resulting in high computational overhead and limited scalability; second, the difference between offline training and online inference environments causes a decline in model performance, making it difficult to balance accuracy and response speed. Furthermore, user behavior in different domains often lacks explicit correlation, increasing the difficulty of modeling.

[0004] Agent technology has advantages in simulating decision-making processes and environmental adaptation; reinforcement learning agents have been applied to single-domain recommendations to simulate long-term user preferences. However, in cross-domain scenarios, agent methods are limited by the sparsity of interaction data and insufficient behavioral modeling. Furthermore, existing agents struggle to align with real user preferences, leading to behavioral simulation biases, reducing the effectiveness of cross-domain knowledge transfer, and limiting their potential application in practical recommendation systems.

[0005] In real-world applications, such as personalized shopping guides that integrate e-commerce and social platforms, users exhibit characteristics like intertwined multiple intentions and preference drift. Implicit connections exist between user behaviors across different platforms, but these are difficult to model due to data silos. Traditional systems, to ensure real-time performance, require frequent cross-domain synchronization, resulting in high computational resource consumption and significant latency. Therefore, there is an urgent need to construct a lightweight, scalable, and high-fidelity user behavior simulation mechanism to achieve efficient cross-domain knowledge transfer and online reasoning, thereby improving the overall performance of recommendation systems. Summary of the Invention

[0006] To address at least one of the aforementioned technical problems, this invention proposes a lightweight cross-domain recommendation method and system based on user alignment agent-driven approach.

[0007] The first aspect of this invention provides a lightweight cross-domain recommendation method based on user alignment agent-driven approach, comprising:

[0008] Acquire user's historical behavior data across multiple domains, and generate a multimodal interest prototype based on the historical behavior data;

[0009] Construct an initial personalized agent based on the aforementioned multimodal interest prototype;

[0010] A lightweight cross-domain knowledge graph is constructed by training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning.

[0011] Lightweight online reasoning is performed based on the aforementioned cross-domain lightweight knowledge graph to generate cross-domain recommendation results.

[0012] In this solution, the step of acquiring user's multi-domain historical behavior data and generating a multimodal interest prototype based on the historical behavior data specifically involves:

[0013] Acquire historical behavioral data of users in multiple independent domains, including e-commerce platforms, social media platforms, and tourism and local life service platforms. The historical behavioral data includes user ID, interaction domain tags, interaction content, interaction content display type, interaction behavior type, timestamp, and contextual information.

[0014] The interaction content and type of each user are sorted according to time sequence to construct a time sequence of user interaction behavior;

[0015] A cross-domain semantic encoder is constructed based on the multimodal Transformer model. The user interaction content of different display types is encoded according to the cross-domain semantic encoder to construct unified representation data of interaction content.

[0016] Based on the time series of user interaction behaviors, the time decay weight of each user interaction content is determined. A user interest vector h is generated based on the unified representation of the interaction content data and the time decay weight, resulting in a multimodal interest prototype, specifically expressed as:

[0017] ,

[0018] in, For users The length of the time series of interactive behaviors; Attention weights; As a time decay weight; For behavior timestamps; Indicates user The One interactive behavior; Indicates user exist Interest vector at any given time.

[0019] In this solution, the step of constructing and initializing a personalized Agent based on the multimodal interest prototype specifically involves:

[0020] The personalized agent is initialized based on the interest prototype, and the user interest vector h is used as the initial state representation of the agent.

[0021] Based on preference alignment loss function Constrain the distribution of the Agent's interactive behavior, and simulate the user's interactive behavior state based on the constraints of the interactive behavior distribution;

[0022] The preference alignment loss function By calculating each user In any interactive behavior α Policy distribution of the agent and the distribution of real user interaction behavior The KL divergence between the agent and real user interaction behavior distributions is determined by summing the results over all users. The specific formula is as follows:

[0023] ,

[0024] Where KL represents the Kullback-Leibler divergence; Indicates the Agent's state for a given user. Selecting interaction behavior α The strategy distribution; This indicates that real users choose interactive behaviors under the same user state. α The probability distribution.

[0025] In this solution, the step of training the Agent's behavioral trajectory based on cross-domain collaborative reinforcement learning to construct a cross-domain lightweight knowledge graph specifically involves:

[0026] User behavior preferences are determined based on a hierarchical policy network. These preferences include general preferences and specific preferences. The behavior preferences are then fused based on a gating mechanism to select the cross-domain action with the highest value.

[0027] A hybrid reward function is constructed to update user behavior preferences in real time, and the Agent behavior trajectory is determined based on the cross-domain actions;

[0028] Based on the Agent's behavioral trajectory, user behavior preferences are distilled and a cross-domain lightweight knowledge graph is constructed.

[0029] In this scheme, the user's behavioral preferences are determined based on a hierarchical policy network. These behavioral preferences include general preferences and specific preferences. A gating mechanism is used to fuse these behavioral preferences, selecting the cross-domain action with the highest value. Specifically:

[0030] A hierarchical policy network is constructed, comprising a bottom-level network and a top-level network. The bottom-level network calculates the user's general preference vector using an activation function and a weight matrix. Based on this general preference vector, the user's general behavioral patterns in different domains are determined. The specific formula is as follows:

[0031] ,

[0032] in, For activation functions; This is the weight matrix; Indicates user At time step The hidden state vector; For bias terms; Indicates user The general preference vector;

[0033] The high-level network calculates a user's specific preference vector using the Softmax function, and determines the user's preference pattern in a specific domain based on this specific preference vector. The specific formula is as follows:

[0034] ,

[0035] The Softmax function is used to convert a vector into a probability distribution. This is the weight matrix; Indicates user At time step The hidden state vector; Indicates user In the field Specific preference vectors;

[0036] Based on a gating mechanism, the general preference vector and the specific preference vector are multiplied element-wise and then summed with weights to select the cross-domain action with the highest value. The specific formula is as follows:

[0037] ,

[0038] in, For the field Action value matrix.

