A recommendation interaction simulation system and method in an online content platform scenario

Abstract

The application discloses a kind of recommended interaction simulation system and method in online content platform scene;The application simulates the interactive dynamics between user and recommendation system in online content platform with agent as basic unit, which includes two kinds of agents, user agent and recommendation system agent respectively;User agent is used to perform six behaviors of publishing, forwarding, liking, commenting, following and unfollowing;The recommendation system agent generates a candidate list to provide post recommendation and friend recommendation to the user agent;The user agent adopts an online content platform user simulation general framework composed of an activity model, a preference model and a generation model to accurately simulate user activity, user interaction preference and user content generation trend, and unifies and abstracts the complex behaviors of users into a standard process of selecting from the candidate list;The application can meet the needs of virtual user experiment, recommendation system test and other needs in online content platform scene.

Description

Technical Field

[0001] This invention belongs to the field of information recommendation technology, specifically, it relates to a recommendation interaction simulation system and method in the context of online content platforms. Background Technology

[0002] Online content platforms are platforms on the internet where individuals with shared interests or goals interact and exchange information. Faced with massive amounts of articles, discussions, comments, and social interactions, recommendation systems filter and provide users with content that best matches their interests and needs, enabling users to more effectively access and exchange information. Recommendation systems can be broadly categorized according to their technological approach as follows:

[0003] 1) Traditional recommendation methods: In early recommendation systems, the system mainly relied on the user's historical browsing and interaction behavior, such as reading, commenting on or liking articles, and recommended relevant content to the user based on rule-based strategies;

[0004] 2) Content-based recommendation technology: This technology mainly analyzes the content attributes of articles or discussions, such as text, tags or topics, and then matches them with the user's historical interests to recommend relevant content;

[0005] 3) Social-based recommendation technology: With the integration of social networks into online content platforms, users' social relationships and interaction data provide valuable information for recommendation systems. For example, when a user's friends or people they follow share or comment on an article, that article may be recommended to that user.

[0006] 4) Collaborative filtering-based recommendation technology: This method analyzes the browsing and interaction behavior of a large number of users to identify similar users or content, and then provides recommendations to the target user. For example, if two users frequently participate in discussions on a specific topic, they may also be interested in other related topics.

[0007] In recent years, deep learning and neural network technologies have been widely used in recommendation systems for online content platforms. Through complex neural network structures, these technologies can capture deep relationships between users and content, thereby providing more accurate recommendations.

[0008] While recommendation systems on online content platforms have significantly improved users' efficiency in accessing information, they have also brought about a series of social responsibility and ethical issues. Within these communities, recommendation systems enhance individual engagement by providing personalized content and social relationships, thus significantly impacting the evolutionary trajectory of these online content platforms. How to maintain the diversity and healthy development of the community while providing personalized recommendations is a crucial challenge facing current and future online content platform recommendation systems.

[0009] Recommendation simulators are designed to generalize the interaction between recommender systems and human activity over a period of time. They are important tools for studying the impact of recommender systems on the evolutionary dynamics of online communities. Typically, existing simulators contain two core agents: a user agent and a recommender system agent. However, these frameworks have two major drawbacks. First, they often struggle to accurately simulate complex user behavior and remain stable during simulations. Second, they are inefficient in modeling recommender system agents, especially when providing extensive recommendations from millions of pieces of content to tens of thousands of users. These limitations hinder the ability to conduct realistic and widespread simulations, thus diminishing the potential practical insights gained from such frameworks.

[0010] Accurate user agent modeling requires precisely predicting various individual user behaviors while ensuring a stable collective distribution. In existing simulation systems, user agent modeling typically employs one of two approaches: rule-based or data-based algorithms. Rule-based methods formulate rules for user behavior based on specific theories. While these methods can reproduce relevant phenomena well with well-defined rules, they struggle to capture individual differences. Conversely, data-based methods utilize machine learning models to fit historical operational data. Although these methods can accurately simulate various user behaviors, data bias due to long-tailed distributions ultimately leads to distribution collapse and poor simulation results. Secondly, effective recommender system agents are crucial for simulating extensive recommendations involving millions of users and content. However, existing simulators, limited by their computationally intensive architecture, can only accommodate simulations involving hundreds of users. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the present invention aims to provide a recommendation interaction simulation system and method for online content platforms, supporting the simulation of dynamic interactions between users and recommendation systems on online content platforms such as Weibo. The method of this invention includes: designing a recommendation interaction simulation framework for online content platforms, defining the process and interface for simulating the dynamic interactions between users and recommendation systems; proposing a general framework for user simulation on online content platforms, supporting accurate simulation of user activity, user interaction preferences, and user content generation trends; and designing a general framework for recommendation system simulation on online content platforms, achieving high scalability support for various recommendation strategies. This invention can meet the needs of virtual user experiments and recommendation system testing in online content platform scenarios.

[0012] The technical solution of the present invention is described in detail below.

