AI-based cultural activity adaptive planning method and system

Abstract

The invention discloses an AI-based cultural activity adaptive planning method and system, and relates to the technical field of information intelligent processing and activity planning, and the method comprises the following steps: obtaining user behavior data from a plurality of campus digital platforms in real time, a dynamic short-term interest vector and a dynamic long-term preference vector are generated through processing of a time sequence model and a graph neural network, and a fusion user portrait set is formed through fusion of a learnable fusion mechanism; through multi-source data real-time acquisition and time sequence model and graph neural network processing, a learnable fusion mechanism is utilized to construct a fusion user portrait, so that a cultural activity planning scheme is established on the basis of continuously perceived real group preference, a traditional subjective mode depending on static experience is changed, and the user experience is improved. Objective and dynamic planning strategies are further realized; a planning scheme is structured into a planning element combination space with spatialization and dimensionalization, and multiple optimization targets are defined and model quantification processing is carried out.

Description

Technical Field

[0001] This invention relates to the field of information intelligent processing and event planning technology, specifically to an AI-based adaptive planning method and system for cultural activities. Background Technology

[0002] When organizing cultural events within universities and other organizations, organizers often face the challenge of accurately attracting and satisfying the diverse and dynamic interests of faculty and students. Whether it's organizing a freshman arts gala, a series of academic lectures, or a cultural festival about campus traditions, organizers hope that the event can reflect mainstream values ​​while also catering to the individual preferences of students from different departments and grades, thereby achieving high participation and a good reputation. Typically, the planning team will develop the event plan by distributing questionnaires, referring to past experiences, or drawing on external case studies.

[0003] However, the planning proposals produced by this method based on limited samples and static experience are often fixed templates designed for an "imagined student group". During the warm-up and event, the actual interests, topic attention and participation behavior of teachers and students are constantly changing, but the established planning scheme is difficult to adjust accordingly. This can easily lead to a gap between the event content and real-time needs, resulting in a decrease in the overall satisfaction of participants and difficulty in achieving the expected results of the event.

[0004] Currently, existing technologies mainly rely on the subjective experience and judgment of planners, or can only perform simple statistics on limited, static individual data and discrete survey data. They cannot extract the evolution trend and subdivided differences of group interests in real time from the continuous and massive behavioral data left by teachers and students on campus digital platforms such as academic affairs systems, social forums, and online registration pages. In other words, there is a lack of a way to continuously and dynamically perceive and quantify the complex preference distribution of the target group, and the nonlinear effects caused by the dynamic interaction of the group during the activity are completely ignored.

[0005] More importantly, existing technologies struggle to systematically quantify the trade-offs between at least two interrelated / conflicting planning objectives, such as "deepening the core audience experience" versus "expanding the event's reach" and "introducing avant-garde cultural elements" versus "maintaining overall harmony." Furthermore, they cannot automatically and efficiently find a planning solution that dynamically approximates the optimal solution within the vast combination space of planning elements, including themes, formats, guests, and promotional channels. This leads to a situation where cultural event planning easily falls into the predicament of "pleasing everyone," either becoming too general and lacking appeal, or too biased towards a specific group, affecting the overall participation scale. The event's effectiveness suffers from significant uncertainty and optimization bottlenecks, potentially resulting in increased pressure on on-site event management, inefficient resource allocation, and the continuous accumulation and spread of negative experiences among participants throughout the event.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide an AI-based adaptive planning system for cultural activities. This invention addresses the problems in the background art by aggregating multi-source campus data in real time, constructing a fusion user profile, spatializing planning elements, accurately quantifying multiple planning objectives, and applying multi-objective reinforcement learning for global optimization.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an AI-based adaptive planning method for cultural activities, comprising the following steps: S1. Real-time acquisition of user behavior data from multiple campus digital platforms, processing of dynamic short-term interest vectors and long-term preference vectors using time series models and graph neural networks respectively, and fusion of these vectors through a learnable fusion mechanism to form a fused user profile set. S2. Deconstruct the cultural activity planning scheme into a finite-dimensional planning element combination space, define at least two interrelated or conflicting quantitative optimization objectives for the planning element combination space, and design corresponding quantitative calculation functions for each objective. S3. Simulate the target group's environmental state by integrating user profile sets, select elements from the planning element combination space as the agent's actions, construct a multi-objective reinforcement learning model for simulation training, drive the agent to conduct multiple rounds of policy exploration in the simulation environment, and finally output a Pareto front solution set representing the trade-off relationship between different optimization objectives. S4. Select the current optimal planning strategy from the Pareto front solution set, generate an initial activity plan based on the actual planning constraints, and dynamically adjust the execution parameters of the planning elements based on real-time user feedback data during the activity warm-up and execution process using an online learning algorithm.

[0009] Furthermore, the steps for forming the fused user profile set in step S1 are as follows: User behavior data acquisition: Real-time streaming acquisition of raw user behavior logs from campus activity registration platforms, online course systems, social forum APIs, and school academic affairs system interfaces; differentiation and processing of raw user behavior logs to obtain short-term high-frequency interaction behavior sequences and long-term stable behavior records. Dynamic short-term interest vector generation: Input short-term high-frequency interaction behavior sequences into the Transformer encoder, and obtain the user's dynamic short-term interest vector through its own self-attention mechanism; Long-term preference vector generation: Based on long-term stable behavior records, a user-activity participation bipartite graph is constructed, and a graph convolutional network is used to learn the user-activity participation bipartite graph to obtain long-term preference vectors; Adaptive fusion with a learnable fusion mechanism: The obtained dynamic short-term interest vector and long-term preference vector are adaptively fused through a fully connected network based on an attention mechanism to obtain the final fused user profile. The formation of the integrated user profile set: A corresponding integrated user profile is generated for each user in the target group, and the integrated user profiles corresponding to all users are combined into a set.

