Schedule management system method and system deeply fused with artificial intelligence

Abstract

This application relates to a method and system for a schedule management system that deeply integrates artificial intelligence. The system includes collecting user schedule data via mobile applications and terminal devices and binding it to virtual avatars; dynamically constructing and incrementally updating a team-level temporal knowledge graph; performing multi-scale temporal evolution analysis on target users: using graph neural networks to learn stable habits from a long-term unified dense graph to form a long-term evolutionary representation; simultaneously, capturing recent dynamics and context through a relationship-aware attention network based on natural slices and recurrent units to generate a short-term evolutionary representation; adaptively fusing the two to obtain a comprehensive user situational representation; when responding to collaborative requests, aggregating the comprehensive situational representations of all participants, calculating the global matching degree of future time windows to avoid conflicts, generating priority-ranked schedule plans, and tracing key historical subgraphs to generate visual recommendation reasons; finally, after confirmation, the final plan is synchronized to the entire platform and a notification is displayed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a schedule management system and method that deeply integrates artificial intelligence. Background Technology

[0002] With the widespread adoption of digital living and collaborative work, schedule management has become a core tool for individuals and teams to plan tasks and improve efficiency. Traditional calendar applications can no longer meet users' needs for intelligent, collaborative, and personalized services. Therefore, schedule management systems that deeply integrate artificial intelligence technology have emerged, aiming to become an intelligent hub for individuals and teams through natural language interaction, intelligent analysis, and visualization.

[0003] Currently, existing intelligent schedule management systems collaborate with mobile applications and fixed terminal devices to provide functions such as natural language schedule creation, multi-user calendar sharing, and basic conflict detection. For example, when a user initiates a request involving multiple participants, the system checks whether each participant's calendar has a schedule marked as "busy" during the target time period. If there is a time overlap, it will indicate a conflict or suggest alternative times.

[0004] The aforementioned technologies can only perform simple time overlap comparisons based on the clearly recorded schedule items in the current calendar. They cannot understand the habitual patterns behind the schedule, the dynamic influence of the relationship network, or the evolution of the event itself on the timeline. As a result, the suggestions they generate are often mechanical and cannot provide scheduling solutions that truly match the internal rhythm of all participants and have a high success rate when faced with complex time constraints or ambiguous needs.

[0005] Based on this, this application provides a method and system for a schedule management system that deeply integrates artificial intelligence. Summary of the Invention

[0006] To address the issue that a schedule management system that can only perform simple time overlap comparisons based on clearly recorded schedule items in the current calendar, failing to understand the habitual patterns, dynamic influence of relationship networks, and the evolution of events themselves on the timeline, often generates mechanical suggestions that cannot provide truly suitable scheduling solutions with high success rates for complex time constraints or ambiguous needs, this application provides a schedule management system method and system that deeply integrates artificial intelligence.

[0007] Firstly, this application provides a schedule management system method that deeply integrates artificial intelligence, employing the following technical solution: including: The system collects schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and generates and binds a virtual avatar for each user. Based on schedule-related data, a team-level time-series knowledge graph is dynamically constructed and maintained, where each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp, and is updated in real time based on the user's subsequent actions on the schedule. Based on the aforementioned temporal knowledge graph, multi-scale temporal evolution analysis is performed on the target user. Graph neural networks are used to learn the long-term evolutionary representation of users and relationships to capture stable habits and cyclical patterns. Sequence models and relationship-aware graph attention networks are used to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target user. In response to a user's schedule creation request, for collaborative requests involving multiple users, the system obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights within the model, the system generates a visual reason for recommending the schedule plan. Receive user confirmation information and synchronize the final confirmed schedule to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provide visual prompts using each user's virtual avatar as an identifier.

[0008] Preferably, the step of collecting schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and generating and binding a virtual avatar identifier for each user, includes: Respond to user registration or avatar creation instructions in the mobile application subsystem and receive front-facing avatar images uploaded by users; The virtual avatar generation unit sends a frontal headshot image to the cloud, uses a face detection algorithm to locate facial feature points, extracts facial feature vectors, and generates a 2D stylized avatar based on a pre-trained style transfer model using a pre-set style template; or, it generates a 3D mesh of the user's face using a 3D face reconstruction model, generates a facial texture map by combining the style prompts selected by the user, and outputs a 3D virtual avatar with customizable clothing, expressions, and accessories through a rendering engine. The generated virtual avatar data is bound to the user's unique system account identifier and stored in a pre-set cloud user profile database; The bound virtual avatar data is synchronized in real time via an encrypted network to the mobile application subsystem and schedule management terminal subsystem associated with the user's account, serving as a unified visual identifier for the corresponding user across the entire system platform.