[0013] This invention provides a recommendation interaction simulation system for online content platforms. It uses intelligent agents as the basic unit to simulate the dynamic interaction between users and the recommendation system on the online content platform. The system includes two types of intelligent agents: a user intelligent agent and a recommendation system intelligent agent. The user intelligent agent performs actions including but not limited to six types: posting, forwarding, liking, commenting, following, and unfollowing. The recommendation system intelligent agent generates a candidate list to provide post recommendations and friend recommendations to the user intelligent agent. Wherein:

[0014] The user agent employs a general framework for simulating online content platform users, consisting of an activity model, a preference model, and a generation model. This framework accurately simulates user activity, user interaction preferences, and user content generation trends, unifying complex user behaviors into a standardized process of selection from a candidate list. The activity model estimates user behavior intentions and predicts behavior density based on historical interaction behavior, social relationships, and interest encoding information. The preference model predicts user preference scores for candidate posts or users based on historical interaction behavior, social relationships, and interest encoding information. The generation model fits the global distribution of posts using a Gaussian mixture model, and then, based on historical interaction behavior and the Gaussian mixture model encoding of candidate posts, predicts user Gaussian mixture model calls and generates content. During operation, candidate content is first ranked based on the preference model. The user agent then interacts with the highest-ranked candidate content based on the behavior density predicted by the activity model. If the interaction involves content generation, the generation model is further invoked.

[0015] In this invention, encoded user interest vectors and post vectors obtained by encoding user real interests and post content based on a discriminative large language model are used as inputs to the model in the general framework for user simulation of online content platforms.

[0016] In this invention, the activity model first estimates user behavior intentions based on UGT calculation rules at the sub-factor granularity, including posting, forwarding, liking, commenting, and following. User intentions include: seeking information, sharing information, seeking self-status, and social interaction. When the factor is seeking information, sub-factors include seeking interesting, useful, or new information; when the user intention is sharing information, sub-factors include self-expression, providing information to others, and contributing to the community; when the user intention is seeking self-status, sub-factors include making a strong impression on others, feeling important, and appearing cool and fashionable; when the user intention is social interaction, sub-factors include staying connected, meeting interesting people, value sharing, and a sense of community belonging.

[0017] Then, a data-driven machine learning approach is used to construct a predictor that learns a mapping function between behavioral intentions and user behavior density to predict behavior density. The predictor is structured as a two-stage composite model, which includes a classification model that predicts the probability of a user engaging in each type of behavior. And a regression model for estimating the density of type a behavior.

[0018] In this invention, the Poisson distribution P(λ) is also used to approximate the prediction of behavior caused by external intentions. The user's final behavior is defined as the sum of the intention-driven behavior predicted by the predictor and the approximate external behavior.

[0019] In this invention, the preference model first concatenates the user's long-term and short-term interest vector encodings with the latent vectors of posts in the user's interaction history. This set of information is then processed by a behavior-sharing user encoder to synthesize a comprehensive user representation vector as the final user encoding vector. Next, each comprehensive user representation vector is paired with the encoding vector of each candidate item (post or user). A rating module targeting specific behaviors is used to predict the user's rating for these candidate items. The preference model is trained using ranking loss, with random negative sampling incorporated during the training process.

[0020] In this invention, when a user presents historical behaviors of different scales, firstly, different types of user behaviors are sampled multiple times with replacement to form a standardized user encoder input. The standardized input obtained by sampling is aggregated through behavior-specific self-attention modules and additional attention modules. Then, the user's long-term interest vector, short-term interest vector, and aggregated interaction representation vector of each behavior are concatenated and used as the input of the user-level self-attention module and additional attention module to perform comprehensive user representation and synthesize a comprehensive user representation vector.

[0021] In this invention, within the general framework for simulating users on an online content platform, a preference model for attention behavior is used to calculate the utility a user derives from their attention relationships. A personalized unfollow threshold is set for each user, which is set to the 0.2 quantile of the user's existing attention relationship utility. When the perceived utility is lower than the unfollow threshold, the user performs the unfollowing action.

[0022] In this invention, a rule-based heuristic algorithm is used in the generation model to manage the activation of the Gaussian mixture model; that is, for posting behavior, user activation signals are collected from long-term and short-term interests, and content is generated by activating the sub-model with the strongest signal; for forwarding and commenting behavior, the signals of the Gaussian mixture model of candidate posts are also integrated to ensure that the generated content remains relevant and realistic.

[0023] In this invention, the recommender agent adopts a general framework simulated by an online content platform recommender system to provide post and friend recommendations to the user agent; the recommendation process of the general framework simulated by the online content platform recommender system is divided into three stages: recall, filtering, and ranking; wherein:

[0024] During the recall phase, each user agent filters out posts that meet specific recall criteria from all posts.

[0025] During the filtering phase, posts that do not meet the basic requirements are removed.

[0026] In the sorting phase, posts that best match the user's interests are selected from those that have passed the recall and filtering phases to form a recommendation list.

[0027] In the general framework of online content platform recommendation system simulation, the recommendation algorithms in the recall and ranking stages include four basic strategies: random recommendation, popularity-based recommendation, social recommendation, and collaborative filtering-based recommendation.

[0028] Random recommendations are based on random sampling for recall and ranking;

[0029] Popularity-based recommendations calculate the popularity of posts and users based on interaction records and social relationships, and then make recommendations accordingly.

[0030] Social-based recommendations utilize graph sampling and graph algorithms for recall and ranking.

[0031] The collaborative filtering-based recommendation uses representation learning algorithms and similarity vector retrieval for recall, and context-aware deep learning algorithms for ranking.