[0010] Furthermore, the design steps for the multi-objective quantization calculation function in step S2 are as follows: Define the target subset and the representation of the planning scheme: Based on the integrated user profile set, the K-Means clustering algorithm is applied to divide all users into a core audience subset and a non-core potential audience subset; For any cultural activity planning scheme to be evaluated, each element in the scheme is encoded into a vector through a pre-trained BERT text embedding model, and then a fixed-dimensional scheme feature vector is generated through a fully connected layer. Design a quantitative calculation function for the core audience's experience depth: train a machine learning regression model to integrate user profiles and scheme feature vectors as combined feature inputs, and output the predicted interaction time and predicted depth task completion rate of users with cultural event planning schemes. For a given cultural event planning scheme, calculate the comprehensive experience depth score on the core audience subset; Design a quantitative calculation function for the breadth of potential audience reach: train an exposure click-through rate prediction model, input the fused user profile and scheme feature vector of the corresponding non-core potential audience subset, and output the predicted probability of users clicking to participate in the cultural activity planning scheme after exposure, i.e., the predicted click-through rate; Train a secondary sharing intention prediction model, input the fused user profile and scheme feature vector of the corresponding non-core potential audience subset, and output the predicted probability that users will share the cultural activity planning scheme a second time, i.e. the predicted sharing rate. For a given cultural event planning scheme, calculate the comprehensive reach score on the non-core potential audience subset based on the predicted click-through rate and predicted share rate of users for the scheme. Design a quantitative calculation function for cultural value fit: Use the encoder part of a pre-trained large-scale BERT language model to encode the complete text description of the elements in the scheme and the preset cultural activity theme text into semantic vectors respectively; The quantified cultural value fit is obtained by calculating the cosine similarity between two semantic vectors. Based on the quantitative optimization goals of the core audience's experience depth, the potential audience's reach, and the cultural value alignment, a three-dimensional target vector was generated for each cultural event planning scheme.

[0011] Furthermore, the construction steps of the multi-objective reinforcement learning model in step S3 are as follows: Construct an environment simulator with an integrated reward prediction model. By receiving a cultural activity planning scheme proposed by a multi-objective reinforcement learning agent and integrating user profiles, predict and output the achievement values ​​of each quantitative optimization objective as multi-dimensional rewards. A multi-objective deep Q-network is designed as a multi-objective reinforcement learning model. The state space is the statistical features of the user profile, and the action space is the discrete selection of the combination of planning elements. During training, the multi-objective reinforcement learning agent explores planning strategies for different cultural activity planning schemes in an environment simulator and updates the network parameters of the multi-objective deep Q network based on the returned multi-dimensional rewards. An external archive is set up to continuously store all non-dominated policies during training iterations. After training, the set of planning policies in the external archive constitutes the Pareto front solution set.

[0012] Furthermore, the step S4 in which the online learning algorithm dynamically adjusts the planning elements is as follows: Based on the direction of activity resource investment and brand strategy focus, a cultural activity planning scheme is manually selected from the Pareto frontier solution set as the current optimal planning strategy. The strategy execution engine is used to transform the selected current optimal planning strategy into a detailed execution plan that includes activity schedule, bill of materials, and push plan. During the pre-event period, A / B testing was conducted on different promotional materials or registration portals using small-scale traffic tests. During the event execution period, on-site sensors and online platforms were used to collect user dwell time, interaction completion rate and social media buzz as real-time user feedback data. The contextual gambling machine algorithm is used to analyze real-time user feedback data online and dynamically adjust the display probability of variations of planning elements, resource allocation of different promotional channels, or push intensity of different experience areas.

[0013] Furthermore, an AI-based adaptive planning system for cultural activities includes a data aggregation and profiling module: used to acquire user behavior data in real time from multiple campus digital platforms through a pre-set data interface, distinguish and process short-term high-frequency interactive behavior sequences and long-term stable behavior records, and use time series models and graph neural networks to construct a fusion set of user profiles. Planning Modeling and Target Quantification Module: This module is used to deconstruct planning schemes into a limited-dimensional space of planning elements, and to define multiple quantitative optimization targets such as the depth of core audience experience, the breadth of potential audience reach, and the degree of cultural value alignment, as well as design quantitative calculation functions for the corresponding optimization targets. Strategy Exploration and Optimization Module: Used to build and run a multi-objective reinforcement learning model with integrated user profiles as the environmental state and planning element configuration as the action, and generate Pareto front solution set through simulated interaction training iteration; Strategy Execution and Adaptive Adjustment Module: This module selects the optimal planning strategy from the Pareto front solution set to generate an initial activity plan, and dynamically fine-tunes the initial activity plan based on real-time user feedback data through A / B testing and online learning algorithms. Database: Connected to various modules via API interface, used to store raw user data, intermediate profiles, strategy solutions, and feedback logs.

[0014] Furthermore, the data aggregation and profiling module includes a data interface unit, which is used to connect to various digital platforms on campus, acquire raw user behavior data in real time, and process it into short-term high-frequency interactive behavior sequences and long-term stable behavior records. The short-term interest analysis unit, with a built-in Transformer encoder, is used to process short-term high-frequency interaction behavior sequences and generate dynamic short-term interest vectors. The long-term preference analysis unit, with a built-in graph convolutional network, is used to process long-term stable behavior records and generate long-term preference vectors. The profile fusion unit has a built-in fully connected network based on an attention mechanism to fuse dynamic short-term interest vectors and long-term preference vectors, and output a fused user profile set.

[0015] Furthermore, the planning modeling and target quantification module includes a planning element library, which is used to standardize the elements of cultural activity planning in a limited dimension to form a planning element combination space, so that any cultural activity planning scheme can be transformed into a scheme feature vector that can be processed by a machine learning model. The target definition unit is used to set multiple optimization targets, including configurable core audience experience depth, potential audience reach breadth, and cultural value alignment. By configuring the weights of the targets, a multi-target optimization algorithm is used to quantify the target optimization orientation. The quantization calculation unit is used to design quantization functions for each optimization objective. It can output multi-objective reward vectors for any cultural activity planning scheme, so as to provide a unified metric for evaluating the quality of cultural activity planning schemes.

[0016] Furthermore, the strategy exploration and optimization module includes an environment simulator that integrates a reward prediction model, used to receive cultural activity planning schemes and integrated user profiles, simulate strategy effects, and output multi-dimensional rewards. A multi-objective reinforcement learning agent, employing a multi-objective deep Q-network architecture, is used for policy exploration and learning in a simulated environment; The Pareto solution set manager is used to maintain an external archive during training, save and manage all non-dominated policies, and finally output the Pareto front solution set.