[0009] Preferably, the step of dynamically constructing and maintaining a team-level time-series knowledge graph based on schedule-related data includes: Receive schedule-related data streams from the mobile application subsystem and the schedule management terminal subsystem. The schedule-related data streams include at least new schedule creation, existing schedule modification, schedule status update and schedule deletion operations and their corresponding time information. The data stream related to the schedule is parsed, the core elements corresponding to each operation are extracted, and they are structured and represented in a four-tuple format. The subject user is the unique identifier of the user who initiates the operation or is the subject of the schedule. The relationship is a semantic tag describing the type of operation or schedule, including creation, modification, completion, and participation. The object is the target to which the operation or schedule is directed, and its type includes other user identifiers, specific task items, and location information. The timestamp is the precise time point or time period of the operation execution or schedule arrangement. The quadruples are stored as new facts in a cloud-based time-series knowledge graph database, and their relationship with historical quadruples is established in the time dimension to form a graph structure that describes the dynamic evolution of the team. Based on the user's subsequent actions on the schedule, the real-time incremental update mechanism of the graph is triggered: when the schedule-related data stream is received, the changed data part is identified, only the changed related quadruples and their associated edges are updated, and the update instructions are pushed to all relevant clients in real time through the subscription message protocol. Through an incremental update mechanism, the time-series knowledge graph continuously evolves and serves as the sole authoritative data source for subsequent multi-scale time-series evolution analysis.

[0010] Preferably, based on the temporal knowledge graph, multi-scale temporal evolution analysis is performed on the target user, and a graph neural network is used to learn the long-term evolutionary representation of the user and relationships to capture stable habits and periodic patterns. A sequence model and a relationship-aware graph attention network are used to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target user, including: From the time-series knowledge graph, extract all historical interaction quadruples of the target user within a long historical time window, and merge them to construct a unified dense knowledge graph, in which all historically appearing entities are retained as nodes, and the relationships between entities across timestamps are merged into the edges of the graph; A relational graph convolutional network is used as a semantic aggregator to perform multi-round message passing and feature aggregation on the unified dense knowledge graph, and to learn the embedding representation of target users and various relationships at a long-term scale as a long-term evolutionary representation. From the time-series knowledge graph, a series of continuous knowledge graph snapshots of the target user within a short historical time window are extracted, and the knowledge graph snapshot sequence is naturally evolved and sliced ​​based on the occurrence pattern of the target entity, dividing it into explicit slices containing the target entity and implicit slices not containing the target entity. For each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies; and a gated recurrent unit is used to capture the temporal dependencies between different snapshots within the slice to generate the local evolutionary representation of the corresponding slice. By introducing a cross-slice attention mechanism with location embedding, the local evolutionary representations of all explicit slices are adaptively weighted and aggregated, wherein the attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, thereby generating a short-term evolutionary representation of the target user. By using learnable weighted parameters, the long-term evolutionary representation and the short-term evolutionary representation are weighted and summed, and then passed through a nonlinear transformation layer to obtain the comprehensive situational representation of the target user.

[0011] Preferably, for each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies; and a gated recurrent unit is used to capture the temporal dependencies between different snapshots within the slice to generate a local evolutionary representation of the corresponding slice, including: For each knowledge graph snapshot within an explicit slice, with the target user entity as the central node, extract its k-hop neighbor entities and their relationships, and construct a locally connected subgraph, where k is a natural number greater than 0; A relation-aware graph attention network is used to encode the local connected subgraph; by calculating the attention weights of the target user node and its neighboring nodes under a specific relationship, the features and relational semantic information of the neighboring nodes are adaptively aggregated to generate a structural embedding representation of the target user under the corresponding snapshot. The structure embedding representation sequence corresponding to multiple snapshots arranged in chronological order within an explicit slice is input into a gated recurrent unit; the hidden state of the previous snapshot and the structure embedding representation of the current snapshot are used as input to update its hidden state, thus modeling the temporal evolution dependency across snapshots within the slice. The hidden state output by the gated loop unit after processing the last snapshot in the explicit slice is used as the local evolution representation of the corresponding explicit slice.