[0032] This invention also provides a method for simulating recommendation interactions in an online content platform scenario. Based on a constructed user agent and a recommendation system agent, it simulates recommendation interactions on an online content platform in multiple rounds. In each round of simulation, the recommendation system agent is first activated to generate a list of recommended posts and a list of recommended friends. Then, the user agent is activated to provide feedback on the recommendation lists, including interactions such as forwarding, liking, commenting, and following. Furthermore, it performs actions such as posting and unfollowing (cleaning the following list) to update its personal intention status. Specifically, the user agent first predicts the user's behavior density based on an activity model, using the user's behavioral intention as input, and simultaneously sorts the candidate list based on a preference model. Then, it generates corresponding interactive behaviors for the top candidates in the sorted candidate list according to the behavior density. If the interactive behavior requires content generation or the posting behavior is triggered, the user agent further calls the generation model to complete the content generation. After all user agents have completed the simulation of all types of behaviors in the current round, the user agents update the user's behavioral intentions based on user interests, user behavior, and social relationships, preparing for the next round of simulation.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention can accurately simulate complex user behaviors and maintain stability during the simulation process. First, it abstracts the complex and diverse user behaviors into a standardized process of selection from a candidate list, and then further breaks it down into three standardized sub-tasks: activity prediction, preference prediction, and generation prediction. This breakdown reduces the technical complexity of user behavior simulation and helps design a stable and accurate user behavior control model. Second, this invention effectively combines rule-based and data-based user control models, solving the shortcomings of inaccurate scale-based user control models and the instability of data-based user control models. Simultaneously, this invention proposes a general framework for simulating online content platform recommendation systems, based on a three-stage process of recall, filtering, and ranking, and four recommendation strategies: random recommendation, popularity-based recommendation, social recommendation, and collaborative filtering. This framework provides post and friend recommendations to the user agent while maintaining a balance between recommendation accuracy and computational efficiency. This invention can meet the needs of virtual user experiments and recommendation system testing in online content platform scenarios. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the recommendation interaction simulation framework in the online content platform scenario of this invention.

[0036] Figure 2 This is a diagram illustrating the preference model architecture within the general framework for simulating users on the online content platform of this invention.

[0037] Figure 3 This is a schematic diagram illustrating the general framework of the online content platform recommendation system of this invention.

[0038] Figure 4 This is a visualization of the distribution stability test results of the activity model of this invention.

[0039] Figure 5 This is a visualization of the distribution stability test results of the model generated in this invention. Detailed Implementation

[0040] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] I. Recommendation Interaction Simulation Framework in Online Content Platform Scenarios

[0042] like Figure 1 As shown, this invention proposes a recommendation interaction simulation framework for online content platform scenarios. This framework consists of two main parts: a user agent and a recommendation system agent. The specific steps are as follows:

[0043] (1) Data sampling is performed on the target online content platform based on Algorithm 1 (data sampling algorithm in online content platform scenario) to obtain dataset D;

[0044] (2) Encode the content and users in dataset D based on a discriminative large language model;

[0045] (3) A general framework for simulating users on online content platforms is proposed. An activity model, a preference model, and a generation model for simulating user behavior on online content platforms are designed. The model is trained and the user agent is initialized based on the dataset D.

[0046] (4) A general framework for simulating online content platform recommendation systems is proposed. Simulated online content platform recommendation systems based on random sampling, popularity, social relationship, and collaborative filtering are designed respectively, and intelligent agent instances of the recommendation system are deployed.

[0047] (5) Based on the user intelligent agent and recommendation system intelligent agent constructed above, multiple rounds of simulation of online content platform recommendation interaction are carried out. In each round of simulation, the recommendation system intelligent agent is activated first to generate a post recommendation list and a friend recommendation list. Then, the user intelligent agent is activated to provide feedback on the recommendation list such as forwarding, liking, commenting, and following, and further to perform actions such as publishing, cleaning up the following list (unfollowing), and updating personal intention status.

[0048] Algorithm 2 (a framework for simulating recommendation interactions in online content platform scenarios) systematically describes the above framework and achieves efficient, accurate, and stable simulation of recommendation interactions on online content platforms.

[0049]

[0050]

[0051]

[0052] II. A General Framework for User Simulation on Online Content Platforms

[0053] like Figure 1 As shown, this invention proposes a general framework for user simulation on online content platforms. The complex and diverse user behaviors on online content platforms place high demands on the accuracy and stability of recommendation interaction simulation. To meticulously simulate this complexity, we abstract various complex user behaviors into a standardized process of selection from a candidate list. Therefore, we propose a general framework for user simulation on online content platforms, consisting of an activity model, a preference model, and a generation model. This framework enables unified simulation modeling of various user behaviors, including but not limited to posting, forwarding, liking, commenting, and following, based on a unified structure.

[0054] 1. Activity Model: This model models users' tendencies toward various behavior types to predict the probability density of user behavior;

[0055] 2. Preference Model: This model models the user's true interests and is used to score and rank each item in the candidate list;

[0056] 3. Generative Model: This model models the content creation behavior of users in the online content platform environment (excluding behaviors that do not generate new content, such as liking or following), and is used to generate vector encodings for new content, such as posting, commenting or forwarding.

