An e-sports store user habit-based personalized recommendation method and system

Abstract

The present application relates to a kind of based on computing power cloud game user personalization recommendation method and system, by collecting the user multidimensional behavior data under the scene of e-sports shop;Receive external input target customer portrait or business target request;Through the ecological load balancing model of fusion trusted data space, task decomposition and resource collaborative scheduling are carried out;Individual and group dynamic interest vector are extracted using cross-session dynamic interest model;Based on request type, by knowledge graph path reasoning or sequence-to-sequence model, output matching target customer list and game recommendation, or generate customized promotion scheme and audience group description;The present application solves the problem that the prior art is difficult to realize data-driven intelligent decision in the offline scene of e-sports shop, realizes end-to-end reasoning from business target to executable scheme, significantly improves operation accuracy and business benefit.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making and data service technology, and in particular to a personalized recommendation method and system based on the user habits of e-sports stores. Background Technology

[0002] In recent years, with the booming development of the e-sports industry and the rapid popularization of offline e-sports stores, e-sports stores have become an important offline venue connecting gamers, content providers, and brands. E-sports stores not only provide high-performance gaming equipment and environments, but also carry multiple functions such as user social interaction, event organization, game experience, and consumption, accumulating massive amounts of user behavior data, including game operation records, device usage preferences, consumption behavior, social interactions, and environmental perception information. How to make full use of this multi-source, high-dimensional, and dynamic data to achieve accurate user habit analysis and intelligent business decision support has become a key need for e-sports store operators and related brand companies to improve service quality and business benefits.

[0003] Currently, data analysis and decision support technologies for esports stores still have the following significant shortcomings: Existing user behavior analysis systems are limited in function and lack deep integration with business objectives: Traditional e-sports store management systems focus on equipment management, membership billing, or simple game time statistics. Their mining of user behavior data remains at the level of basic reports. They cannot achieve end-to-end reasoning from "target customer profile" to "personalized content recommendation" or from "business objectives" to "matching promotional plans with the target audience". E-sports store operators often rely on experience or general marketing strategies for operational decisions, lacking the ability to make accurate reasoning based on data-driven approaches. This results in poor promotional activities, low customer retention rates, and unclear returns on resource investment. Existing personalized recommendation technologies are not adapted to the characteristics of offline e-sports stores: Although cloud gaming recommendation systems can recommend games based on users' online behavior, their models are usually geared towards individual online users and do not fully consider the spatiotemporal characteristics of user behavior in e-sports stores (such as store area heatmaps, concurrent device usage, peak hour characteristics), the intensity of social interaction, and offline consumption connections. In addition, existing recommendation systems focus more on real-time interest matching and lack modeling of users' long-term interest evolution and group behavior trends, making it difficult to support e-sports stores in formulating group-level operation strategies. Lack of a computing power-data collaborative scheduling mechanism for e-sports store scenarios: Data analysis tasks in e-sports stores are characterized by high real-time performance, high concurrency, and multiple types (such as real-time recommendations, group analysis, and business reasoning), requiring efficient processing by distributed computing power clusters. However, existing resource scheduling methods are mostly general load balancing strategies that do not combine e-sports store-specific data sources (such as local device logs and IoT sensor data) with external trusted data spaces for collaborative scheduling. This results in high data processing latency and insufficient reliability of reasoning results, failing to meet real-time decision-making needs.

[0004] Therefore, there is an urgent need for a comprehensive analysis method and system for e-sports stores that can integrate multi-source behavioral data, support real-time dynamic interest modeling, have computing power-data collaborative scheduling capabilities, and provide customer matching and business decision-making reasoning through open interfaces, in order to fill the current gaps in technology in terms of scenario adaptability, reasoning depth, and system openness. Summary of the Invention

[0005] In view of this, the present invention provides a personalized recommendation method and system based on the user habits of e-sports stores, which solves the problem that existing e-sports stores are unable to deeply integrate multi-source user behavior data with business objectives, and realizes end-to-end decision support from target customers to accurate recommendations, and from business objectives to customized promotional plans.