[0017] Furthermore, the strategy execution and adaptive adjustment module includes a strategy execution engine, used to transform the optimal strategy selected from the Pareto front solution set into operable activity execution orders and instructions; The A / B testing unit is used to conduct concurrent grouping tests on different variations of the planning elements of a cultural event planning scheme during the event's pre-heating and execution phases. The feedback collection unit collects real-time feedback data on user participation behavior through online tracking points and offline IoT devices. The online learning adjustment unit incorporates a built-in contextual gambling machine algorithm to analyze real-time user feedback data and dynamically adjust the execution parameters of planning elements.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention utilizes real-time multi-source data acquisition and temporal modeling with graph neural networks to construct a fusion user profile through a learnable fusion mechanism. This allows cultural activity planning schemes to be based on continuously perceived real group preferences, changing the traditional subjective model that relies on static experience and further realizing the objectification and dynamism of planning strategies. By structuring the planning scheme into a spatial and dimensional combination space of planning elements, defining multiple optimization objectives and performing model quantification, and applying a multi-objective reinforcement learning model, it can automatically explore and obtain Pareto front solutions with multiple scientifically balanced schemes. That is, it automatically finds a scientific balance point among complex conflicting objectives, getting rid of the limitations of relying on subjective human trade-offs, and providing decision-makers with clear and quantifiable strategic options after trade-offs, making the decision-making process more transparent and scientific. Furthermore, during the activity preparation period, high-quality global optimization is performed through simulation to generate a baseline scheme, while during the activity warm-up and execution periods, dynamic fine-tuning is performed based on real-time feedback data using an online learning contextual algorithm. This significantly improves the ability of cultural activity planning schemes to cope with dynamic changes, not only enabling the schemes to resist dynamic interference and continuously optimize, but also significantly improving the certainty of activity effects and resource utilization efficiency. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart of the AI-based adaptive planning method for cultural activities according to the present invention.

[0021] Figure 2 This is a block diagram of the AI-based adaptive planning method for cultural activities according to the present invention. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0023] Example 1 This invention provides, for example Figure 1 The AI-based adaptive planning method for cultural activities, as shown, includes the following steps: S1. Real-time acquisition of user behavior data from multiple campus digital platforms, including short-term high-frequency interaction behavior sequences and long-term stable behavior records. Using time series models to process short-term high-frequency interaction behavior sequences to generate dynamic short-term interest vectors, using graph neural networks to process long-term stable behavior records to generate long-term preference vectors, and using a learnable fusion mechanism to fuse dynamic short-term interest vectors and long-term preference vectors to form a fused user profile set that is both timely and stable, serving as a digital representation of group preferences. In a preferred embodiment, the steps for forming the fused user profile set in step S1 are as follows: User behavior data acquisition: Real-time streaming acquisition of raw user behavior logs, including but not limited to user click streams, comment sequences, historical participation records, and community-related data, from campus activity registration platforms, online course systems, social forum APIs, and school academic affairs system interfaces. The raw user behavior logs are differentiated and processed, and the timestamp events such as click streams and comment sequences that occur within a short time window for each user are arranged in ascending order of occurrence time to form a short-term high-frequency interactive behavior sequence. At the same time, the user's historical activity participation record list, affiliated club and major information, and long-term followed forum section list are extracted to form a long-term stable behavior record. Dynamic short-term interest vector generation: Each event in the short-term high-frequency interaction behavior sequence is mapped into a dense vector through an embedding layer. The mapped sequence is input into the Transformer encoder. Through its own self-attention mechanism, the correlation weight of the corresponding dense vector between any two events in the mapped sequence is calculated, thereby capturing the long-distance dependencies and patterns between behaviors. The hidden state vectors of all time steps output by the Transformer encoder are pooled to obtain the user's dynamic short-term interest vector. Long-term preference vector generation: Based on long-term stable behavior records, a user-activity participation bipartite graph is constructed. The node set of the bipartite graph includes user nodes and entity nodes related to activity participation. Edges are established between users and entities participating in activities. A graph convolutional network is used to learn the user-activity participation bipartite graph. Through multi-layer message passing, the feature vector of each user node will aggregate the information of neighboring entities. After multi-layer convolution, the final embedding vector of the user node is used as the long-term preference vector. Adaptive fusion with a learnable fusion mechanism: The obtained dynamic short-term interest vector and long-term preference vector are adaptively fused through a fully connected network based on an attention mechanism. First, the two vectors are concatenated and then passed through a single-layer feedforward network and a Softmax function to generate a pair of normalized attention weights. Then, the calculated attention weights are applied to the original vectors respectively, and the weighted original vectors are concatenated and passed through a fully connected layer for dimensionality reduction and feature integration to obtain the final fused user profile. The formation of the integrated user profile set: A corresponding integrated user profile is generated for each user in the target group, and the integrated user profiles corresponding to all users are combined into a set, forming an integrated user profile set that describes the complex preference distribution of the current target group.

[0024] To further clarify, the temporal model is a Transformer encoder, the graph neural network is a graph convolutional network, and the learnable fusion mechanism is a fully connected network based on an attention mechanism.

[0025] The working principle of the above technical solution is as follows: User behavior logs are continuously and automatically collected from heterogeneous data sources such as campus card consumption systems, online course platforms, activity registration mini-programs, and campus social media via API interfaces. Based on the time attributes and patterns of the behavior, high-frequency interactions such as recent likes, comments, and shares are automatically marked as short-term high-frequency interaction behavior sequences. Low-frequency student attributes and behaviors with long-term cumulative effects, such as all historical lecture records and long-term interest groups, are marked as long-term stable behavior records. A Transformer time-series model is used to capture the continuous patterns of interest drift within the short-term high-frequency interaction behavior sequences, generating a dynamic short-term interest vector representing immediate and volatile interests. For long-term stable behavior records, a bipartite graph with user-entity (activity) nodes is constructed, and a graph neural network (GCN) is used for representation learning. The rich connections in the graph are used to enhance the inference of sparse user preferences, generating a long-term preference vector representing stable and deep preferences. Finally, a neural network with an attention mechanism is used to assign appropriate weights to the short-term interest vector and the long-term preference vector according to the specific context of the current planning task, and they are merged into a unified and structured fused user profile. The above technical solution achieves the following results: it solves the problem of traditional methods relying on static and partial user data, realizes dynamic, continuous and comprehensive perception of user preferences, effectively alleviates the problem of data sparsity by distinguishing between short-term and long-term behaviors and utilizing the association inference ability of graph neural networks, and can generate relatively accurate preference profiles even for new users or users with few behavior records. Furthermore, the fusion profiles provide high-quality and computable basic data for subsequent personalized recommendations and group analysis.