[0057] This framework achieves standardized decomposition of complex user behaviors, transforming difficult-to-model complex user behaviors into goal-oriented computational tasks, thus reducing the modeling difficulty. The core of these three sub-models lies in computing encoded vectors in the latent interest space. These encoded vectors deeply capture users' interests, preferences, and behavioral characteristics. By integrating rule-based heuristics and data-driven machine learning strategies, these three sub-models ensure both individual-level accuracy and maintain the overall stability of the simulation.

[0058] like Figure 1As shown, we first initialize user behavior intentions based on historical data and train the activity model, preference model, and generation model to complete the initialization of the user agent. Then, in the simulation environment, when the recommendation system agent sends a candidate list to the user agent through post recommendations or friend recommendations, the user agent first predicts the user's behavior density based on the activity model, using the user behavior intention as input; simultaneously, it sorts the candidate list based on the preference model; then, based on the behavior density, it generates corresponding interactive behaviors for the top candidates in the sorted candidate list; if the interactive behavior requires content generation or the posting behavior is triggered, the user agent further calls the generation model to complete the content generation; after all user agents have completed the simulation of all types of behaviors in the current round, the user agent updates the user's behavior intentions based on user interests, user behavior, social relationships, etc., to prepare for the next round of simulation.

[0059] 1.1 Content and User Code

[0060] Simulating abstract, real-world user interests and post content is a fundamental challenge when building simulators. To address this, we leverage the concept of "representation" from machine learning, mapping these elements to low-dimensional vectors in a latent space.

[0061] When trained on large amounts of online text data, LLMs (Large-Scale Language Models) generate a latent space that reflects typical online user behavior, making them effective tools for understanding and simulating user behavior. Specifically, we use SentenceBERT, a variant of discriminative LLM, which transforms each post into a representation vector, denoted as P. User preferences are then computed using the joint representation of all posts the user interacts with or publishes. We model long-term user interest Z by averaging the representation vectors of all posts within a predefined time window. l And by using a weighted average to capture short-term interest Z s The post vectors and encoded user interest vectors will serve as input to the subsequent model. This method accurately and efficiently simulates user behavior by performing vector operations in the latent interest space.

[0062] 1.2 Activity Model

[0063] The activity model aims to predict the density of user behavior by analyzing user behavioral intentions. Using encoded vectors of posts and user preferences, along with social and interaction relationships as input, we construct a set of rules based on the Uses and Gratification Theory (UGT) to model behavioral intentions. Following this rule-based modeling, we employ machine learning models to further refine the behavioral density predictions based on these modeled intentions, ensuring a more accurate representation of user behavior in the simulation. UGT helps estimate user intentions through various motivational factors, including seeking information (e.g., seeking interesting, useful, or new information), sharing information (e.g., self-expression, providing information to others, contributing to the community), seeking self-status (e.g., impressing others, feeling important, appearing cool or trendy), and social interaction (e.g., staying connected, meeting interesting people, sharing value, feeling a sense of community). However, UGT is descriptive, making quantitative calculation and analysis challenging. As shown in Table 1, based on computable metrics such as interest matching, social connections, and interaction relationships, we propose a set of universal sub-factor granularity UGT calculation rules for user intention estimation.

[0064]

[0065]

[0066] Then, we employ a data-driven machine learning approach to construct a predictor, learning a mapping function between behavioral intentions and user behavior density. The predictor is structured as a two-stage composite model: first, a classification model focused on behavior type a. Predict the likelihood of users engaging in each type of behavior (e.g., posting); then, focus on a regression model for behavior type a. Estimate the density of type A behavior, such as the number of posts. Given that user behavioral intentions tend to be relatively consistent over time, we divide the simulation process into different rounds. This design helps us to base our simulation on historical round data R. his Create a training dataset containing the behavioral intentions of user u in the current round r. As a feature, the behavior density of behavior type a in the next round. The model is trained using the following objective function, with labels as input:

[0067]

[0068]

[0069] These two equations describe the optimization objectives of the classification model and the regression model, respectively. Here, a, r, and u represent the behavior type, round ID, and user ID, respectively. Set a function for the upper boundary of the data (set data greater than 1 to 1). θ represents the model parameters, with subscripts and superscripts indicating the corresponding behavior type and model type, and so on. U is the set of all users. a+ Let be the set of all users with behavior 'a'. The term ω is the weight of the regularization term, used to prevent overfitting and maintain the model's generalization ability.

[0070] During the simulation, the predictor operates in two steps: assessing the probability of a behavior based on intent, and then predicting the behavior density of these confirmed behaviors. Furthermore, external intent factors may exist that are not considered in the UGT theory. To improve the stability and universality of behavior density predictions, we use a Poisson distribution. The behavior induced by these external intentions is approximated and predicted. The user's final behavior is defined as the sum of the intention-driven behavior predicted by the machine learning model and the approximate external behavior:

[0071]

[0072] This represents the final behavior density of user u in the next round r+1, behavior type a. By combining rule-based user intention modeling with a data-driven prediction model, the activity model achieves both efficiency and accuracy in predicting user behavior.