[0039] In this scheme, the construction of a hybrid reward function updates the research value of user behavior preferences in real time, and the determination of the Agent's behavior trajectory based on the cross-domain actions are specifically as follows:

[0040] A hybrid reward function is constructed to analyze the research value of user behavior preferences. The research value of each user's behavior preferences is determined based on the total value of the hybrid function, and users whose research value is lower than the preset value are removed.

[0041] The hybrid reward function is constructed based on the user's immediate feedback, cross-domain collaborative signals, and long-term value. The specific formula is as follows:

[0042] ,

[0043] in, These are the weighting coefficients; This indicates a reward for clicking to provide feedback; Cosine similarity representing cross-domain interests; Rewards for purchasing behavior; Indicates the discount factor; Discount rewards indicating long-term purchasing behavior; This represents the total value of the mixed reward function;

[0044] The Agent's behavior trajectory is constructed based on the user's cross-domain actions at each point in time.

[0045] In this solution, the step of distilling user behavior preferences based on the Agent's behavior trajectory and constructing a cross-domain lightweight knowledge graph specifically involves:

[0046] User behavior trajectory analysis based on graph convolutional networks and items in different fields The hidden state vectors are aggregated to extract the user's implicit preferences. The specific formula is as follows:

[0047] ,

[0048] in, Represents the user's hidden state vector; In different fields Chinese items The set of hidden state vectors; Indicates user The implicit preference vector;

[0049] The user preference features of the Agent's behavior trajectory are optimized based on the implicit preference vector. A cross-domain lightweight knowledge graph is constructed based on the optimized Agent behavior trajectory, and the edge weights of the cross-domain lightweight knowledge graph are updated through a meta-learning adapter.

[0050] In this solution, the step of training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning to construct a cross-domain lightweight knowledge graph also includes:

[0051] Obtain user interaction frequency information in each independent domain, determine user interaction activity in each independent domain based on the interaction frequency information, and determine the amount of user behavior data obtained in each independent domain based on the interaction activity.

[0052] Interaction behavior data from independent domains with user behavior data volume lower than the preset data volume are labeled as sparse behavior data.

[0053] If the behavioral data is sparse, a graph autoencoder is used to complete the missing behavioral data of the user in the corresponding independent domain using a cross-domain adjacency matrix, resulting in an implicit semantic completion strategy. The specific formula is as follows:

[0054] ,

[0055] in, Indicates user The original behavior vector; This represents a cross-domain adjacency matrix, indicating the relationships between different domains. Indicates the user after completion The behavior vector;

[0056] Based on the hidden state of the source domain and the probability distribution of associations between domains, a synthesized action is generated. This synthesized action is then used to supplement the user's behavioral data in the corresponding domain, resulting in a cross-domain association sampling strategy. The specific formula is as follows:

[0057] ,

[0058] in, Represents the hidden state in a given source domain. Under the conditions, target domain actions The probability distribution; This indicates the action of composition.

[0059] In this solution, the step of generating cross-domain recommendation results through lightweight online reasoning based on the cross-domain lightweight knowledge graph specifically involves:

[0060] The system acquires the real-time behavior of online users in multiple independent domains, sorts the real-time behavior according to timestamps, and generates a real-time behavior sequence.

[0061] The real-time behavior sequence is temporally encoded using an LSTM model to generate a hidden state vector for each time step. The specific formula is as follows:

[0062] ,

[0063] in, Indicates time step Hidden state vector; Embed( α t (Time step) behavior The embedding vector; Indicates time step The hidden state vector;

[0064] The hidden state vector of real-time online user behavior is obtained through linear transformation. Projecting this information into the embedded space of the cross-domain lightweight knowledge graph used for training generates implicit preference vectors for online users. The specific formula is as follows:

[0065] ,

[0066] in, This is the weight matrix; For bias terms; Indicates user Implicit preference vectors under real-time behavior;

[0067] By integrating cross-domain collaborative signals and real-time context information, a recommendation score is calculated, and a Top-K recommendation list is generated from highest to lowest score. The specific formula is as follows:

[0068] ,

[0069] in, Represents items eigenvectors; Weighting coefficients are used to adjust the impact of cross-domain click-through rate; Cross-domain click-through rate (CTR) represents the probability of a user clicking on an item in a cross-domain scenario. Indicates user For items Recommended score.

[0070] A second aspect of the present invention also provides a lightweight cross-domain recommendation system based on user alignment agent driving. The system includes a memory and a processor. The memory includes a lightweight cross-domain recommendation method program based on user alignment agent driving. When the lightweight cross-domain recommendation method program based on user alignment agent driving is executed by the processor, it implements the steps of the lightweight cross-domain recommendation method based on user alignment agent driving as described in any of the preceding claims.