[0026] S2. Deconstruct the cultural event planning scheme into a limited space of planning elements consisting of event theme, content form, guest attributes and publicity channels. Define multiple quantitative optimization objectives for the planning element combination space, including at least the depth of core audience experience, the breadth of potential audience reach and the degree of cultural value fit. Design quantitative calculation functions based on prediction or statistics for each objective. In a preferred embodiment, the design steps of the multi-objective quantization calculation function in step S2 are as follows: Define the target subset and the representation of the planning scheme: Based on the fused user profile set formed in S1, the K-Means clustering algorithm is applied to divide all users into a core audience subset and a non-core potential audience subset; Establish a space for combining planning elements that are optional and include limited dimensions such as event theme, content format, guest attributes, and publicity channels. Any cultural event planning scheme can be represented as a specific coordinate point in the space for combining planning elements. Then, any cultural event planning scheme to be evaluated is vectorized according to the space for combining planning elements. That is, each element in the cultural event planning scheme is encoded into a vector through a pre-trained BERT text embedding model, and then a fixed-dimensional scheme feature vector is generated through a fully connected layer. Design a function to quantify the depth of the core audience experience: Train a machine learning regression model with a combination of user profiles and program feature vectors as input, outputting the predicted interaction duration and predicted depth task completion rate of users' cultural activity planning programs. Supervised learning training is performed using historical activity data. The expression for the machine learning regression model is as follows: In the formula, Represented as user Regarding the plan Predicted interaction duration, Represented as user Regarding the plan Predicting the completion rate of deep tasks, This is represented as a machine learning regression model. Represented as user The integration of user profiles Represented as the feature vector of the scheme, Represented as a user index, This is presented as an index of cultural event planning schemes; For a given cultural event planning scheme, calculate the comprehensive experience depth score on a subset of the core audience. The comprehensive experience depth score is a weighted geometric mean of predicted interaction time and deep task completion rate, emphasizing the balanced importance of both. The formula for calculating the comprehensive experience depth score is as follows: ,and In the formula, Represented as the first Each cultural event planning proposal is calculated within the core audience subset. The overall experience depth score on the platform This represents the number of users within the core audience subset. The weight index is represented by the interaction duration. This is represented by a weighted index indicating the completion rate of deep tasks. Design a quantitative calculation function for the breadth of potential audience reach: Train an exposure-click-through rate (CTR) prediction model, input the fused user profile and scheme feature vector corresponding to the non-core potential audience subset, and output the predicted probability that users will click to participate in the cultural activity planning scheme after exposure, i.e., the predicted CTR. The expression for the exposure-click-through rate prediction model is: In the formula, Represented as user Regarding the plan Predicted click-through rate, This is represented as an exposure click-through rate prediction model; The secondary sharing intention prediction model is trained by inputting a fused user profile and scheme feature vector corresponding to a subset of non-core potential audiences, and outputting the predicted probability, i.e., the predicted sharing rate, of users sharing a cultural activity planning scheme after participation. The expression for the secondary sharing intention prediction model is as follows: In the formula, Represented as user Regarding the plan Predicted sharing rate This is represented as a secondary sharing willingness prediction model; For a given cultural event planning scheme, based on users' predicted click-through rate and predicted share rate, a comprehensive reach score is calculated on a non-core potential audience subset. The formula for calculating the comprehensive reach score is as follows: ,and In the formula, Represented as the first Each cultural event planning proposal is calculated within a non-core potential audience subset. The overall radiation breadth score, This is represented as a weighting coefficient balancing direct reach and amplified propagation. This represents the number of users in the non-core potential audience subset. The design of a quantitative calculation function for cultural value alignment involves using the encoder part of a pre-trained large-scale BERT language model to encode the complete textual descriptions of the elements within the scheme and the pre-defined thematic text of the cultural activity into semantic vectors. The expression for the semantic vector is as follows: In the formula, Represented as a scheme The semantic vector of the complete text description Represented as the BERT language model, Represented as a scheme A complete textual description of the internal element. Represented as a semantic vector of the main text of a cultural activity. This is presented as the main text of a cultural activity. The quantified cultural value fit is obtained by calculating the cosine similarity between two semantic vectors. The formula for calculating the cultural value fit is as follows: In the formula, This is represented as the degree of cultural value alignment between two semantic vectors. Represented as the L2 norm of semantic vectors; Based on the quantitative optimization goals of core audience experience depth, potential audience reach breadth, and cultural value alignment, a three-dimensional target vector was generated for each cultural activity planning scheme. The expression for the target vector is as follows: In the formula, Represented as a scheme The target vector is the solution. Predicted values ​​for each optimization objective.

[0027] To further clarify, the quantification of the core audience's experience depth is based on a predictive model of the expected interaction time and deep task completion rate of the target user group under the cultural event planning scheme. The quantification of the potential audience reach is based on a predictive model of the dissemination conversion indicators of predicted exposure click-through rate and willingness to share again for non-core user groups under the cultural event planning scheme. The quantification of cultural value alignment is achieved by calculating the semantic similarity between the planning proposal description text and the pre-set cultural activity theme text.

[0028] The working principle of the above technical solution is as follows: the system maintains a limited-dimensional and selectable planning element library. The planning idea of ​​any specific cultural activity planning method can be mapped into a vector in this multi-dimensional "planning element combination space". At the same time, multiple optimization objectives are defined for the core audience experience depth, potential audience reach breadth and cultural value fit of this activity. By designing a quantitative prediction model corresponding to the multiple optimization objectives, an objective and unified measurement standard is provided for evaluating the quality of any planning scheme.