[0073] 1.3 Preference Model

[0074] like Figure 2 As shown, the preference model aims to model the interaction trends between the user agent and the candidate items generated by the recommender system agent. For posts, we directly use their latent vectors as their feature encodings. First, posts are encoded as latent vectors. For users, we first concatenate the user's long-term and short-term interest vector encodings with the latent vectors of posts from the user's interaction history, and then pass them through a behavior-shared user encoder (…). These sets of information are processed and synthesized into a comprehensive user representation vector as the final user encoding vector. When users exhibit historical behaviors of varying scales, we first perform multiple samplings with replacement on different types of user behaviors to form a standardized user encoder input. We perform two rounds of sampling from the perspectives of long-term and short-term interaction behaviors. The sampling weight for the former is set to 1, while the sampling weight for the latter is set to the reciprocal of the round difference. The standardized input obtained from the sampling is aggregated through behavior-specific self-attention modules and additional attention modules. This process can be formulated as follows:

[0075]

[0076]

[0077]

[0078] in This is the sampling result for user a, and P represents the latent interest space vector encoding of all posts. and It is the self-attention projection matrix. φ a It is a projection function, implemented in the form of a dot product operation. It is the aggregated interaction representation vector of user u in terms of behavior a. and We calculate intermediate variables. Next, we concatenate the user's long-term interest vector, short-term interest vector, and aggregated interaction representation vector for each behavior, and use this as input to the user-level self-attention module and additional attention module to perform comprehensive user representation.

[0079] We then pair each comprehensive user representation vector with the encoded vector of each candidate item, which can be a post or a user, depending on the context. We utilize a behavior-specific scoring module ( This is used to predict user ratings for these candidate items, as described below:

[0080]

[0081] Where P refers to the latent space vector of all posts. H u This represents the historical behavior of user u. and These are the long-term and short-term interest vectors of user u, respectively. [-] represents the indexing operation. The preference model is trained using ranking loss. To improve the model's stability and accuracy, we incorporate random negative sampling during training. The above process can be described as follows:

[0082]

[0083] Among them, i + and i - These represent positive and negative samples, respectively. Here, θ refers to θ0. UE and sum.

[0084] 1.4 Generative Model

[0085] Generative models aim to generate content for the behavior of creating new content. Typically, on online content platforms, content is presented around prominent topics, with users creating a large amount of content around each focus. Gaussian Mixture Models (GMMs) accurately reflect this distribution, providing effective boundary constraints for content generation, making them suitable for creating content that aligns with the actual topic distribution observed on online content platforms. Therefore, the content generation task can be transformed into a GMM submodel activation prediction task. However, the inherent long-tail data distribution can lead to biases in the activation process modeled by deep learning methods. Specifically, this bias can be exacerbated by creating a recursive loop where the output of the generative model serves as input for future rounds. To prevent this negative impact, we implement a rule-based heuristic algorithm to manage the activation of GMM submodels. Specifically, for posting behavior, we collect user activation signals from long-term and short-term interests and generate content by activating the submodel with the strongest signal. For forwarding and commenting behavior, we also integrate the signals from the GMM submodels of candidate posts to ensure that the generated content remains relevant and realistic. The above process can be formally represented as:

[0086] argmaxω1softmax(sum(G[H u ]))+ω2softmax(sum((rT[u,H u ])·G[H u ]) / +ω3G[i]

[0087] Where G represents the one-hot encoding of all post GMM submodel indices. r is the index of the current round. T[u,i] is the round index when user u interacts with post i. ω1, ω2, and ω3 are the weights of different parts of the expression to balance their contributions.

[0088] 1.5 Heuristic Unfollowing Strategy

[0089] Since user unfollowing data is often unavailable, we cannot model unfollowing behavior using our proposed activity model. Therefore, we construct an alternative heuristic algorithm based on an expectation model. The basic assumption is that users have a psychological threshold for whether to stop following other users. When perceived utility falls below this threshold, users will unfollow; otherwise, they will not. We use a preference model of following behavior to calculate the utility users derive from following relationships. Since each user exhibits a different unfollowing tendency, we set a personalized threshold for each user. We assume that users will only unfollow when they are clearly aware that the utility of their existing relationship is low; therefore, the unfollowing threshold is set to the 0.2 quantile of the user's existing following relationship utility. This process can be formulated as follows:

[0090]

[0091]

[0092] in It refers to all users that user u follows. It is a comprehensive user representation vector. This indicates the action decision made by user u to unfollow user i.

[0093] This invention simulates user behavior based on three sub-models: after ranking candidate content using a preference model, the user agent interacts with the highest-ranked candidate content based on the behavior density predicted by the activity model. If the interaction involves content generation, the generation model is further invoked. This architecture for simulating user agent behavior based on the activity model, preference model, and generation model is universal and not limited to simulating a specific type of behavior. This architecture can simulate various specific behavior types on online content platforms, including but not limited to posting, forwarding, liking, commenting, following, and unfollowing. The activity model, preference model, and generation model logically decompose user behavior; any specific user behavior can be considered a combination of the three models. For example, posting is a combination of the activity model and the generation model, and liking is a combination of the activity model and the preference model. This division simplifies complex user behavior into three clear sub-tasks, thus achieving accuracy and stability in simulating user behavior.