[0071] This invention discloses a lightweight cross-domain recommendation method and system based on user-aligned agent-driven architecture. The method first acquires historical user behavior data across multiple domains, integrates multimodal content such as text and images, generates fine-grained interest prototypes through a cross-domain semantic encoder, and constructs a personalized agent to simulate user intent. Next, in a multi-domain collaborative environment, reinforcement learning and a hybrid reward mechanism are used to optimize the agent's behavioral strategy. A hierarchical policy network models general preferences and domain-specific preferences, and a gating mechanism is used to achieve knowledge fusion. Subsequently, preference distillation technology is combined to extract transferable representations from agent behavior, constructing a lightweight cross-domain knowledge graph. Finally, behavioral trajectory compression and cross-domain preference mapping are employed to achieve efficient and low-cost personalized recommendations. This method effectively alleviates the problems of cross-domain data sparsity and model complexity, improves recommendation accuracy and system response efficiency, and is suitable for real-time recommendation services in multiple scenarios. Attached Figure Description

[0072] Figure 1 A flowchart of a lightweight cross-domain recommendation method based on user alignment agent driven by the present invention is shown;

[0073] Figure 2 The flowchart of the cross-domain collaborative training of the present invention is shown;

[0074] Figure 3 The flowchart illustrating the lightweight online inference method for generating cross-domain recommendation results according to the present invention is shown.

[0075] Figure 4 A flowchart of a lightweight cross-domain recommendation method based on user alignment agent driven by the present invention is shown.

[0076] Figure 5 The diagram illustrates a block diagram of a lightweight cross-domain recommendation system based on user alignment agent driven by the present invention. Detailed Implementation

[0077] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0078] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0079] Figure 1 The flowchart of a lightweight cross-domain recommendation method based on user alignment agent driven by the present invention is shown.

[0080] Figure 4 The flowchart of a lightweight cross-domain recommendation method based on user alignment agent driven by the present invention is shown.

[0081] like Figure 1 and Figure 4 As shown, the first aspect of this invention provides a lightweight cross-domain recommendation method based on user alignment agent-driven approach, comprising:

[0082] S102, acquire user's multi-domain historical behavior data, and generate a multimodal interest prototype based on the historical behavior data;

[0083] S104, Construct an initial personalized agent based on the multimodal interest prototype;

[0084] S106, Based on cross-domain collaborative reinforcement learning, the agent's behavioral trajectory is trained to construct a cross-domain lightweight knowledge graph;

[0085] S108, Generate cross-domain recommendation results by performing lightweight online reasoning based on the cross-domain lightweight knowledge graph.

[0086] It's important to note that by acquiring historical user behavior data across multiple domains and integrating multimodal content (such as text, images, and interaction types), fine-grained interest prototypes are generated, effectively mitigating the problems of data sparsity and behavior fragmentation. Subsequently, personalized agents are constructed based on these interest prototypes, and KL divergence is used to constrain their policy distribution to align with real user behavior, thereby enhancing the realism of the behavior simulation. During the training phase, the agent optimizes its behavioral trajectory through reinforcement learning in a multi-domain collaborative environment. A hybrid reward function is constructed by combining click-through rate, cross-domain collaborative signals, and purchase conversion rate to improve policy optimization. Simultaneously, transferable features are extracted from the behavioral trajectory through dynamic preference distillation to construct a cross-domain lightweight knowledge graph, achieving knowledge compression and transfer. Finally, in the online inference phase, behavioral trajectory compression and cross-domain preference projection techniques are used to achieve efficient representation and rapid recommendation of real-time behavior. Combined with a multi-objective ranking strategy, personalized recommendation results that balance diversity, relevance, and consistency are output, significantly improving recommendation performance and system scalability while reducing computational costs.

[0087] According to an embodiment of the present invention, the step of acquiring user multi-domain historical behavior data and generating a multimodal interest prototype based on the historical behavior data specifically includes:

[0088] Acquire historical behavioral data of users in multiple independent domains, including e-commerce platforms, social media platforms, and tourism and local life service platforms. The historical behavioral data includes user ID, interaction domain tags, interaction content, interaction content display type, interaction behavior type, timestamp, and contextual information.

[0089] The interaction content and type of each user are sorted according to time sequence to construct a time sequence of user interaction behavior;

[0090] A cross-domain semantic encoder is constructed based on the multimodal Transformer model. The user interaction content of different display types is encoded according to the cross-domain semantic encoder to construct unified representation data of interaction content.

[0091] Based on the time series of user interaction behaviors, the time decay weight of each user interaction content is determined. A user interest vector h is generated based on the unified representation of the interaction content data and the time decay weight, resulting in a multimodal interest prototype, specifically expressed as:

[0092] ,

[0093] in, For users The length of the time series of interactive behaviors; Attention weights; As a time decay weight; For behavior timestamps; Indicates user The One interactive behavior; Indicates user exist Interest vector at any given time.

[0094] It should be noted that the interactive content includes products, posts, services, etc.; the display types within the interaction include text, video, and images; the interactive behavior types are actions performed on the interactive content, including clicking, saving, purchasing, browsing, rating, etc.; the contextual information includes geographical location, user device type, etc.; the time decay weight is used to measure the degree of influence of user interaction behavior on current interest over time, and is obtained through an exponential decay function; the user interest vector is a fixed-dimensional numerical vector encoded from the user's historical behavior (such as clicking, purchasing, browsing, etc.) to quantify the user's interest preferences at a certain moment. Unified representation refers to converting data of different modalities (such as text, images, videos, etc.) in user interaction content into semantic representations located in the same vector space through a multimodal coding model.