[0012] Preferably, the local evolutionary representations of all explicit slices are adaptively weighted and aggregated by introducing a cross-slice attention mechanism with location embedding, wherein the attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, generating a short-term evolutionary representation of the target user, including: Assign a learnable position index to all explicit slices arranged in chronological order and generate corresponding position embedding vectors to encode the order information of each slice in the overall time series; The local evolution representation of each explicit slice is added to the corresponding position embedding vector to generate the enhanced slice representation of the corresponding slice; Based on the current schedule query to be predicted, a query vector is generated; the semantic similarity between the query vector and the enhanced slice representation of each explicit slice is calculated as an initial score to measure the relevance of the corresponding slice to the current prediction target topic. The enhanced slice representation of each explicit slice is used as the key and value, the query vector is used as the query input, the scaled dot product attention calculation is performed, the initial score is normalized, and the adaptive attention weights corresponding to each explicit slice are generated. Based on the adaptive attention weights, the enhanced slice representations of all explicit slices are weighted and summed to obtain the aggregated representation vector, which serves as the short-term evolutionary representation of the target user.

[0013] Preferably, in responding to a user's schedule creation request, for collaborative requests involving multiple users, the system obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights within the model, a visualized recommendation reason for the schedule plan is generated, including: In response to a schedule creation request initiated by the mobile application subsystem, the natural language processing unit parses the request content and extracts the target time range, event type, and set of participating users associated with identifiers; For each user in the participating user set, the corresponding comprehensive situational representation is obtained to form a multi-user situational representation set; Based on the multi-user situational representation set, for multiple candidate time windows within the target time range, a global matching score for each candidate time window is calculated; wherein, the calculation process comprehensively evaluates the individual matching degree of each user in the corresponding window, and introduces a penalty factor to avoid time conflicts with each user's existing confirmed schedule. Select at least one candidate time window with the highest global matching score, combine it with the corresponding event type and participants, generate a structured schedule, and sort it by priority from high to low according to the global matching score; For each generated schedule plan, based on the attention weight used when generating the schedule plan, several historical quadruplets that contribute more than a preset threshold to the prediction result are located from the time-series knowledge graph, and key historical subgraphs are constructed with these quadruplets as the core. Alignment analysis is performed between the key historical subgraphs and the elements of the corresponding schedule plan to generate natural language describing the correlation between historical patterns and current decisions, serving as a visual recommendation reason for the corresponding schedule plan.

[0014] Secondly, this application discloses a schedule management system device that deeply integrates artificial intelligence, employing the following technical solution, including: The virtual avatar module is used to collect schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and to generate and bind a virtual avatar identifier for each user. The data association module is used to dynamically build and maintain a team-level time-series knowledge graph based on schedule-related data. Each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp. The module is updated in real time based on the user's subsequent actions on the schedule. The data analysis module is used to perform multi-scale temporal evolution analysis on target users based on the temporal knowledge graph, and to learn the long-term evolutionary representation of users and relationships using graph neural networks to capture stable habits and periodic patterns. It also uses sequence models and relationship-aware graph attention networks to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target users. The schedule recommendation module is used to respond to users' schedule creation requests. For collaborative requests involving multiple users, it obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights inside the model, it generates a visual reason for recommending the schedule plan. The plan confirmation module is used to receive user confirmation information and synchronize the final confirmed schedule plan to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provide visual prompts using the virtual avatars of each user.

[0015] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described above in the method of a schedule management system with deep integration of artificial intelligence.

[0016] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding a schedule management system method that deeply integrates artificial intelligence.