[0094] III. A General Framework for Simulating Recommendation Systems on Online Content Platforms

[0095] like Figure 3 As shown, referencing common recommendation architectures in the industry, we propose a general framework for simulating recommendation systems on online content platforms to maintain a balance between recommendation accuracy and computational efficiency. This framework comprises three stages:

[0096] 1. Recall Phase: The recommender system agent uses efficient recall algorithms to recall a batch of thousands of candidate items from millions of pieces of community content;

[0097] 2. Filtering phase: The recommender agent optimizes the list by filtering out candidates that do not meet basic criteria (such as previously recommended posts);

[0098] 3. Ranking stage: The recommender system agent uses accurate but computationally complex algorithms to rank the filtered candidate items, forming the final list of hundreds of recommendations.

[0099] Based on this framework, we further define four basic recommendation strategies: random recommendation, popularity-based recommendation, social relationship-based recommendation, and collaborative filtering-based recommendation. These four strategies represent a basic classification of current recommendation algorithms based on their technical approaches, as summarized in this invention. Modern recommendation systems can all be considered combinations of these four strategies, such as proportional combinations. Most existing recommendation system agents focus on supporting different collaborative filtering-based recommendation algorithms, and can be considered special cases of the agent proposed in this invention. This classification helps us to more comprehensively understand the impact of recommendation technology on the evolution of online content communities within a larger-scale technical system. Based on this classification, and without sacrificing representativeness, this invention provides a default template for each recommendation strategy, as follows:

[0100] 1. Random Recommendation

[0101] Both recall and ranking are based on random sampling. We omitted factors beyond the content that might influence user interaction decisions, such as the author's influence. Therefore, our tested random recommendation strategy does not directly correspond to completely random recommendations in the real world, but is closer to random recommendations for high-quality content, because high-quality content is less affected by factors outside the content itself.

[0102] 2. Popularity-based recommendations

[0103] This strategy recalls the users with the most followers and the posts with the most interaction, ranking them by popularity. Due to the cold start dilemma of new posts, recommendations based entirely on popularity will continue to recommend posts from the historical dataset, as these posts have significantly higher popularity compared to new posts, until they expire.

[0104] 3. Recommendations based on social relationships

[0105] We first merge social networks and interaction networks to form an enhanced relationship network. Then, we recall users and posts within two hops of the seed user within this enhanced relationship network and sort them based on the number of connection paths. This approach incorporates many widely used social recommendation paths, such as new posts from friends, posts liked by friends, users followed by friends, and authors of historically interacting posts.

[0106] 4. Recommendation based on collaborative filtering

[0107] Most deep learning-based recommendation algorithms fall into this category. First, based on user and post interaction records, a feature model encodes users and posts into embedding vectors. Then, recall is performed through similarity vector search. We can train multiple feature models simultaneously to achieve multi-space recall. Next, we concatenate all feature vectors of users and posts with other features such as demographic information and utilize a context-aware ranking model (also known as a click-through rate prediction model, CTR model) to predict ranking scores. In our example, the feature models include BPR (Bayesian Personalized Ranking) and LightGCN (Lightweight Graph Neural Network), and the ranking model uses NFM (Factorization Neural Network). Each model will be trained for a maximum of 100 epochs, and an early stopping strategy based on the validation set is used to improve system efficiency.

[0108] The recommender system's agent takes raw data, including content representations, interaction records, and user relationships, as input. It sequentially goes through three stages: recall, filtering, and ranking. The final output is a ranked list of posts or friend recommendation candidates, which is then pushed to the user's agent. The recall and ranking stages can be implemented using one or more of four recommendation strategies in combination, while the filtering stage is implemented using predefined rules, such as filtering already recommended candidates.

[0109] The post recommendation and friend recommendation functions of the recommendation system agent are general in nature, but have different meanings on different specific platforms. For example, on social media, content such as microblogs and questions can be modeled as one of the specific types of posts, while in community forums, users such as developers and answerers can be modeled as one of the specific types of friends. Based on the above interpretation, specific recommendation services such as question-and-answer recommendation and expert recommendation can be modeled as post recommendation and friend recommendation, and thus supported by the proposed recommendation system agent.

[0110] Example 1

[0111] Implementation Environment

[0112] The large-model-driven content recommendation ecological social intelligent agent experimental system proposed in this invention is applicable to various computer systems, with a Linux system server environment (Ubuntu 20.04.5LTS operating system, 128G or more of memory, Nvidia GeForce series dedicated graphics card) recommended. The example is developed using Python 3.9 and includes graph computing and machine learning algorithm toolkits such as dgl, torch, and sklearn.

[0113] Data Instance

[0114] This invention constructs a data instance using Sina Weibo as the target platform. Data sampling employs the data sampling algorithm for online content platforms proposed in this patent. The specific process is as follows: On seed date D, select the top ten trending topics from the trend list, and then randomly sample the user groups that interact with these topics, denoted as U1. For each user in U1, further sample all users followed by these users, denoted as U2. Then, complete the follow relationships from U2 to U1, and merge U1 and U2 to form the final user sampling list U. Based on this, capture all posts published by these users within the time range [D-7, D+14] to form the final sampled post list P. Finally, capture all user interactions with posts within the same time range, including forwarding, liking, and commenting. The sampling seed time was set to 2023 / 05 / 23, and a total of 20,065 user data entries, 875,671 post data entries, 49,720 forward data entries, 81,492 like data entries, 172,922 comment data entries, and 134,014 follow data entries were collected.