[0006] To achieve the above objectives, the present invention provides a personalized recommendation method based on the user habits of e-sports stores, comprising the following steps: S1. Collect multi-dimensional behavioral data of e-sports store users, including local game session data, device usage preferences, consumption records, social interaction graphs and environmental perception data, perform data cleaning, remove noise and normalize the data, and then aggregate the data to generate an initial feature vector. S2. Receive service requests from esports stores and esports brand companies, including specific business objectives and target user profiles; S3. Input the initial feature vector and service request into the ecological load balancing model. In the cluster where distributed computing power nodes and trusted data space work together, decompose the task into personalized recommendation sub-tasks and management sub-tasks, and perform dynamic resource allocation and parallel processing to generate a fusion scheduling output. S4. Use the cross-conversation dynamic interest model to extract individual real-time conversation features from the initial feature vector, and fuse the interaction graph and historical behavior sequence. Through time decay integral and group interest aggregation, generate individual dynamic feature vectors and target group dynamic interest vectors. S5. Perform targeted analysis based on the type of the service request and output the inference results: S501. When the service request is a target user profile of an e-sports brand company, the individual dynamic feature vector is matched with the target user profile, and the matching target user list and their preferred game content recommendation information are output through knowledge graph path reasoning. S502. When the service request is a specific business objective of the e-sports store, based on the dynamic interest vector of the target group and the specific business objective, a targeted promotion plan and a recommended audience characteristic description are generated through a sequence-to-sequence reasoning model. S6. Collect user feedback data on the inference results, and optimize and update the parameters of the ecological load balancing model and the cross-session dynamic interest model through the influence function.

[0007] Preferably, the environmental perception data, including store area heat maps, peak-hour traffic sequences, and device idle status, is collected through IoT sensors and cameras and spatiotemporally aligned and fused with user game behavior data.

[0008] Preferably, step S3 specifically includes the following steps: S301. Decompose the task into a set of personalized recommendation subtasks and a set of business reasoning subtasks, and calculate the dynamic weight and load for each subtask based on the task urgency, data dependency and computing cost. S302. Real-time assessment of the comprehensive cost of each computing node, including computing cost, network latency, memory usage, and historical failure rate; S303. Based on the principles of ecological game competition and force field equilibrium, calculate the resource allocation ratio of each computing node, and optimize the overall processing latency and maximize the credibility of inference data. S304. Based on historical task execution efficiency, current load dynamics, and node reliability, a multi-agent reinforcement learning model is used to adaptively adjust resource allocation strategies and fault fallback mechanisms. S305. Schedule the subtasks to the optimal node in parallel, and synchronize the required multi-dimensional user behavior data from the trusted data space. Finally, aggregate the outputs of each node to generate a scheduling output vector and an inference report for e-sports stores, including recommended target user groups and promotional plans.

[0009] Preferably, the expression for calculating the resource allocation ratio of each computing node based on the ecological game competition and force field equilibrium principle is as follows: in, Represents a node The proportion of resources allocated, Indicates the weight of the subtask. Indicates the load size. Represents a node Current computing power cost, Indicates the attenuation coefficient. Indicates the number of subtasks. Indicates the number of nodes.

[0010] Preferably, the cross-conversation dynamic interest model generates individual dynamic feature vectors and target group dynamic interest vectors by including the following steps: S401. Use a location-aware attention mechanism to extract real-time conversation features; S402. Construct a user-game-social knowledge graph in the e-sports store scenario, and use a graph neural network to encode the graph structure information; S403: Integrate real-time conversation features, knowledge graph encoding vectors, and time decay integrals of historical behavior to generate individual dynamic feature vectors; S404. Cluster and aggregate the individual dynamic feature vectors of all users within a specific target user group to generate a target group dynamic interest vector that represents the overall interest tendency and trend of the specific target user group.

[0011] Preferably, the expression for the time decay integral is: in, Indicates user In time Interest points The function representing the importance weight of a session. Indicates the attenuation rate. Indicates characteristics of historical sessions.

[0012] Preferably, the knowledge graph path reasoning specifically involves: in the user-interest-game content knowledge graph, starting from the target user profile, calculating the path probability to various game content nodes through a pre-trained graph attention network, and outputting the top N game contents with the highest probability as recommendation information.

[0013] Preferably, the sequence-to-sequence reasoning model takes the business target text description as the input sequence, and generates a conditional sequence by combining the target group's dynamic interest vector through an encoder-decoder architecture. The output sequence is a structured promotional plan text and audience user characteristic keywords.