[0029] The effect of the above technical solution is to transform the unstructured, inspiration-dependent or subjective planning and creative process into a search and optimization problem in a structured space. By defining at least two interrelated or conflicting quantitative optimization objectives, it provides flexible objective configuration capabilities for different activities, thereby enhancing the system's universality and configurability.

[0030] S3. Simulate the target group's environmental state by integrating user profile sets, select elements from the planning element combination space as the agent's actions, construct a multi-objective reinforcement learning model for simulation training, drive the agent to conduct multiple rounds of strategy exploration in the simulated environment, calculate multi-dimensional rewards based on the quantitative optimization objective function after each round of exploration, and finally output the Pareto front solution set after the multi-objective reinforcement learning model iterates, where each solution represents a candidate planning strategy that achieves a specific balance between different optimization objectives; In a preferred embodiment, the construction steps of the multi-objective reinforcement learning model in step S3 are as follows: An environment simulator integrating a reward prediction model is constructed. This simulator receives a cultural activity planning scheme and a fused user profile proposed by the multi-objective reinforcement learning agent, predicts and outputs the achievement values ​​of each quantified optimization objective as a multi-dimensional reward. The definitions of environment, state, and action are as follows: Define the environment: Build a reward prediction model using an environment simulator, with the target vector from step S2 as the output and a fusion user profile set and a cultural event planning scheme as the input; Define state: In Time, State It is an aggregated statistical summary of the preferences of the current target group, rather than the original set of all user profiles, in order to control the dimensionality of the state space, where the expression for the state is: ,and In the formula, Indicated as current Current state Represented as a statistical feature extraction function, Indicated as current Moment-based integrated user profile set , Represented as A fusion user profile set formed by individual users It is represented as the mean vector of the merged user profile set vectors. Represented as the standard deviation vector of the merged user profile set vectors. This represents the result after principal component analysis of the merged user profile set. The variance contribution rate of each principal component; Define action: Action Corresponding to a specific selection made in the planning element combination space defined in step S2, each planning element dimension has multiple possible values, such as action. Furthermore, the size of the action space is equal to the product of the number of selectable values ​​for each element; A multi-objective deep Q-network is designed as a multi-objective reinforcement learning model. The state space is a fusion of statistical features from user profiles, and the action space is a discrete selection of combinations of planning elements. The design steps of the multi-objective deep Q-network are as follows: Multi-objective deep Q-network: A multi-objective reinforcement learning agent contains a deep Q-network. The input is the state. The output is the size of the action space. A vector of dimension, where each element Indicates the state Next, take action The expected cumulative return that can be obtained in scalar metrics; Scalar functions: These correspond to the three optimization objectives defined in the objective vector, specifically the preference weight vector. ,and ,as well as In the formula, Represented as corresponding respectively , , The preference weights for the three objectives are determined at each interaction step, and based on the reward prediction model, the environment returns a multi-objective instantaneous reward vector. Then, through linear scalarization, the multi-objective instant reward vector is transformed into a scalar instant reward. ; During training, the multi-objective reinforcement learning agent explores planning strategies for different cultural activity planning schemes in an environment simulator, and updates the network parameters of the multi-objective deep Q-network based on the returned multi-dimensional rewards. The steps for updating the network parameters are as follows: Objective function and loss function: Following the standard DQN algorithm, the goal is to learn an optimal action-value function. Such that the Bellman optimal equation is satisfied. In the formula, Represented as a discount factor, This is indicated as the next state. Indicates the next action; Due to the environmental state that characterizes the statistical features of group preferences It is static during an event planning cycle, that is... Then the Bellman optimal equation simplifies to ; The corresponding loss function is then used to train the deep Q-network. The expression for the loss function is: In the formula, Represented as a loss function, This is represented as periodically obtaining target network parameters. Copied; An external archive is set up to continuously store all non-dominated policies during training iterations. After training, the set of planning policies in the external archive constitutes the Pareto front solution set. The steps for generating the Pareto front are as follows: Multi-weight training loop: Prepare a set of N weight vectors uniformly distributed on a 3D weight simplex. For each weight vector The steps for performing iterative training are as follows: Initialization: Reset the environment based on the merged user profile set. Get the initial state ; Exploration and Experience Replay: Based on Current Deep Q-Networks ,use Greedy strategy selects action The environmental reward prediction model calculates the multi-objective reward vector. and through Convert to scalar rewards, and convert experience tuples Store in shared experience replay buffer ; Network Updates: From Shared Experience Replay Buffer A small batch of data is randomly sampled, and the loss function is applied. Update current weight Corresponding deep Q network parameters ; Strategy Evaluation and Archiving: Regularly evaluate current cultural event planning schemes. The reward prediction model in the environment calculates the target vector corresponding to the cultural activity planning scheme. ; Update external archives The current cultural event planning proposal obtained from the evaluation and target vector If the current cultural activity planning scheme is not dominated by any strategy in the archive, it is added to the archive and sorted by non-dominance, that is, all strategies in the archive that are dominated by the current cultural activity planning scheme are removed. Pareto front solution set output: After fully training all N weight vectors, the external archive... All strategies saved in and its corresponding multi-objective vector values Together, they form an approximate Pareto front solution set.

[0031] To further clarify, the multi-objective reinforcement learning model is a multi-objective deep Q-network. Its training process calculates multi-dimensional rewards through an environment simulator that integrates prediction models, and utilizes an external archive manager to maintain and output the Pareto front solution set.

[0032] The working principle of the above technical solution is as follows: by using the fused user profile set as the state of the simulated environment, a specific solution is selected from the combination space of planning elements and defined as the action that the multi-objective reinforcement learning agent can perform; a multi-objective deep Q network is adopted so that the multi-objective reinforcement learning agent can continuously try different planning solutions in the simulated environment to achieve the corresponding optimization objectives, and the results are fed back to the multi-dimensional reward as a multi-dimensional reward. The multi-objective reinforcement learning agent outputs a Pareto front solution set through learning and training.