[0095] According to an embodiment of the present invention, the step of constructing an initial personalized agent based on the multimodal interest prototype specifically includes:

[0096] The personalized agent is initialized based on the interest prototype, and the user interest vector h is used as the initial state representation of the agent.

[0097] Based on preference alignment loss function Constrain the distribution of the Agent's interactive behavior, and simulate the user's interactive behavior state based on the constraints of the interactive behavior distribution;

[0098] The preference alignment loss function By calculating each user In any interactive behavior α Policy distribution of the agent and the distribution of real user interaction behavior The KL divergence between the agent and real user interaction behavior distributions is determined by summing the results over all users. The specific formula is as follows:

[0099] ,

[0100] Where KL represents the Kullback-Leibler divergence; Indicates the Agent's state for a given user. Selecting interaction behavior α The strategy distribution; This indicates that real users choose interactive behaviors under the same user state. α The probability distribution.

[0101] It should be noted that initializing a personalized agent based on an interest prototype refers to using the multimodal interest prototype (i.e., the user's cross-domain interest vector representation) generated in previous steps as initial parameters to construct a unique agent for each user; the agent interaction behavior distribution refers to the probability distribution of the personalized agent choosing different recommendation actions (such as clicking, purchasing, and favoriting) given a user state (represented by the interest prototype); the goal of simulating user interaction behavior states for the agent according to the interaction behavior distribution constraints is to minimize the preference alignment loss function, making the agent's behavior closer to real user behavior; the user interaction behavior distribution refers to the statistical regularity of a user's historical behavior towards different content or actions (such as clicking, purchasing, and rating) in a specific context (such as time, scene, and domain), which is represented as a probability distribution; the KL divergence is a method to measure the difference between two probability distributions; the user state is any user interest vector initialized by the interest prototype; the policy distribution is the probability distribution of the agent towards all possible actions (such as recommending items and user operations) given the current user state.

[0102] Figure 2The flowchart of the cross-domain collaborative training of the present invention is shown.

[0103] According to an embodiment of the present invention, the step of training the Agent's behavioral trajectory based on cross-domain collaborative reinforcement learning to construct a cross-domain lightweight knowledge graph specifically includes:

[0104] S202, determine the user's behavioral preferences based on a hierarchical policy network, the behavioral preferences include general preferences and specific preferences, fuse the behavioral preferences based on a gating mechanism, and select the cross-domain action with the highest value;

[0105] S204, Construct a hybrid reward function to update the user's behavioral preferences in real time, and determine the Agent's behavioral trajectory based on the cross-domain actions;

[0106] S206, Distill user behavior preferences based on the Agent behavior trajectory and construct a cross-domain lightweight knowledge graph.

[0107] It should be noted that this approach drives the evolution of Agent cross-domain behavior by constructing a collaborative training framework based on reinforcement learning and meta-learning. Specifically, for the sparse interaction problem in cross-domain scenarios, a sparse behavior enhancement mechanism is proposed. This mechanism generates synthetic behavior data by using implicit semantic completion within the domain and probabilistic association sampling across domains, thus enriching the Agent's training samples. A hierarchical policy network is designed, with the lower-level network capturing general domain preferences (such as price sensitivity) and the higher-level network dynamically adjusting domain-specific preferences (such as e-commerce category preferences and social topic preferences). A gating mechanism is used to achieve cross-domain knowledge fusion. A meta-learning optimizer is introduced to quickly adjust the Agent's policy parameters in multiple cross-domain tasks, adapting to data distribution shifts in new domain environments.

[0108] According to an embodiment of the present invention, the user's behavioral preferences are determined based on a hierarchical policy network. These behavioral preferences include general preferences and specific preferences. The behavioral preferences are then fused based on a gating mechanism to select the cross-domain action with the highest value. Specifically:

[0109] A hierarchical policy network is constructed, comprising a bottom-level network and a top-level network. The bottom-level network calculates the user's general preference vector using an activation function and a weight matrix. Based on this general preference vector, the user's general behavioral patterns in different domains are determined. The specific formula is as follows:

[0110] ,

[0111] in, For activation functions; This is the weight matrix; Indicates user At time step The hidden state vector; For bias terms; Indicates user The general preference vector;

[0112] The high-level network calculates a user's specific preference vector using the Softmax function, and determines the user's preference pattern in a specific domain based on this specific preference vector. The specific formula is as follows:

[0113] ,

[0114] The Softmax function is used to convert a vector into a probability distribution. This is the weight matrix; Indicates user At time step The hidden state vector; Indicates user In the field Specific preference vectors;

[0115] Based on a gating mechanism, the general preference vector and the specific preference vector are multiplied element-wise and then summed with weights to select the cross-domain action with the highest value. The specific formula is as follows:

[0116] ,

[0117] in, For the field Action value matrix.