[0014] This invention also provides a personalized recommendation system based on the user habits of e-sports stores, comprising: The data acquisition and preprocessing module collects and processes multi-source user behavior data in the e-sports store scenario to generate initial feature vectors. The public service interface module is used to receive and process inference service requests from esports stores or brand companies; The computing power and data collaborative scheduling module, based on the ecological load balancing model, performs task decomposition and dynamic resource allocation between the distributed computing power cluster and the trusted data space. The User and Group Dynamics Modeling Module is used to build cross-session dynamic interest models and generate individual and group dynamic interest vectors. The business decision reasoning engine module is used to perform customer matching or solution generation reasoning based on the request type. The results feedback and model update module is used to output inference results and continuously optimize the system model using feedback data.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention introduces a business decision-making reasoning engine module, which can proactively respond to the business goals input by users (e-sports stores / brands). It can not only output accurate lists of target customers and their preferences, but also automatically generate customized promotional strategies and matching audience descriptions. It directly transforms data analysis capabilities into executable and evaluable business decision-making solutions, greatly improving operational efficiency and the accuracy of marketing activities. This invention constructs a cross-session dynamic interest model that integrates graph structure encoding and time decay integral. This model can not only accurately depict the real-time interest changes of individual users, but also gain insights into the behavioral trends and evolutionary patterns of specific customer groups through the aggregation of group interests. This dual "individual-group" modeling capability enables the system to serve personalized recommendations as well as support group-level business trend analysis and target audience positioning, thus expanding its application dimensions. Attached Figure Description

[0016] Figure 1 is a flowchart of the method of the present invention; Figure 2 This is a system module framework diagram of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] Example 1

[0019] This embodiment proposes a personalized recommendation method based on the user habits of e-sports stores, including the following steps: S1. Collect multi-dimensional behavioral data of e-sports store users, including local game session data, device usage preferences, consumption records, social interaction graphs and environmental perception data, perform data cleaning, remove noise and normalize the data, and then aggregate the data to generate an initial feature vector. Environmental perception data, including store area heat maps, peak-hour traffic sequences, and device idle status, is collected through IoT sensors and cameras and spatiotemporally aligned and fused with user game behavior data. S2. Receive service requests from esports stores and esports brand companies, including specific business objectives and target user profiles; S3. Input the initial feature vector and service request into the ecological load balancing model. In the cluster where distributed computing power nodes and trusted data space work together, decompose the task into personalized recommendation sub-tasks and management sub-tasks, and perform dynamic resource allocation and parallel processing to generate a fusion scheduling output. S301. Decompose the task into a set of personalized recommendation subtasks and a set of business reasoning subtasks, and calculate the dynamic weight and load for each subtask based on the task urgency, data dependency and computing cost. S302. Real-time assessment of the comprehensive cost of each computing node, including computing cost, network latency, memory usage, and historical failure rate; S303. Based on the principles of ecological game competition and force field equilibrium, calculate the resource allocation ratio of each computing node, and optimize the overall processing latency and maximize the credibility of inference data. The expression for calculating the resource allocation ratio of each computing node based on the principles of ecological game competition and force field equilibrium is as follows: in, Represents a node The proportion of resources allocated, Indicates the weight of the subtask. Indicates the load size. Represents a node Current computing power cost, Indicates the attenuation coefficient. Indicates the number of subtasks. Indicates the number of nodes; S304. Based on historical task execution efficiency, current load dynamics, and node reliability, a multi-agent reinforcement learning model is used to adaptively adjust resource allocation strategies and fault fallback mechanisms. S305. Schedule the subtasks in parallel to the optimal node, and synchronize the required multi-dimensional user behavior data from the trusted data space. Finally, aggregate the outputs of each node to generate a scheduling output vector and an inference report for e-sports stores, including recommended target user groups and promotional plans. S4. Use the cross-conversation dynamic interest model to extract individual real-time conversation features from the initial feature vector, and fuse the interaction graph and historical behavior sequence. Through time decay integral and group interest aggregation, generate individual dynamic feature vectors and target group dynamic interest vectors. S401. Use a location-aware attention mechanism to extract real-time conversation features; S402. Construct a user-game-social knowledge graph in the e-sports store scenario, and use a graph neural network to encode the graph structure information; S403. By fusing real-time conversation features, knowledge graph encoding vectors, and time decay integrals of historical behavior, an individual dynamic feature vector is generated. The expression for the time decay integral is: in, Indicates user In time Interest points The function representing the importance weight of a session. Indicates the attenuation rate. Indicates characteristics of historical conversations; S404. Cluster and aggregate the individual dynamic feature vectors of all users within a specific target user group to generate a target group dynamic interest vector that represents the overall interest tendency and trend of the specific target user group. S5. Perform targeted analysis based on the type of the service request and output the inference results: S501. When the service request is a target user profile of an e-sports brand company, the individual dynamic feature vector is matched with the target user profile, and the matching target user list and their preferred game content recommendation information are output through knowledge graph path reasoning. The knowledge graph path reasoning is as follows: In the knowledge graph of user-interest-game content, starting from the target user profile, the path probability to various game content nodes is calculated through a pre-trained graph attention network, and the top N game contents with the highest probability are output as recommendation information. S502. When the service request is a specific business objective of the e-sports store, based on the dynamic interest vector of the target group and the specific business objective, a targeted promotion plan and a recommended audience characteristic description are generated through a sequence-to-sequence reasoning model. The sequence-to-sequence reasoning model takes the business objective text description as the input sequence, and generates a conditional sequence by combining the dynamic interest vector of the target group through an encoder-decoder architecture. The output sequence is a structured promotional plan text and audience user characteristic keywords. S6. Collect user feedback data on the inference results, and optimize and update the parameters of the ecological load balancing model and the cross-session dynamic interest model through the influence function.