[0017] In summary, this application collects users' schedule data through mobile applications and terminal devices and binds it to their virtual avatars to construct and continuously update a team-level temporal knowledge graph based on four-tuples as fundamental facts. Based on this graph, the core innovation of the system lies in performing multi-scale temporal evolutionary analysis: It utilizes graph neural networks to learn from a unified dense graph constructed from long-term historical data, capturing users' stable habits and cyclical patterns to form a long-term evolutionary representation; simultaneously, through natural evolutionary slicing and employing a relation-aware graph attention network and gated recurrent units, it learns users' recent dynamics and contextual information influenced by related users to generate a short-term evolutionary representation; subsequently, the long-term and short-term representations are adaptively fused to form a comprehensive representation reflecting the user's overall situation. When responding to a user's collaborative schedule request, the system aggregates the comprehensive situational representations of all participants, calculates the global matching degree of future candidate time windows to avoid conflicts, and generates priority-ranked schedule plans. After the plan is generated, the system automatically generates visualized recommendation reasons by tracing key historical subgraphs in the temporal knowledge graph and combining them with model attention weights. Ultimately, the user-confirmed plan will be synchronized to all relevant terminals and visually prompted using each user's virtual avatar. This achieves a fundamental shift from passive recording to proactive prediction, from simple conflict checking to understanding asynchronous rhythms, and from black-box decision-making to transparent and reliable explanation by simulating deep temporal reasoning in human cognition, significantly improving the collaborative efficiency of schedule management. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a schedule management system method that deeply integrates artificial intelligence.

[0019] Figure 2 This is a flowchart of the virtual character generation process in this application.

[0020] Figure 3 This is a structural block diagram of a schedule management system device that deeply integrates artificial intelligence. Detailed Implementation

[0021] The following combination Figures 1-3 This application will be described in further detail.

[0022] In this application, the system consists of a mobile app subsystem, a terminal subsystem, and a cloud AI service module: Mobile App Subsystem: Deployed on iOS or Android smart terminals, it provides users with a personalized schedule management portal and supports functions such as AI interaction, virtual avatar customization, schedule creation and synchronization; Terminal Subsystem: A screen-equipped device running the Android system, supporting multiple sizes and models, which can be placed on a desktop or wall-mounted, serving as a central hub for displaying and interacting with family / team schedules; AI core service modules: Deployed in the cloud or on the local terminal, including natural language processing unit, 3D virtual image generation unit, intelligent schedule analysis unit and cross-platform synchronization unit, providing AI capability support for the two major subsystems.

[0023] Reference Figure 1 The embodiments of this application include at least steps S10 to S50.

[0024] S10 collects schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and generates and binds a virtual avatar for each user.

[0025] S20 dynamically constructs and maintains a team-level time-series knowledge graph based on schedule-related data. Each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp. It is also updated in real time based on the user's subsequent actions on the schedule.

[0026] S30, based on a temporal knowledge graph, performs multi-scale temporal evolution analysis on target users and uses graph neural networks to learn long-term evolutionary representations of users and relationships to capture stable habits and periodic patterns. It also uses sequence models and relationship-aware graph attention networks to learn short-term evolutionary representations to capture recent changes and contextual influences. Finally, it adaptively weights and fuses the long-term and short-term evolutionary representations to generate a comprehensive situational representation of the target users.

[0027] S40 responds to user schedule creation requests. For collaborative requests involving multiple users, it obtains a comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, it generates a visualized recommendation reason for the schedule plan by analyzing key historical subgraphs in the time-series knowledge graph and the attention weights within the model.

[0028] S50 receives user confirmation information and synchronizes the final confirmed schedule plan to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provides visual prompts using the virtual avatars of each user.

[0029] Specifically, the system first gathers all members' past schedule data from the app and team terminals, and assigns each person a unique virtual avatar. Then, the cloud dynamically constructs and maintains a detailed team time-series knowledge graph based on this data, recording every past collaboration. Based on this graph, the system performs deep behavioral analysis on each member—it mines members' long-term stable rhythms through graph neural networks, while simultaneously perceiving their recent dynamics and collaboration context through attention networks. It then merges these long- and short-term patterns to form a comprehensive status profile for each member. When processing the aforementioned collaboration requests, the system concurrently calls the status profiles of all participants, intelligently assesses the global matching degree for different time periods in the coming days, automatically avoids all conflicts, and generates a ranking scheme such as "Next Tuesday at 2 PM is best." Simultaneously, the system traces key historical evidence in the graph to generate clear and concise recommendation reasons. Once the scheme is confirmed, it is immediately synchronized to everyone's calendar and the team's public screen, and a vivid virtual avatar is used as a prompt. This transforms passive response into proactive prediction, achieving collaborative scheduling that truly understands the team's rhythm.