[0033] The above technical solution achieves the following results: it avoids the inefficiency of traditional planning that relies on manual exhaustive search or trial and error, and can efficiently and automatically explore high-quality candidate solutions in a vast strategy space. It systematically solves the problem of multi-objective trade-offs, and its output Pareto front solution set provides decision-makers with clear, data-driven options for planning decisions.

[0034] S4. Select a current optimal planning strategy from the Pareto front solution set, generate an initial activity plan based on the actual planning constraints, and collect real-time user feedback data during the activity preheating and execution process based on the initial activity plan. Use online learning algorithms to dynamically adjust the configuration parameters, display priorities, or resource allocation weights of planning elements to ensure that the cultural activity planning plan continuously approaches the optimal effect in actual operation.

[0035] In a preferred embodiment, the step S4, in which the online learning algorithm dynamically adjusts the planning elements, is as follows: Based on the direction of activity resource investment and brand strategy focus, a cultural activity planning scheme is manually selected from the Pareto frontier solution set as the current optimal planning strategy. The strategy execution engine is used to transform the selected current optimal planning strategy into a detailed execution plan that includes activity schedule, bill of materials, and push plan. During the pre-event period, A / B testing was conducted on different promotional materials or registration portals using small-scale traffic tests. During the event execution period, on-site sensors and online platforms were used to collect user dwell time, interaction completion rate and social media buzz as real-time user feedback data. The contextual gambling machine algorithm is used to analyze real-time user feedback data online, dynamically adjusting the display probability of variations of planned elements, resource allocation for different promotional channels, or push intensity for different experience areas. The application steps of the contextual gambling machine algorithm are as follows: Define the variant of the planning element or resource configuration option to be adjusted as an arm; At each decision-making moment, a contextual feature vector is observed corresponding to environmental characteristics, activity status, and user characteristics. The decision is personalized and contextualized based on the contextual feature vector to adjust the strategy to adapt to different users and scenarios. At the current moment, select an arm based on the context feature vector and execute it to obtain an observed immediate reward; The linear upper confidence bound algorithm is applied to maintain two data structures for each arm to estimate the feature weight vector of the arm to be learned; The linear upper confidence bound UCB algorithm can be run in parallel or serially at multiple decision points during activity warm-up and execution.

[0036] To further clarify, the online learning algorithm is a contextual gambling machine algorithm.

[0037] The working principle of the above technical solution is as follows: Decision-makers select the current optimal strategy from the Pareto solution set generated in step S3 based on the actual resources and strategic priorities of the cultural activity, and transform it into an executable detailed plan; during the pre-heating period of the activity, A / B testing is conducted on two different versions of promotional posters with small-scale traffic; on the day of the activity, group experiments are conducted on different interactive sequence sequences; real-time feedback data on user participation behavior, such as click-through rate, dwell time, and interaction completion rate, are collected through online tracking, mini-program logs, and on-site sensors. This real-time feedback data is input into an online learning contextual gambling machine algorithm, which continuously evaluates the expected effects of different planning element variations in the current context and dynamically adjusts subsequent traffic allocation or execution strategies.

[0038] The effects of the above technical solution are as follows: it realizes the transformation from "preset static solution" to "dynamic evolution solution", enabling cultural activities to be dynamically optimized based on real-time feedback, and to cope with various uncertainties on site. It not only ensures the foresight and global optimization of the solution, but also has the agility and adaptability on site. Furthermore, through data-driven real-time optimization, it can maximize the immediate utility of activity resources, improve the real-time user experience, and accumulate high-quality data for subsequent iterative optimization of the model.

[0039] Example 2 This invention provides, for example Figure 1 The system shown is an AI-based adaptive planning system for cultural activities, including a data aggregation and profile building module: including a data interface unit, a short-term interest analysis unit, a long-term preference analysis unit, and a profile fusion unit. It is used to obtain user behavior data in real time from multiple campus digital platforms through a pre-set data interface, distinguish and process short-term high-frequency interactive behavior sequences and long-term stable behavior records, and use time series models and graph neural networks to build a fused user profile set. Planning Modeling and Target Quantification Module: Includes planning element library, target definition unit and quantitative calculation unit, used to deconstruct planning scheme into a limited-dimensional planning element combination space, and define multiple quantitative optimization targets such as core audience experience depth, potential audience reach breadth and cultural value fit, and design corresponding quantitative calculation functions for optimization targets; The strategy exploration and optimization module includes an environment simulator, a multi-objective reinforcement learning agent, and a Pareto solution set manager. It is used to build and run a multi-objective reinforcement learning model with the fused user profile as the environment state and the planning element configuration as the action. It generates Pareto frontier solution sets through simulated interactive training and iteration. The strategy execution and adaptive adjustment module includes a strategy execution engine, an A / B testing unit, a feedback collection unit, and an online learning and adjustment unit. It is used to select the optimal planning strategy from the Pareto front solution set to generate an initial activity plan, and to dynamically fine-tune the initial activity plan based on real-time user feedback data through A / B testing and online learning algorithms. Database: Connected to various modules via API interface, used to store raw user data, intermediate profiles, strategy solutions, and feedback logs.

[0040] In a preferred embodiment, the data aggregation and profiling module includes a data interface unit for connecting various digital platforms on campus, acquiring raw user behavior data in real-time streams, and processing it into short-term high-frequency interactive behavior sequences and long-term stable behavior records. The short-term interest analysis unit, with a built-in Transformer encoder, is used to process short-term high-frequency interaction behavior sequences and generate dynamic short-term interest vectors. The long-term preference analysis unit, with a built-in graph convolutional network, is used to process long-term stable behavior records and generate long-term preference vectors. The profile fusion unit has a built-in fully connected network based on an attention mechanism to fuse dynamic short-term interest vectors and long-term preference vectors, and output a fused user profile set.

[0041] To further clarify, the data aggregation and profiling module is deployed on the data server and continuously ingests real-time streaming data through the data interface unit. The short-term interest analysis unit and the long-term preference analysis unit are run by a GPU-equipped server using the corresponding deep learning framework to complete model inference, while the profiling fusion unit provides a fusion API as a microservice.