[0020] Example 2

[0021] This embodiment provides a personalized recommendation system based on the user habits of e-sports stores, including: The data acquisition and preprocessing module collects and processes multi-source user behavior data in the e-sports store scenario to generate initial feature vectors. The public service interface module is used to receive and process inference service requests from esports stores or brand companies; The computing power and data collaborative scheduling module, based on the ecological load balancing model, performs task decomposition and dynamic resource allocation between the distributed computing power cluster and the trusted data space. The User and Group Dynamics Modeling Module is used to build cross-session dynamic interest models and generate individual and group dynamic interest vectors. The business decision reasoning engine module is used to perform customer matching or solution generation reasoning based on the request type. The results feedback and model update module is used to output inference results and continuously optimize the system model using feedback data.

[0022] Example 3

[0023] 1. Data Collection This embodiment collects real user behavior logs from a gaming store from January 2023 to June 2023, a total of 6 months. The dataset includes the following parts: User behavior data: Covering the game behavior records of 1 million active users, including game duration sequences. Operational preference sequence (such as hero selection frequency, operation frequency, etc.), and social interaction data; Social graph data: a graph of social relationships among users. , where nodes Represents the user and their friends, side It represents the intensity of interaction (such as the frequency of team formation) and contains approximately 5 million edges. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio. All data has been anonymized and personally identifiable information has been removed to ensure privacy protection.

[0024] 2. Experimental Environment The experiment was conducted on a private cloud cluster, with the following cluster configuration: Hardware environment: 50 compute nodes, each node is equipped with a 32-core CPU, 128GB of memory, and 4 NVIDIA A100 GPUs (40GB of video memory). Software environment: Operating system is Ubuntu 20.04, deep learning framework is PyTorch 1.9.0, and distributed scheduling framework is based on Apache Spark 3.2.0; Network environment: Nodes are interconnected via a 100Gbps InfiniBand network to ensure low-latency data transmission.

[0025] Evaluation indicators The experiment designed evaluation indicators from two aspects: recommendation effectiveness and system performance. Recommendation performance: Precision, Recall, and Normalized Discounted Cumulative Gain (NDCG) are used to evaluate the matching degree between the recommended content and the user's interests. In this embodiment, N=5 and 10. User experience: The long-term effectiveness of the recommendation system is indirectly evaluated through user retention rate (%) and average click-through rate (CTR,%).