[0118] It should be noted that the weight matrix is ​​a parameter matrix automatically learned through backpropagation during model training, used to map the input hidden state vector to user representations across different preference dimensions; the general preferences refer to stable behavioral characteristics commonly exhibited by users in multiple domains, including at least price sensitivity, interaction activity, brand loyalty, content complexity preference, time preference (e.g., daytime or nighttime activity), and device usage habits; the general behavioral pattern is the underlying behavioral rule shared by users in multiple domain scenarios, which is generated by the underlying network of the hierarchical policy network through activation functions and weight matrices to calculate the general preference vector. It represents the stable preference characteristics that users exhibit across different domains, such as price sensitivity or activity within fixed time periods whether users are shopping on e-commerce platforms or browsing content on social platforms; the specific preference refers to the unique behavioral characteristics that users exhibit within a single domain, such as category preferences in e-commerce (e.g., electronics vs. clothing), topic interests on social platforms (e.g., technology vs. entertainment), or scenario preferences for local life services (e.g., dining vs. tourist attractions); the specific domain preference pattern refers to the unique behavioral tendencies that users exhibit in a specific domain (e.g., e-commerce, social media, etc.), such as e-commerce category preferences, social media topic preferences, etc. The algorithm is derived from a multimodal interest prototype; the bias term is used to adjust the threshold of the activation function; the Softmax function ensures that the sum of all elements is 1; the hidden state vector is the user interest vector, and the action is the user's interaction behavior in different domains, i.e., the interaction content and the interaction behavior type of the interaction content; the cross-domain action is the interaction behavior (such as clicking on a product, browsing posts, etc.) selected by the agent in different domains based on the user's current state (such as the interest vector), and the optimal recommended operation is generated by fusing general and specific preferences through a gating mechanism; the action value matrix is ​​a domain-specific value table obtained through reinforcement learning training, which stores the expected long-term reward of each action (such as clicking, purchasing) in a specific state, and is used to guide the agent to select the optimal cross-domain action.

[0119] According to an embodiment of the present invention, the construction of the hybrid reward function to update the research value of user behavior preferences in real time, and the determination of the Agent behavior trajectory based on the cross-domain action, specifically includes:

[0120] A hybrid reward function is constructed to analyze the research value of user behavior preferences. The research value of each user's behavior preferences is determined based on the total value of the hybrid function, and users whose research value is lower than the preset value are removed.

[0121] The hybrid reward function is constructed based on the user's immediate feedback, cross-domain collaborative signals, and long-term value. The specific formula is as follows:

[0122] ,

[0123] in, These are the weighting coefficients; This indicates a reward for clicking to provide feedback; Cosine similarity representing cross-domain interests; Rewards for purchasing behavior; Indicates the discount factor; Discount rewards indicating long-term purchasing behavior; This represents the total value of the mixed reward function;

[0124] The Agent's behavior trajectory is constructed based on the user's cross-domain actions at each point in time.

[0125] It should be noted that by calculating a hybrid reward function (integrating click-through rate, cross-domain collaborative signals, and purchase conversion rate (i.e., long-term value)), more comprehensive and personalized user behavior modeling and decision-making are achieved. The hybrid reward function combines immediate feedback, cross-domain collaborative signals, and long-term value, ensuring the system can quickly respond to users' immediate needs, understand users' cross-domain preferences, and predict and optimize users' long-term value. In the hybrid reward function, immediate feedback (such as click rewards) ensures the short-term accuracy of recommendations, cross-domain collaborative signals (such as interest similarity) enhance knowledge transfer between domains, and long-term value (such as discount purchase rewards) guides the Agent to focus on users' continuous satisfaction. This multi-objective fusion mechanism can dynamically balance the short-term effects and long-term benefits of the recommendation system; the immediate feedback refers to the user's click-through rate for each domain; the discount purchase reward refers to the decaying weight reward assigned to the user's potential future purchase behavior, gradually reducing the impact of future purchase behavior on the current decision through a discount factor γ; the weight coefficients are determined through offline training; the click feedback reward refers to the immediate positive incentive given to the Agent when the user clicks on the recommended content. It measures the similarity between users' interest vectors in the source and target domains. Agent behavior trajectories refer to a series of decision paths formed by a personalized agent continuously selecting cross-domain actions (such as clicking on products or browsing posts) based on user states (such as interest vectors) during interactions with multi-domain environments. These trajectories record time-series data of states, actions, and rewards, which are used to train reinforcement learning models and construct cross-domain knowledge graphs.

[0126] According to an embodiment of the present invention, the step of distilling user behavior preferences based on the Agent behavior trajectory and constructing a cross-domain lightweight knowledge graph specifically includes:

[0127] User behavior trajectory analysis based on graph convolutional networks and items in different fields The hidden state vectors are aggregated to extract the user's implicit preferences. The specific formula is as follows:

[0128] ,

[0129] in, Represents the user's hidden state vector; In different fields Chinese items The set of hidden state vectors; Indicates user The implicit preference vector;

[0130] The user preference features of the Agent's behavior trajectory are optimized based on the implicit preference vector. A cross-domain lightweight knowledge graph is constructed based on the optimized Agent behavior trajectory, and the edge weights of the cross-domain lightweight knowledge graph are updated through a meta-learning adapter.

[0131] It should be noted that the system needs to implement dynamic preference distillation and knowledge graph construction. Specifically, this involves extracting cross-domain transferable representations from agent trajectories and constructing a lightweight knowledge graph. This process includes two parts: implicit preference extraction and graph construction rules. The construction of the cross-domain lightweight knowledge graph specifically involves: abstracting user and item entities in the agent's behavioral trajectory as graph nodes, where user nodes are represented by implicit preference vectors, and item nodes are encoded according to hidden state vectors from different domains; secondly, defining edge relationships by analyzing the cross-domain interaction behavior of users and items, and initializing edge weights using indicators such as behavior frequency and time decay weights; next, using a graph neural network to jointly learn representations of nodes and edges to optimize the semantic consistency of cross-domain associations; finally, dynamically adjusting edge weights through a meta-learning adapter to form a lightweight graph structure that can be incrementally updated, stored as a combination of compressed adjacency lists and low-dimensional embedding vectors. Implicit preference vectors are dense, low-dimensional vectors generated by deep aggregation of multi-domain interaction data in a user's historical behavior trajectory through graph convolutional networks. They quantify and represent a user's potential, unobserved, cross-domain interest preferences (e.g., "tech product enthusiast" may implicitly imply a preference for "following electronic social topics").