[0042] In a preferred embodiment, the planning modeling and target quantification module includes a planning element library, which is used to standardize the elements of cultural activity planning in a limited dimension to form a planning element combination space, so that any cultural activity planning scheme can be transformed into a scheme feature vector that can be processed by a machine learning model. The target definition unit is used to set multiple optimization targets, including configurable core audience experience depth, potential audience reach breadth, and cultural value alignment. By configuring the weights of the targets, a multi-target optimization algorithm is used to quantify the target optimization orientation. The quantization calculation unit is used to design quantization functions for each optimization objective. It can output multi-objective reward vectors for any cultural activity planning scheme, so as to provide a unified metric for evaluating the quality of cultural activity planning schemes.

[0043] In a preferred embodiment, the strategy exploration and optimization module includes an environment simulator that integrates a reward prediction model, used to receive cultural activity planning schemes and integrated user profiles, simulate the effect of the strategy and output multi-dimensional rewards. A multi-objective reinforcement learning agent, employing a multi-objective deep Q-network architecture, is used for policy exploration and learning in a simulated environment; The Pareto solution set manager is used to maintain an external archive during training, save and manage all non-dominated policies, and finally output the Pareto front solution set.

[0044] To further clarify, the policy exploration and optimization module can be deployed on a high-performance computing cluster. The multi-objective reinforcement learning agent requires a large amount of parallel computing resources, which can be implemented using distributed reinforcement learning frameworks such as Ray. The Pareto solution set manager is maintained in an in-memory database for real-time querying.

[0045] In a preferred embodiment, the strategy execution and adaptive adjustment module includes a strategy execution engine for converting the optimal strategy selected from the Pareto front solution set into actionable activity execution orders and instructions. The A / B testing unit is used to conduct concurrent grouping tests on different variations of the planning elements of a cultural event planning scheme during the event's pre-heating and execution phases. The feedback collection unit collects real-time feedback data on user participation behavior through online tracking points and offline IoT devices. The online learning adjustment unit incorporates a built-in contextual gambling machine algorithm to analyze real-time user feedback data and dynamically adjust the execution parameters of planning elements.

[0046] To further clarify, the strategy execution engine can synchronize cultural event planning schemes to the event management backend, the A / B testing unit and feedback collection unit can be connected to the user behavior analysis platform, and the online learning adjustment unit serves as an independent service, consuming feedback data streams in real time for calculation.

[0047] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0048] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0049] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0050] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based adaptive planning method for cultural activities, characterized in that, Includes the following steps: S1. Real-time acquisition of user behavior data from multiple campus digital platforms, processing of dynamic short-term interest vectors and long-term preference vectors using time series models and graph neural networks respectively, and fusion of these vectors through a learnable fusion mechanism to form a fused user profile set. S2. Deconstruct the cultural activity planning scheme into a finite-dimensional planning element combination space, define at least two interrelated or conflicting quantitative optimization objectives for the planning element combination space, and design corresponding quantitative calculation functions for each objective. S3. Simulate the target group's environmental state by integrating user profile sets, select elements from the planning element combination space as the agent's actions, construct a multi-objective reinforcement learning model for simulation training, drive the agent to conduct multiple rounds of policy exploration in the simulation environment, and finally output a Pareto front solution set representing the trade-off relationship between different optimization objectives. S4. Select the current optimal planning strategy from the Pareto front solution set, generate an initial activity plan based on the actual planning constraints, and dynamically adjust the execution parameters of the planning elements based on real-time user feedback data during the activity warm-up and execution process using an online learning algorithm.

2. The AI-based adaptive planning method for cultural activities according to claim 1, characterized in that, The steps for forming the merged user profile set in step S1 are as follows: User behavior data acquisition: Real-time streaming acquisition of raw user behavior logs from campus activity registration platforms, online course systems, social forum APIs, and school academic affairs system interfaces; differentiation and processing of raw user behavior logs to obtain short-term high-frequency interaction behavior sequences and long-term stable behavior records. Dynamic short-term interest vector generation: Input short-term high-frequency interaction behavior sequences into the Transformer encoder, and obtain the user's dynamic short-term interest vector through its own self-attention mechanism; Long-term preference vector generation: Based on long-term stable behavior records, a user-activity participation bipartite graph is constructed, and a graph convolutional network is used to learn the user-activity participation bipartite graph to obtain long-term preference vectors; Adaptive fusion with a learnable fusion mechanism: The obtained dynamic short-term interest vector and long-term preference vector are adaptively fused through a fully connected network based on an attention mechanism to obtain the final fused user profile. The formation of the integrated user profile set: A corresponding integrated user profile is generated for each user in the target group, and the integrated user profiles corresponding to all users are combined into a set.

3. The AI-based adaptive planning method for cultural activities according to claim 1, characterized in that, The design steps for the multi-objective quantization calculation function in step S2 are as follows: Define the target subset and the representation of the planning scheme: Based on the integrated user profile set, the K-Means clustering algorithm is applied to divide all users into a core audience subset and a non-core potential audience subset; For any cultural activity planning scheme to be evaluated, each element in the scheme is encoded into a vector through a pre-trained BERT text embedding model, and then a fixed-dimensional scheme feature vector is generated through a fully connected layer. Design a quantitative calculation function for the core audience's experience depth: train a machine learning regression model to integrate user profiles and scheme feature vectors as combined feature inputs, and output the predicted interaction time and predicted depth task completion rate of users with cultural event planning schemes. For a given cultural event planning scheme, calculate the comprehensive experience depth score on the core audience subset; Design a quantitative calculation function for the breadth of potential audience reach: train an exposure click-through rate prediction model, input the fused user profile and scheme feature vector of the corresponding non-core potential audience subset, and output the predicted probability of users clicking to participate in the cultural activity planning scheme after exposure, i.e., the predicted click-through rate; Train a secondary sharing intention prediction model, input the fused user profile and scheme feature vector of the corresponding non-core potential audience subset, and output the predicted probability that users will share the cultural activity planning scheme a second time, i.e. the predicted sharing rate. For a given cultural event planning scheme, calculate the comprehensive reach score on the non-core potential audience subset based on the predicted click-through rate and predicted share rate of users for the scheme. Design a quantitative calculation function for cultural value fit: Use the encoder part of a pre-trained large-scale BERT language model to encode the complete text description of the elements in the scheme and the preset cultural activity theme text into semantic vectors respectively; The quantified cultural value fit is obtained by calculating the cosine similarity between two semantic vectors. Based on the quantitative optimization goals of the core audience's experience depth, the potential audience's reach, and the cultural value alignment, a three-dimensional target vector was generated for each cultural event planning scheme.