[0026] Experimental process 4.1 Ecological Load Balancing Model User behavior data is broken down into sub-tasks by type (such as duration analysis, operation analysis, and social analysis), and the weight of each sub-task is calculated. (Based on task priority) and load (Based on data volume), for example, the weight of the duration analysis subtask is set to 0.3, and the load is 1.2TB; Calculate resource allocation ratio based on improved ecological game competition formula : in, , , Competition penalties and diversity punishment Calculated dynamically using historical scheduling data; Subtasks are assigned to nodes for parallel processing, and aggregated to generate a scheduling output vector. As input to the next module, record the resource allocation ratio of each node. Task processing latency and resource utilization, for example, node 1's =0.12, processing latency is 23ms, utilization rate is 85%; 4.2 Cross-Session Dynamic Interest Model Features are extracted from each user's session data using a location-aware attention mechanism. : Among them, position encoding Query vector generated using the sine function. Initialized by user's historical behavior; User interest representation is calculated based on the improved dynamic model formula. : ,in, , The integral term is approximated using discrete time steps in the social graph. Encoded via a graph convolutional network (GCN); Generate user feature vectors This is used for subsequent interest tag generation and records each user's session characteristics. Cross-session interest representation and final feature vector For example, user 1's session feature dimension is 128, and the final feature vector dimension is 64; 5. Experimental Results and Analysis (1) The ecological load balancing model EGRS proposed in this invention is compared with the following three methods: Round-Robin (RR): round-robin scheduling, uniformly distributing tasks; Least-Cost-First (LCF): prioritizing the allocation of nodes with the lowest cost; Load-Balancing (LB): a scheduling strategy based on load balancing; the experimental results are shown in Table 1: Table 1 Comparison of Computational Power Scheduling Performance The Ecological Load Balancing Model (EGRS) proposed in this invention significantly outperforms benchmark methods in terms of average latency, resource utilization, and throughput. Latency is reduced by approximately 20%-37%, utilization is increased by approximately 10%-33%, and throughput is increased by approximately 20%-56%. This is attributed to the effective combination of ecological game competition mechanism and diversity constraints, which avoids excessive resource concentration and single-point overload. (2) The cross-session dynamic interest model proposed in this invention is compared with the following three methods: Session-Based RNN (SBRNN): a session recommendation model based on RNN; Graph Neural Network (GNN): pure graph structure modeling without time decay; Attention-Based Model (ABM): session feature extraction based on attention; the experimental results are shown in Table 2: Table 2 Comparison of Recommendation Effects Our proposed method outperforms the baseline method on all recommendation metrics, improving Precision@5 by approximately 12%-32% and NDCG@5 by approximately 15%-31%. This demonstrates that the combination of graph structure encoding and time decay integral effectively captures the short-term dynamics and long-term evolution of user interests. (3) User experience was evaluated through online A / B testing. The experimental group used the method described in this paper, while the control group used the traditional collaborative filtering method. The experimental results are as follows: User retention rate: 78.5% in the experimental group and 65.3% in the control group, an increase of approximately 20%; Average click-through rate: 12.7% in the experimental group and 8.9% in the control group, an increase of approximately 43%.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A personalized recommendation method based on the user habits of e-sports stores, characterized in that, Includes the following steps: S1. Collect multi-dimensional behavioral data of e-sports store users, including local game session data, device usage preferences, consumption records, social interaction graphs and environmental awareness data, perform data cleaning, remove noise and normalize the data, and then aggregate the data to generate an initial feature vector. S2. Receive service requests from esports stores and esports brand companies, including specific business objectives and target user profiles; S3. Input the initial feature vector and service request into the ecological load balancing model. In the cluster where distributed computing power nodes and trusted data space collaborate, decompose the task into personalized recommendation sub-tasks and management sub-tasks, and perform dynamic resource allocation and parallel processing to generate a fusion scheduling output. S4. Use the cross-conversation dynamic interest model to extract individual real-time conversation features from the initial feature vector, and fuse the interaction graph and historical behavior sequence. Through time decay integral and group interest aggregation, generate individual dynamic feature vector and target group dynamic interest vector. S5. Perform targeted analysis based on the type of the service request and output the inference results: S501. When the service request is a target user profile of an e-sports brand company, the individual dynamic feature vector is matched with the target user profile, and the matching target user list and their preferred game content recommendation information are output through knowledge graph path reasoning. S502. When the service request is a specific business objective of the e-sports store, based on the dynamic interest vector of the target group and the specific business objective, a targeted promotion plan and a recommended audience characteristic description are generated through a sequence-to-sequence reasoning model. S6. Collect user feedback data on the inference results, and optimize and update the parameters of the ecological load balancing model and the cross-session dynamic interest model through the influence function.