[0132] According to an embodiment of the present invention, the step of training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning to construct a cross-domain lightweight knowledge graph further includes:

[0133] Obtain user interaction frequency information in each independent domain, determine user interaction activity in each independent domain based on the interaction frequency information, and determine the amount of user behavior data obtained in each independent domain based on the interaction activity.

[0134] Interaction behavior data from independent domains with user behavior data volume lower than the preset data volume are labeled as sparse behavior data.

[0135] If the behavioral data is sparse, a graph autoencoder is used to complete the missing behavioral data of the user in the corresponding independent domain using a cross-domain adjacency matrix, resulting in an implicit semantic completion strategy. The specific formula is as follows:

[0136] ,

[0137] in, Indicates user The original behavior vector; This represents a cross-domain adjacency matrix, indicating the relationships between different domains. Indicates the user after completion The behavior vector;

[0138] Based on the hidden state of the source domain and the probability distribution of associations between domains, a synthesized action is generated. This synthesized action is then used to supplement the user's behavioral data in the corresponding domain, resulting in a cross-domain association sampling strategy. The specific formula is as follows:

[0139] ,

[0140] in, Represents the hidden state in a given source domain. Under the conditions, target domain actions The probability distribution; This indicates the action of composition.

[0141] It's important to note that in real-world applications, user behavior distribution across different domains is often uneven (e.g., frequent use of e-commerce but infrequent use of travel services), resulting in a significant lack of interaction data in some domains (labeled as sparse behavioral data). This data sparsity hinders the accuracy of agent behavior simulation in low-activity domains, thus affecting the effectiveness of cross-domain knowledge transfer. Traditional methods struggle to effectively fill in these missing behaviors, while directly ignoring sparse domains risks losing potential cross-domain collaborative signals (e.g., a user's purchase in e-commerce may be indirectly influenced by preferences in the travel domain). Therefore, it's necessary to utilize implicit semantic completion and cross-domain association sampling strategies to generate synthetic data based on inter-domain relationships, thereby enhancing the agent's training performance in sparse domains. A cross-domain adjacency matrix is ​​a matrix representing the relationships between different domains, where each element quantifies the probability of user behavior transfer or semantic similarity between two domains (e.g., e-commerce and social media). Missing behavioral data corresponding to a specific domain refers to user behavior records that are not recorded or are severely insufficient in a specific domain (such as tourism services) due to low interaction frequency. These records include specific interactive actions such as clicking, purchasing, and browsing, as well as their associated contextual information (such as timestamps and content types). The source domain is the domain in which sparse behavioral data exists.

[0142] Figure 3 The flowchart of the lightweight online inference method for generating cross-domain recommendation results according to the present invention is shown.

[0143] According to an embodiment of the present invention, the step of generating cross-domain recommendation results through lightweight online reasoning based on the cross-domain lightweight knowledge graph specifically includes:

[0144] S302, acquire the real-time behavior of online users in multiple independent domains, sort the real-time behavior according to timestamps, and generate a real-time behavior sequence;

[0145] S304, Based on the LSTM model, the real-time behavior sequence is temporally encoded to generate a hidden state vector for each time step. The specific formula is as follows:

[0146] ,

[0147] in, Indicates time step Hidden state vector; Embed( α t (Time step) behavior The embedding vector; Indicates time step The hidden state vector;

[0148] S306 uses a linear transformation to hide the state vector of real-time online user behavior. Projecting this onto the embedded space of the cross-domain lightweight knowledge graph used for training generates implicit preference vectors for online users, using the following formula:

[0149] ,

[0150] in, This is the weight matrix; For bias terms; Indicates user Implicit preference vectors under real-time behavior;

[0151] S308 integrates cross-domain collaborative signals and real-time context information to calculate recommendation scores, and generates a Top-K recommendation list from highest to lowest score. The specific formula is as follows:

[0152] ,

[0153] in, Represents items eigenvectors; Weighting coefficients are used to adjust the impact of cross-domain click-through rate; Cross-domain click-through rate (CTR) represents the probability of a user clicking on an item in a cross-domain scenario. Indicates user For items Recommended score.