[0115] User agent instance

[0116] Encoding was performed using a SentenceTransformer (distiluse-base-multilingual-cased-v1) large language model. The activity model's classification and regression models were constructed using an MLP (Multi-Layer Perceptron) model, each with two intermediate layers of size 100. The classification model was trained using the Cross-Entropy Loss (CE Loss), and the regression model using the Mean Squared Error (MSE) loss function. A two-layer attention model consisting of a behavior layer and a user layer was used to construct the user encoding model within the preference model. An MLP with intermediate layer shapes of (1024, 512, 256) was used to construct the rating model within the preference model, trained based on the BPR loss. In the generative model, the number of GMM sub-models was set to 20. For posting behavior, ω1 and ω2 were both set to 0.5. For forwarding and commenting behavior, ω1 and ω2 were both set to 0.3, and ω3 was set to 0.4. For users with no behavioral records, both long-term and short-term interest signals were set to zero.

[0117] Recommendation system agent instance

[0118] In each round, the recommendation system agent recalls 1000 posts and 100 users, and finally recommends the 100 posts and 10 users with the highest ranking scores. Four recommendation strategies are deployed separately.

[0119] Test Results

[0120] 1. Performance Testing of the Activity Model: First, we conducted accuracy tests on the classification and regression models of the activity model, as shown in Tables 2 and 3. For the classification model, we comprehensively evaluated it using accuracy, recall, precision, and F1 score. For the regression model, we evaluated it using the prediction error Error@X (i.e., the proportion of users whose error does not exceed X) for both weak and strong user groups. Since historical data on unfollowing behavior is not publicly available, we only tested five user agent behaviors: posting, forwarding, liking, commenting, and following. The test results show that the activity model proposed in this invention achieves good accuracy and good technical results. Then, representatively, we tested the distribution stability of the activity model using posting behavior as an example, as shown in Tables 2 and 3. Figure 4 As shown in the figure, we compared the predicted and actual user activity distributions through visualization, and the two showed a high degree of consistency, demonstrating the distribution stability of the activity model proposed in this invention.

[0121]

[0122]

[0123] 2. Performance Testing of the Preference Model: We tested the accuracy of the preference model based on widely used metrics in two ranking tasks: Normalized Discounted Cumulative Gain (NDCG) and Mean Reciprocal Rank (MRR). The results are shown in Table 4. The ranking model achieved good results in all four behaviors. It is worth noting that posting behavior does not involve ranking, so no testing is needed; unfollowing and following-shared ranking models do not require separate testing. Since the preference model is only used for candidate ranking and does not directly affect the quantity and distribution of content generation, stability testing is unnecessary. Table 5 presents a performance comparison between the preference model proposed in this invention and the baseline model. Most existing baseline models consist of two parts: user modeling and similarity measurement. For user modeling, most existing solutions use dynamic user interest vectors or user history sequences to model user interaction preferences. For similarity measurement, most use dot product or cosine similarity. The results in Table 5 show that the preference model proposed in this invention significantly outperforms existing solutions in terms of accuracy, demonstrating the advancement of this invention.

[0124]

[0125]

[0126] 3. Performance Testing of the Generative Model: First, we tested the accuracy of the generative model, and the results are shown in Table 6. Using users as the basic unit, we statistically analyzed the average cosine similarity of content generation and the true cosine similarity in user history. The two showed a high degree of consistency, indicating that the generative model proposed in this invention can accurately simulate user content generation behavior. Then, we tested the distributional stability of the generative model using visualization methods, and the results are as follows: Figure 5 As shown in the figure, the red dots represent the historical distribution of user content generation, while the blue dots represent the content generation distribution of the generation model at the end of the simulation or the user content generation distribution of the real dataset 7 days later. From the figure, we can see that user content generation behavior exhibits distributional stability in real data. The generation model proposed in this invention successfully ensures the stability of the overall distribution during the simulation process by combining data-based distribution fitting and policy-based generation control. In contrast, common end-to-end generation techniques and data-based distribution fitting and generation control techniques in existing technologies cannot guarantee the stability of the distribution during the simulation.

[0127]