[0030] In some embodiments, step S10 specifically includes the following steps: responding to a user's registration or avatar creation instruction in the mobile application subsystem, receiving a frontal avatar image uploaded by the user; sending the frontal avatar image to a virtual avatar generation unit in the cloud, locating facial feature points through a face detection algorithm, extracting facial feature vectors, and generating a 2D stylized avatar based on a preset style template and a pre-trained style transfer model; or, generating a three-dimensional mesh of the user's face through a 3D face reconstruction model, generating a facial texture map by combining the style prompt words selected by the user, and outputting a 3D virtual avatar with customizable clothing, expressions, and accessories through a rendering engine; binding the generated virtual avatar data with the user's unique system account identifier and storing it in a preset cloud user profile library; and synchronizing the bound virtual avatar data in real time through an encrypted network to the mobile application subsystem and schedule management terminal subsystem associated with the user's account, serving as a unified visual identifier for the corresponding user across the entire system platform.

[0031] Reference Figure 2 In practice, the mobile app subsystem supports both iOS and Android systems, and its core functions are implemented through the following technologies: User registration and virtual avatar generation: After completing registration, users can generate AI virtual avatars (2D stylized portraits / 3D avatars). The generated result is bound to the user's account ID and synchronized to the terminal subsystem as a unified visual identifier across the entire system. The specific implementation is as follows: 1. 2D Stylized Portrait Generation Method: After the user uploads a front-facing portrait photo, the 2D feature extraction unit in the AI ​​core service module first locates key facial regions (containing 68 feature points, covering facial features and contours) using the MTCNN face detection algorithm, and generates a 1024-dimensional facial feature vector using the ArcFace deep feature extraction network (cosine similarity matching threshold ≥0.85 to ensure feature uniqueness); then, combined with the system's selected preset style (such as "3D cartoon" or "illustration" style templates), the features are fused through the pre-trained StyleGAN2 style transfer model. After the input photo is resized to 512×512 resolution, a 2D stylized portrait is generated through 50 iterations; the generated result supports fine-tuning of brightness and contrast, and finally outputs a 1024×1024 resolution WebP format image, maintaining the user's core facial features (facial proportions, eyebrow and eye features, etc.) while matching the selected style.

[0032] 2. 3D Avatar Generation Method: After the user selects this method, the AI ​​core service module generates an interactive 3D virtual avatar through the following steps: ① Face Reconstruction: A pre-trained model based on a 3D Deformation Model (3DMM) is used, with 5023 vertices, including 199 shape parameters and 29 expression parameters. A 256×256 resolution RGB portrait photo is input, and the model iterates for 100 rounds using the Adam optimizer (initial learning rate 1e-4, decaying to 0.5 of the previous round every 20 rounds), outputting 128 three-dimensional facial feature point coordinates (error controlled within 0.3 pixels) to reconstruct the three-dimensional structure of the face; ② Texture Generation: An improved Stable texture is used. The Diffusion model uses the reconstructed 3D facial mesh as a geometric constraint, integrates user-selected style prompts, sets the diffusion process sampling step size to 20 steps, uses the DDIM sampler, and has a text guidance weight of 7.5, generating a 1024×1024 resolution high-fidelity facial texture map (texture and mesh vertex mapping error ≤0.1mm); ③ Super-resolution optimization: introduces the ESRGAN network (8 residual blocks, each containing 2 3×3 convolutional layers), magnifying the texture to 2048×2048 resolution (PSNR≥35dB); ④ Rendering and customization: outputs 3D models through the PBR rendering engine, supporting clothing (10+ sets of preset templates, custom import in OBJ format), expressions (12 basic expressions, driven by Blend Shape), and accessories (glasses, headwear, etc.) customization, supporting different angle and pitch rendering and 10-second looping animations at 30fps (waving, nodding, etc.).

[0033] The two generation methods are configured and used according to different versions of the system product. The generated results are stored in the cloud user profile library. Users can switch generation methods or modify image details at any time. After modification, the system will synchronize to the terminal device in real time via the MQTT protocol to ensure consistent visual identity across the platform.