[0154] It should be noted that the process involves several key steps. First, online user behavior trajectories are compressed, specifically by using an LSTM network to encode the real-time behavior sequences of online users, generating a hidden state vector for each time step. Second, the compressed vectors are projected onto an offline knowledge graph space. This is achieved through a linear transformation, projecting the hidden state vectors of real-time online user behavior onto the offline knowledge graph space to generate implicit preference vectors for online users. Finally, by integrating cross-domain collaborative signals and real-time context information, recommendation scores are calculated, and a final recommendation list is generated through Top-K ranking, balancing diversity, relevance, and cross-domain consistency. In the online service, a behavior trajectory compression algorithm is used to incrementally encode real-time user behavior, generating low-dimensional temporal representation vectors. A cross-domain preference projection module maps the real-time representations to the offline-trained cross-domain knowledge graph embedding space, and combines this with a domain adapter for local fine-tuning, generating a cross-domain recommendation candidate set for the current moment. By constructing a multi-objective ranking learning framework, cross-domain collaborative representations, dynamic context-aware features, and user-item multimodal interaction relationships are jointly modeled. Recommendation scores are calculated based on an adaptive weight allocation strategy, ultimately generating personalized recommendation results that balance diversity gain, semantic relevance, and cross-domain preference consistency. Real-time behavior refers to the latest interactive actions (such as clicking on products, browsing posts, etc.) that a user generates in different areas during the current session. It includes timestamps, content types, and contextual information, and is used for immediate preference updates and recommendation adjustments during the online inference phase.

[0155] Figure 5 The diagram illustrates a block diagram of a lightweight cross-domain recommendation system based on user alignment agent driven by the present invention.

[0156] A second aspect of the present invention also provides a lightweight cross-domain recommendation system based on user alignment agent driving. The system includes a memory 51 and a processor 52. The memory includes a lightweight cross-domain recommendation method program based on user alignment agent driving. When the lightweight cross-domain recommendation method program based on user alignment agent driving is executed by the processor, it implements the steps of the lightweight cross-domain recommendation method based on user alignment agent driving as described in any of the preceding claims.

[0157] This invention discloses a lightweight cross-domain recommendation method and system based on user-aligned agent-driven architecture. The method first acquires historical user behavior data across multiple domains, integrates multimodal content such as text and images, generates fine-grained interest prototypes through a cross-domain semantic encoder, and constructs a personalized agent to simulate user intent. Next, in a multi-domain collaborative environment, reinforcement learning and a hybrid reward mechanism are used to optimize the agent's behavioral strategy. A hierarchical policy network models general preferences and domain-specific preferences, and a gating mechanism is used to achieve knowledge fusion. Subsequently, preference distillation technology is combined to extract transferable representations from agent behavior, constructing a lightweight cross-domain knowledge graph. Finally, behavioral trajectory compression and cross-domain preference mapping are employed to achieve efficient and low-cost personalized recommendations. This method effectively alleviates the problems of cross-domain data sparsity and model complexity, improves recommendation accuracy and system response efficiency, and is suitable for real-time recommendation services in multiple scenarios.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0159] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0161] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0162] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0163] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A lightweight cross-domain recommendation method based on user alignment agent-driven approach, characterized in that, Includes the following steps: Obtain user's historical behavior data across multiple domains, and generate a multimodal interest prototype based on the historical behavior data, specifically as follows: Acquire historical behavioral data of users in multiple independent domains, including e-commerce platforms, social media platforms, and tourism and local life service platforms. The historical behavioral data includes user ID, interaction domain tags, interaction content, interaction content display type, interaction behavior type, timestamp, and contextual information. The interaction content and type of each user are sorted according to time sequence to construct a time sequence of user interaction behavior; A cross-domain semantic encoder is constructed based on the multimodal Transformer model. The user interaction content of different display types is encoded according to the cross-domain semantic encoder to construct unified representation data of interaction content. Based on the time series of user interaction behaviors, the time decay weight of each user interaction content is determined. A user interest vector h is generated based on the unified representation of the interaction content data and the time decay weight, resulting in a multimodal interest prototype, specifically expressed as: ， in, For users The length of the time series of interactive behaviors; Attention weights; As a time decay weight; For behavior timestamps; Indicates user The One interactive behavior; Indicates user exist Interest vector at time step; The initialization of the personalized agent is constructed based on the aforementioned multimodal interest prototype, specifically as follows: The personalized agent is initialized based on the interest prototype, and the user interest vector h is used as the initial state representation of the agent. Based on preference alignment loss function Constrain the distribution of the Agent's interactive behavior, and simulate the user's interactive behavior state based on the constraints of the interactive behavior distribution; The preference alignment loss function By calculating each user In any interactive behavior α Policy distribution of the agent and the distribution of real user interaction behavior The KL divergence between the agent and real user interaction behavior distributions is determined by summing the results over all users. The specific formula is as follows: ， Where KL represents the Kullback-Leibler divergence; Indicates the Agent's state for a given user. Selecting interaction behavior α The strategy distribution; This indicates that real users choose interactive behaviors under the same user state. α The probability distribution; A lightweight cross-domain knowledge graph is constructed by training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning. Lightweight online reasoning is performed based on the aforementioned cross-domain lightweight knowledge graph to generate cross-domain recommendation results.

2. The lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 1, characterized in that, The method of training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning and constructing a cross-domain lightweight knowledge graph is as follows: User behavior preferences are determined based on a hierarchical policy network. These preferences include general preferences and specific preferences. The behavior preferences are then fused based on a gating mechanism to select the cross-domain action with the highest value. A hybrid reward function is constructed to update user behavior preferences in real time, and the Agent behavior trajectory is determined based on the cross-domain actions; Based on the Agent's behavioral trajectory, user behavior preferences are distilled and a cross-domain lightweight knowledge graph is constructed.