2. The personalized recommendation method based on the user habits of e-sports stores according to claim 1, characterized in that, The environmental perception data, including heat maps of the store area, peak-hour traffic sequences, and device idle status, is collected through IoT sensors and cameras and spatiotemporally aligned and fused with user game behavior data.

3. The personalized recommendation method based on e-sports store user habits according to claim 1, characterized in that, Step S3 specifically includes the following steps: S301. Decompose the task into a set of personalized recommendation subtasks and a set of business reasoning subtasks, and calculate the dynamic weight and load for each subtask based on the task urgency, data dependency and computing cost. S302. Real-time assessment of the comprehensive cost of each computing node, including computing cost, network latency, memory usage, and historical failure rate; S303. Based on the principles of ecological game competition and force field equilibrium, calculate the resource allocation ratio of each computing node, and optimize the overall processing latency and maximize the credibility of inference data. S304. Based on historical task execution efficiency, current load dynamics, and node reliability, a multi-agent reinforcement learning model is used to adaptively adjust resource allocation strategies and fault fallback mechanisms. S305. Schedule the subtasks to the optimal node in parallel, and synchronize the required multi-dimensional user behavior data from the trusted data space. Finally, aggregate the outputs of each node to generate a scheduling output vector and an inference report for e-sports stores, including recommended target user groups and promotional plans.

4. The personalized recommendation method based on e-sports store user habits according to claim 3, characterized in that, The expression for calculating the resource allocation ratio of each computing node based on the principle of ecological game competition and force field equilibrium is as follows: in, Represents a node The proportion of resources allocated, Indicates the weight of the subtask. Indicates the load size. Represents a node Current computing power cost, Indicates the attenuation coefficient. Indicates the number of subtasks. Indicates the number of nodes.

5. The personalized recommendation method based on the user habits of e-sports stores according to claim 1, characterized in that, The cross-conversation dynamic interest model generates individual dynamic feature vectors and target group dynamic interest vectors through the following steps: S401. Use a location-aware attention mechanism to extract real-time conversation features; S402. Construct a user-game-social knowledge graph in the e-sports store scenario, and use a graph neural network to encode the graph structure information; S403: Integrate real-time conversation features, knowledge graph encoding vectors, and time decay integrals of historical behavior to generate individual dynamic feature vectors; S404. Cluster and aggregate the individual dynamic feature vectors of all users within a specific target user group to generate a target group dynamic interest vector that represents the overall interest tendency and trend of the specific target user group.

6. The personalized recommendation method based on the user habits of e-sports stores according to claim 5, characterized in that, The expression for the time decay integral is: in, Indicates user In time Interest points The function representing the importance weight of a session. Indicates the attenuation rate. Indicates characteristics of historical sessions.

7. The personalized recommendation method based on e-sports store user habits according to claim 1, characterized in that, The knowledge graph path reasoning is specifically as follows: In the user-interest-game content knowledge graph, starting from the target user profile, a pre-trained graph attention network is used to calculate the path probability to various game content nodes, and the top N game contents with the highest probability are output as recommendation information.

8. The personalized recommendation method based on the user habits of e-sports stores according to claim 7, characterized in that, The sequence-to-sequence reasoning model takes the business objective text description as the input sequence, and generates a conditional sequence by combining the target group's dynamic interest vector through an encoder-decoder architecture. The output sequence is a structured promotional plan text and audience user characteristic keywords.

9. A personalized recommendation system based on the user habits of e-sports stores, using the method described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module collects and processes multi-source user behavior data in the e-sports store scenario to generate initial feature vectors. The public service interface module is used to receive and process inference service requests from esports stores or brand companies; The computing power and data collaborative scheduling module, based on the ecological load balancing model, performs task decomposition and dynamic resource allocation between the distributed computing power cluster and the trusted data space. The User and Group Dynamics Modeling Module is used to build cross-session dynamic interest models and generate individual and group dynamic interest vectors. The business decision reasoning engine module is used to perform customer matching or solution generation reasoning based on the request type. The results feedback and model update module is used to output inference results and continuously optimize the system model using feedback data.

Images