[0034] In some embodiments, step S20 specifically includes the following steps: receiving schedule-related data streams from the mobile application subsystem and the schedule management terminal subsystem, wherein the schedule-related data streams at least include new schedule creation, existing schedule modification, schedule status update, and schedule deletion operations and their corresponding time information; parsing the schedule-related data streams, extracting the core elements corresponding to each operation, and representing them in a structured format as quadruples, wherein the subject user is the unique identifier of the user who initiates the operation or is the subject of the schedule, the relationship is a semantic tag describing the type of operation or schedule, including creation, modification, completion, and participation, the object is the target to which the operation or schedule is directed, and its type includes other user identifiers, specific task items, and location information, and the timestamp is the precise time point or time period of the operation execution or schedule arrangement; storing the quadruples as new facts in a cloud-based time-series knowledge graph database, and establishing a time-dimensional association with historical quadruples to form a graph structure describing the dynamic evolution of the team; Based on subsequent user actions on the schedule, the real-time incremental update mechanism of the graph is triggered: when a schedule-related data stream is received, the changed data part is identified, and only the changed related quadruplets and their associated edges are updated. The update instructions are pushed to all relevant clients in real time through a subscription message protocol. Through the incremental update mechanism, the time-series knowledge graph continues to evolve and serves as the sole authoritative data source for subsequent multi-scale time-series evolution analysis.

[0035] Specifically, the system receives real-time data streams from applications and terminals, and through parsing, accurately extracts each operation (such as creating a "Monday team meeting" or marking "Task A completed") into a standard four-tuple of (user, relationship, task, time). This is then stored as new facts in a cloud-based knowledge graph database and linked with historical data on a timeline, thereby constructing a knowledge graph that dynamically reflects the entirety of team collaboration. More importantly, through an incremental update mechanism, it only updates and pushes updates to the changed parts in real time, ensuring the knowledge graph is always up-to-date. This provides the system with a real-time, evolving data source of team behavior, enabling all subsequent analyses (such as habit learning and conflict prediction) to be based on accurate and timely global information, overcoming the shortcomings of traditional systems with scattered and lagging data.

[0036] In some embodiments, step S30 specifically includes the following steps: Extracting all historical interaction quadruples of the target user within a long historical time window from the temporal knowledge graph, fusing them to construct a unified dense knowledge graph, where all historically occurring entities are retained as nodes, and cross-timestamp entity relationships are fused into graph edges; employing a relational graph convolutional network as a semantic aggregator, performing multi-round message passing and feature aggregation on the unified dense knowledge graph, learning the embedding representation of the target user and various relationships at a long-term scale, as a long-term evolutionary representation; extracting a series of continuous knowledge graph snapshots of the target user within a short historical time window from the temporal knowledge graph, and performing natural evolutionary slicing on the knowledge graph snapshot sequence based on the occurrence pattern of the target entities, dividing it into segments containing... The system identifies explicit slices of the target entity and implicit slices that do not contain the target entity. For each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies. A gated recurrent unit is used to capture the temporal dependencies between different snapshots within the slice, generating a local evolutionary representation of the corresponding slice. By introducing a cross-slice attention mechanism with position embedding, the local evolutionary representations of all explicit slices are adaptively weighted and aggregated. The attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, generating a short-term evolutionary representation of the target user. Through learnable weighting parameters, the long-term and short-term evolutionary representations are weighted and summed, and then passed through a nonlinear transformation layer to obtain a comprehensive situational representation of the target user.

[0037] Specifically, by constructing a unified, dense knowledge graph (i.e., integrating a static network of relationships formed by all interactions over a long historical period), and using a relational graph convolutional network to mine the implicit stable habits and periodic patterns within it, a long-term evolutionary representation is formed. The semantic aggregation process is implemented through the following relational graph convolutional network: ; in, and Let S and O represent the feature representations of entities s and o in the unified dense knowledge graph at layer l, respectively. The embedding vector representing relation r; It is the set of all fact triples (s,r,o) in the knowledge graph; and ϕ is the learnable parameter matrix of the l-th layer; ϕ is a one-dimensional convolution operator used to model the semantic translation properties of entities through relations; As a normalization factor, it is usually taken as the in-degree of entity o; It is a non-linear activation function.