3. A lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 2, characterized in that, The hierarchical policy network determines user behavioral preferences, which include general preferences and specific preferences. These preferences are then fused using a gating mechanism to select the cross-domain action with the highest value. Specifically: A hierarchical policy network is constructed, comprising a bottom-level network and a top-level network. The bottom-level network calculates the user's general preference vector using an activation function and a weight matrix. Based on this general preference vector, the user's general behavioral patterns in different domains are determined. The specific formula is as follows: ， in, For activation functions; This is the weight matrix; Indicates user At time step The hidden state vector; For bias terms; Indicates user The general preference vector; The high-level network calculates a user's specific preference vector using the Softmax function, and determines the user's preference pattern in a specific domain based on this specific preference vector. The specific formula is as follows: ， The Softmax function is used to convert a vector into a probability distribution. This is the weight matrix; Indicates user At time step The hidden state vector; Indicates user In the field Specific preference vectors; Based on a gating mechanism, the general preference vector and the specific preference vector are multiplied element-wise and then summed with weights to select the cross-domain action with the highest value. The specific formula is as follows: ， in, For the field Action value matrix.

4. A lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 2, characterized in that, The construction of the hybrid reward function updates the research value of user behavior preferences in real time, and determines the agent's behavior trajectory based on the cross-domain actions, specifically as follows: A hybrid reward function is constructed to analyze the research value of user behavior preferences. The research value of each user's behavior preferences is determined based on the total value of the hybrid function, and users whose research value is lower than the preset value are removed. The hybrid reward function is constructed based on the user's immediate feedback, cross-domain collaborative signals, and long-term value. The specific formula is as follows: ， in, These are the weighting coefficients; This indicates a reward for clicking to provide feedback; Cosine similarity representing cross-domain interests; Rewards for purchasing behavior; Indicates the discount factor; Discount rewards indicating long-term purchasing behavior; This represents the total value of the mixed reward function; The Agent's behavior trajectory is constructed based on the user's cross-domain actions at each point in time.

5. A lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 2, characterized in that, The process of distilling user behavior preferences based on the Agent's behavior trajectory and constructing a cross-domain lightweight knowledge graph specifically involves: User behavior trajectory analysis based on graph convolutional networks and items in different fields The hidden state vectors are aggregated to extract the user's implicit preferences. The specific formula is as follows: ， in, Represents the user's hidden state vector; In different fields Chinese items The set of hidden state vectors; Indicates user The implicit preference vector; The user preference features of the Agent's behavior trajectory are optimized based on the implicit preference vector. A cross-domain lightweight knowledge graph is constructed based on the optimized Agent behavior trajectory, and the edge weights of the cross-domain lightweight knowledge graph are updated through a meta-learning adapter.

6. A lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 1, characterized in that, The method of training the agent's behavioral trajectory based on cross-domain collaborative reinforcement learning to construct a cross-domain lightweight knowledge graph also includes: Obtain user interaction frequency information in each independent domain, determine user interaction activity in each independent domain based on the interaction frequency information, and determine the amount of user behavior data obtained in each independent domain based on the interaction activity. Interaction behavior data from independent domains with user behavior data volume lower than the preset data volume are labeled as sparse behavior data. If the behavioral data is sparse, a graph autoencoder is used to complete the missing behavioral data of the user in the corresponding independent domain using a cross-domain adjacency matrix, resulting in an implicit semantic completion strategy. The specific formula is as follows: ， in, Indicates user The original behavior vector; This represents a cross-domain adjacency matrix, indicating the relationships between different domains. Indicates the user after completion The behavior vector; Based on the hidden state of the source domain and the probability distribution of associations between domains, a synthesized action is generated. This synthesized action is then used to supplement the user's behavioral data in the corresponding domain, resulting in a cross-domain association sampling strategy. The specific formula is as follows: ， in, Represents the hidden state in a given source domain. Under the conditions, target domain actions The probability distribution; This indicates the action of composition.

7. A lightweight cross-domain recommendation method based on user alignment agent-driven approach according to claim 1, characterized in that, The step of generating cross-domain recommendation results through lightweight online reasoning based on the cross-domain lightweight knowledge graph specifically involves: The system acquires the real-time behavior of online users in multiple independent domains, sorts the real-time behavior according to timestamps, and generates a real-time behavior sequence. The real-time behavior sequence is temporally encoded using an LSTM model to generate a hidden state vector for each time step. The specific formula is as follows: ， in, Indicates time step Hidden state vector; Embed( α t (Time step) behavior The embedding vector; Indicates time step The hidden state vector; The hidden state vector of real-time online user behavior is obtained through linear transformation. Projecting this information into the embedded space of the cross-domain lightweight knowledge graph used for training generates implicit preference vectors for online users. The specific formula is as follows: ， in, This is the weight matrix; For bias terms; Indicates user Implicit preference vectors under real-time behavior; By integrating cross-domain collaborative signals and real-time context information, a recommendation score is calculated, and a Top-K recommendation list is generated from highest to lowest score. The specific formula is as follows: ， in, Represents items eigenvectors; Weighting coefficients are used to adjust the impact of cross-domain click-through rate; Cross-domain click-through rate (CTR) represents the probability of a user clicking on an item in a cross-domain scenario. Indicates user For items Recommended score.

8. A lightweight cross-domain recommendation system driven by a user alignment agent, characterized in that, The system includes a storage device and a processor. The storage device includes a lightweight cross-domain recommendation method program driven by a user alignment agent. When the user alignment agent-driven lightweight cross-domain recommendation method program is executed by the processor, it implements the steps of the lightweight cross-domain recommendation method driven by a user alignment agent as described in any one of claims 1 to 7.

Images