4. The AI-based adaptive planning method for cultural activities according to claim 1, characterized in that, The construction steps of the multi-objective reinforcement learning model in step S3 are as follows: Construct an environment simulator with an integrated reward prediction model. By receiving a cultural activity planning scheme proposed by a multi-objective reinforcement learning agent and integrating user profiles, predict and output the achievement values ​​of each quantitative optimization objective as multi-dimensional rewards. A multi-objective deep Q-network is designed as a multi-objective reinforcement learning model. The state space is the statistical features of the user profile, and the action space is the discrete selection of the combination of planning elements. During training, the multi-objective reinforcement learning agent explores planning strategies for different cultural activity planning schemes in an environment simulator and updates the network parameters of the multi-objective deep Q network based on the returned multi-dimensional rewards. An external archive is set up to continuously store all non-dominated policies during training iterations. After training, the set of planning policies in the external archive constitutes the Pareto front solution set.

5. The AI-based adaptive planning method for cultural activities according to claim 1, characterized in that, The steps in step S4 where the online learning algorithm dynamically adjusts the planning elements are as follows: Based on the direction of activity resource investment and brand strategy focus, a cultural activity planning scheme is manually selected from the Pareto frontier solution set as the current optimal planning strategy. The strategy execution engine is used to transform the selected current optimal planning strategy into a detailed execution plan that includes activity schedule, bill of materials, and push plan. During the pre-event period, A / B testing was conducted on different promotional materials or registration portals using small-scale traffic tests. During the event execution period, on-site sensors and online platforms were used to collect user dwell time, interaction completion rate and social media buzz as real-time user feedback data. The contextual gambling machine algorithm is used to analyze real-time user feedback data online and dynamically adjust the display probability of variations of planning elements, resource allocation of different promotional channels, or push intensity of different experience areas.

6. An AI-based adaptive planning system for cultural activities, and an AI-based adaptive planning method for cultural activities according to any one of claims 1-5, characterized in that, This includes a data aggregation and profiling module: used to acquire user behavior data in real time from multiple campus digital platforms through a pre-built data interface, distinguish and process short-term high-frequency interaction behavior sequences and long-term stable behavior records, and use time series models and graph neural networks to construct a fused user profile set; Planning Modeling and Target Quantification Module: This module is used to deconstruct planning schemes into a limited-dimensional space of planning elements, and to define multiple quantitative optimization targets such as the depth of core audience experience, the breadth of potential audience reach, and the degree of cultural value alignment, as well as design quantitative calculation functions for the corresponding optimization targets. Strategy Exploration and Optimization Module: Used to build and run a multi-objective reinforcement learning model with integrated user profiles as the environmental state and planning element configuration as the action, and generate Pareto front solution set through simulated interaction training iteration; Strategy Execution and Adaptive Adjustment Module: This module selects the optimal planning strategy from the Pareto front solution set to generate an initial activity plan, and dynamically fine-tunes the initial activity plan based on real-time user feedback data through A / B testing and online learning algorithms. Database: Connected to various modules via API interface, used to store raw user data, intermediate profiles, strategy solutions, and feedback logs.

7. The AI-based adaptive planning system for cultural activities according to claim 6, characterized in that, The data aggregation and profiling module includes a data interface unit, which connects to various digital platforms on campus, acquires raw user behavior data in real time, and distinguishes and processes it into short-term high-frequency interactive behavior sequences and long-term stable behavior records. The short-term interest analysis unit, with a built-in Transformer encoder, is used to process short-term high-frequency interaction behavior sequences and generate dynamic short-term interest vectors. The long-term preference analysis unit, with a built-in graph convolutional network, is used to process long-term stable behavior records and generate long-term preference vectors. The profile fusion unit has a built-in fully connected network based on an attention mechanism to fuse dynamic short-term interest vectors and long-term preference vectors, and output a fused user profile set.

8. The AI-based adaptive planning system for cultural activities according to claim 6, characterized in that, The planning modeling and target quantification module includes a planning element library, which is used to standardize the elements of cultural activity planning in a limited dimension to form a planning element combination space, so that any cultural activity planning scheme can be transformed into a scheme feature vector that can be processed by a machine learning model. The target definition unit is used to set multiple optimization targets, including configurable core audience experience depth, potential audience reach breadth, and cultural value alignment. By configuring the weights of the targets, a multi-target optimization algorithm is used to quantify the target optimization orientation. The quantization calculation unit is used to design quantization functions for each optimization objective. It can output multi-objective reward vectors for any cultural activity planning scheme, so as to provide a unified metric for evaluating the quality of cultural activity planning schemes.

9. The AI-based adaptive planning system for cultural activities according to claim 6, characterized in that, The strategy exploration and optimization module includes an environment simulator, which integrates a reward prediction model. It is used to receive cultural activity planning schemes and integrate user profiles, simulate the effect of strategies, and output multi-dimensional rewards. A multi-objective reinforcement learning agent, employing a multi-objective deep Q-network architecture, is used for policy exploration and learning in a simulated environment; The Pareto solution set manager is used to maintain an external archive during training, save and manage all non-dominated policies, and finally output the Pareto front solution set.

10. The AI-based adaptive planning system for cultural activities according to claim 6, characterized in that, The strategy execution and adaptive adjustment module includes a strategy execution engine, which is used to transform the optimal strategy selected from the Pareto front solution set into operable activity execution orders and instructions; The A / B testing unit is used to conduct concurrent grouping tests on different variations of the planning elements of a cultural event planning scheme during the event's pre-heating and execution phases. The feedback collection unit collects real-time feedback data on user participation behavior through online tracking points and offline IoT devices. The online learning adjustment unit incorporates a built-in contextual gambling machine algorithm to analyze real-time user feedback data and dynamically adjust the execution parameters of planning elements.

Images