[0038] Meanwhile, based on the recent data stream being divided into explicit slices (times when users are active) and implicit slices (times when users are inactive but their contacts are active) according to user activity status, within each explicit slice, a relationship-aware graph attention network is used to analyze the real-time relationship structure. The calculation of its attention weight can be expressed as: ; in, This represents the attention coefficient of entity s to its neighbor entity o under relation r at layer l; and Let represent the feature representations of entity s, entity o, and relation r at timestamp t, respectively; a is a learnable attention vector. Let represent the set of neighbors of entity s.

[0039] Furthermore, by leveraging gated loop units to capture the temporal evolution within a slice, a slice-level local representation is generated, and its state update process is as follows: ; in, This represents the hidden state matrix of all entities at timestamp t; This represents the entity representation matrix output by the relation-aware graph attention network at the previous timestamp.

[0040] Then, the slice representations are fused through a cross-slice attention mechanism to form a short-term evolutionary representation that reflects recent dynamics and context; the calculation of the attention mechanism is as follows: ; in, A query vector generated based on the current prediction target; and These are matrices obtained by adding the local representations (as keys and values) of all M explicit slices to their positional embeddings; Let the dimension be the vector. Used for scaling to avoid gradient vanishing.

[0041] Finally, the long-term and short-term representations are adaptively fused to generate a comprehensive situational representation that fully depicts the user's current state and potential patterns. This fusion process can be represented as follows: ; in, and These represent long-term evolution and short-term evolution, respectively. It is a learnable weighted parameter used to balance the influence of long-term habits and short-term dynamics on the final representation.

[0042] This overcomes the limitations of traditional models that only focus on recent events or only statistically analyze long-term frequencies. Through deep fusion of multi-scale and asynchronous perception, the system can not only identify users' long-term habits, but also perceive the short-term state of "users needing a buffer after finishing high-intensity tasks," thus laying a precise cognitive foundation for subsequent collaborative scheduling.

[0043] Furthermore, considering the problem of constructing local evolutionary representations, the corresponding processing steps are as follows: For each knowledge graph snapshot within an explicit slice, with the target user entity as the central node, extract its k-hop neighbor entities and their relationships to construct a locally connected subgraph, where k is a natural number greater than 0; use a relation-aware graph attention network to encode the locally connected subgraph; by calculating the attention weights of the target user node and its neighbor nodes under specific relationships, adaptively aggregate the features and relational semantic information of the neighbor nodes to generate the structural embedding representation of the target user under the corresponding snapshot; input the sequence of structural embedding representations corresponding to multiple snapshots arranged in chronological order within the explicit slice into a gated recurrent unit; use the hidden state of the previous snapshot and the structural embedding representation of the current snapshot as input to update its hidden state, modeling the temporal evolutionary dependency across snapshots within the slice; use the hidden state output by the gated recurrent unit after processing the last snapshot in the explicit slice as the local evolutionary representation of the corresponding explicit slice.

[0044] Specifically, a local subgraph is constructed by extracting the k-hop neighbors (i.e., multi-layered relationship networks directly or indirectly connected to the target user) centered on the target user. A relationship-aware graph attention network is used to dynamically calculate the interaction weights with different neighbors, thereby aggregating and generating a structural embedding representation reflecting the complex relationship structure in the current snapshot. Subsequently, a series of time-ordered structural embeddings are input into a gated recurrent unit, utilizing its hidden state (memory unit) to transmit and fuse historical information, thus modeling the continuous evolution of behavioral patterns within the explicit slice. Finally, the final state of this unit is output as a local evolutionary representation. This approach simultaneously captures key structural dependencies and temporal dynamics within complex graph data, consolidating the scattered, multi-relationship interaction information of users in each active period into a coherent vector that comprehensively represents the behavioral characteristics of that period, providing fine-grained input for subsequent cross-slice fusion and long- and short-term situational analysis.

[0045] Furthermore, considering the following processing steps: Assign a learnable position index to all explicit slices arranged in chronological order and generate corresponding position embedding vectors to encode the order information of each slice in the overall time series; add the local evolutionary representation of each explicit slice to the corresponding position embedding vector to generate an enhanced slice representation of the corresponding slice; generate a query vector based on the current schedule query to be predicted; calculate the semantic similarity between the query vector and the enhanced slice representation of each explicit slice as an initial score to measure the relevance of the corresponding slice to the current prediction target topic; use the enhanced slice representation of each explicit slice as the key and value, and the query vector as the query input, perform scaled dot product attention calculation, normalize the initial score, and generate adaptive attention weights corresponding to each explicit slice; based on the adaptive attention weights, perform a weighted summation of the enhanced slice representations of all explicit slices to obtain an aggregated representation vector, which serves as the short-term evolutionary representation of the target user.