Claims

1. A recommendation interaction simulation system for an online content platform scenario, characterized in that: It uses intelligent agents as the basic unit to simulate the dynamic interaction between users and recommendation systems in online content platforms. It includes two types of intelligent agents: user intelligent agents and recommendation system intelligent agents. The user intelligent agent is used to perform six types of behaviors, including posting, forwarding, liking, commenting, following, and unfollowing. The recommendation system agent generates candidate post recommendation lists and friend recommendation lists to provide post and friend recommendations to the user agent; where: The user agent employs a general framework for simulating online content platform users, consisting of an activity model, a preference model, and a generation model. This framework accurately simulates user activity, user interaction preferences, and user content generation trends, unifying complex user behaviors into a standardized process of selection from a candidate list. The activity model estimates user behavior intentions and predicts behavior density based on historical interaction behavior, social relationships, and interest encoding information. The preference model predicts user preference scores for candidate posts or users based on historical interaction behavior, social relationships, and interest encoding information. The generation model fits the global distribution of posts using a Gaussian mixture model, and then predicts user content generation based on historical interaction behavior and the Gaussian mixture model encoding of candidate posts. During operation, candidate content is first ranked based on the preference model. The user agent then generates interaction behavior for the highest-ranked candidate content based on the behavior density predicted by the activity model. If the interaction behavior involves content generation, the generation model is further invoked. In the activity model, user behavior intentions are first estimated based on the UGT calculation rules at the sub-factor granularity level, including posting, forwarding, liking, commenting, and following. User intentions include: seeking information, sharing information, seeking self-status, and social interaction. When the user intention is to seek information, sub-factors include seeking interesting, useful, or new information; when the user intention is to share information, sub-factors include self-expression, providing information to others, and contributing to the community; when the user intention is to seek self-status, sub-factors include making a good impression on others, feeling important, and appearing cool and trendy; when the user intention is to social interaction, sub-factors include staying connected, meeting interesting people, sharing value, and a sense of community belonging. Then, a data-based machine learning method is used to build a predictor, learning the mapping function between behavioral intentions and user behavior density to predict behavior density. The predictor structure is a two-stage composite model, which includes a classification model that predicts the probability of users participating in each type of behavior. And a regression model for estimating the density of type a behavior. ; In the preference model, the user's long-term and short-term interest vectors are first concatenated with the latent vectors of posts in the user's interaction history. These aggregated information are then processed by a behavior-sharing user encoder to synthesize a comprehensive user representation vector as the final user encoding vector. Next, each comprehensive user representation vector is paired with the encoding vector of each candidate item, which can be a post or a user. A rating module targeting specific behaviors is used to predict the user's rating for these candidate items. The preference model is trained using ranking loss, with random negative sampling incorporated during the training process. In the generative model, a rule-based heuristic algorithm is used to manage the activation of the Gaussian mixture model; that is, for posting behavior, user activation signals are collected from long-term and short-term interests, and content is generated by activating the sub-model with the strongest signal; for forwarding and commenting behavior, the signals of the Gaussian mixture model of candidate posts are also integrated to ensure that the generated content remains relevant and realistic. The recommendation system agent adopts a general framework simulated by an online content platform recommendation system to provide post and friend recommendations to the user agent. The recommendation process of this general framework is divided into three stages: recall, filtering, and ranking. During the recall phase, each user agent filters out posts that meet specific recall criteria from all posts. During the filtering phase, posts that do not meet the basic requirements are removed. In the sorting phase, posts that best match the user's interests are selected from those that have passed the recall and filtering phases to form a recommendation list; The recommendation algorithms used in the recall and ranking stages of the general framework for online content platform recommendation systems include random recommendation, popularity-based recommendation, social-based recommendation, and collaborative filtering-based recommendation.

2. The recommendation interaction simulation system in the online content platform scenario according to claim 1, characterized in that: After encoding users' real interests and post content using a discriminative large language model, encoded user interest vectors and post vectors are obtained, which serve as inputs to the model in the general framework for user simulation on online content platforms.

3. The recommendation interaction simulation system in the online content platform scenario according to claim 1, characterized in that: The Poisson distribution P(λ) is also used to approximate the prediction of behavior caused by external intentions. The user's final behavior is defined as the sum of the intention-driven behavior predicted by the predictor and the approximate external behavior.

4. The recommendation interaction simulation system in the online content platform scenario according to claim 1, characterized in that: When users exhibit historical behaviors at different scales, the system first performs multiple samplings of different types of user behaviors with replacement to form standardized user encoder inputs. The standardized inputs obtained from the sampling are then aggregated through behavior-specific self-attention modules and additional attention modules. Subsequently, the user's long-term interest vector, short-term interest vector, and aggregated interaction representation vector of each behavior are concatenated and used as inputs to the user-level self-attention module and additional attention module to perform comprehensive user representation and synthesize a comprehensive user representation vector.

5. The recommendation interaction simulation system in the online content platform scenario according to claim 1, characterized in that: In the general framework for user simulation on online content platforms, a preference model of attention behavior is used to calculate the utility that users obtain from attention relationships, and a personalized unfollow threshold is set for each user. The unfollow threshold is set to the 0.2 quantile of the user's existing attention relationship utility. When the perceived utility is lower than the unfollow threshold, the user performs the unfollow behavior.

6. A method for simulating recommendation interaction in a recommendation interaction simulation system based on the online content platform scenario described in claim 1, characterized in that: Based on the constructed user agent and recommendation system agent, multiple rounds of simulations are conducted on the recommendation interaction of an online content platform. In each round of simulation, the recommendation system agent is first activated to generate a list of recommended posts and a list of recommended friends. Then, the user agent is activated to provide feedback on the recommended lists, including interactions such as forwarding, liking, commenting, and following, and further to take actions such as posting and unfollowing. Specifically, the user agent first predicts the user's behavior density based on an activity model, using user behavior intentions as input, and sorts the candidate list based on a preference model. Then, based on the behavior density, it generates corresponding interactive behaviors for the top candidates in the sorted candidate list. If the interactive behavior requires content generation or the posting behavior is triggered, the user agent further calls the generation model to complete the content generation. After all user agents have completed the simulation of all types of behaviors in the current round, the user agents update the user's behavior intentions based on user interests, user behavior, and social relationships to prepare for the next round of simulation.

Images