[0046] Specifically, the representation of each explicit slice is enhanced by adding a location embedding vector, thus carrying temporal information. Then, a query vector (representing the current prediction intent) is generated based on the current pending schedule request, and the similarity between this vector and the representation of each enhanced slice is calculated to assess the relevance of each historical slice to the current task. Next, scaled dot product attention is used to dynamically weight all slice representations, giving higher weight to historical periods highly relevant to the current meeting schedule or task type (such as project work periods from last week), and lower weight to irrelevant periods (such as vacation periods). Finally, a short-term evolutionary representation of the user is generated through weighted aggregation. This achieves dynamic filtering and fusion of historical information, ensuring that the system can most effectively draw upon the historical experience most relevant to the current context when predicting recent user behavior.

[0047] In some embodiments, step S40 specifically includes the following steps: responding to a schedule creation request initiated by the mobile application subsystem, the natural language processing unit parses the request content and extracts the target time range, event type, and set of participating users associated with identifiers; for each user in the set of participating users, the corresponding comprehensive situational representation is obtained to form a multi-user situational representation set; based on the multi-user situational representation set, for multiple candidate time windows within the target time range, the global matching score of each candidate time window is calculated; wherein, the calculation process comprehensively evaluates the individual matching degree of each user in the corresponding window and introduces a penalty factor to avoid time conflicts with the existing confirmed schedules of each user; at least one candidate time window with the highest global matching score is selected, and a structured schedule plan is generated by combining the corresponding event type and participants, and prioritized according to the global matching score from high to low; for each generated schedule plan, based on the attention weight used when generating the schedule plan, several historical quadruplets whose contribution to the prediction result exceeds a preset threshold are located from the temporal knowledge graph, and key historical subgraphs are constructed with them as the core; this process achieves the extraction of key subgraphs by optimizing a mask matrix: ; Where M is the adjacency matrix mask to be learned; Here, is the indicator function; N is the total number of entities; y is the actual label; Let c be the adjacency matrix of the computational subgraph centered on the target entity c. The objective function is the sigmoid activation function, used to constrain the mask values ​​to the interval [0,1]; ⊙ represents element-wise multiplication. Optimizing this objective function yields a sparse mask M, whose edges corresponding to non-zero elements constitute the key history subgraph that has a critical impact on the current prediction decision.

[0048] Finally, the key historical subgraphs are aligned with the elements of the corresponding schedule options to generate natural language describing the correlation between historical patterns and current decisions, serving as a visual recommendation reason for the corresponding schedule options.

[0049] Specifically, the system first parses user requests and obtains a comprehensive situational representation of all participants. Then, for multiple candidate time windows, it automatically avoids existing schedules by comprehensively calculating the individual matching degree of each participant and introducing a penalty factor, thereby calculating a global matching degree score for each window. Next, the system selects the window with the highest score to generate a structured schedule plan and ranks it. Finally, the plan analyzes the attention weights (parameters reflecting the importance of different historical information) within the model, locates and extracts key historical subgraphs (the set of past events that have the greatest impact on the current decision), and then generates a natural language explanation clarifying the basis for the recommendation. This achieves a complete process from conflict avoidance to generating explainable recommendations, ensuring a high success rate of collaborative scheduling and decision-making transparency.

[0050] The implementation principle of a schedule management system method deeply integrated with artificial intelligence, as described in this application, is as follows: User schedule data is collected through mobile applications and terminal devices and linked to their virtual avatars to construct and continuously update a team-level temporal knowledge graph based on four-tuple-based facts. Based on this graph, the core innovation of the system lies in performing multi-scale temporal evolution analysis: A graph neural network is used to learn from a unified dense graph constructed from long-term historical data, capturing stable user habits and cyclical patterns to form a long-term evolutionary representation. Simultaneously, through natural evolutionary slicing and using a relation-aware graph attention network and gated recurrent units, the system learns the user's recent dynamics and contextual information influenced by related users to generate a short-term evolutionary representation. Subsequently, the long-term and short-term representations are adaptively fused to form a comprehensive representation that fully reflects the user's situation. When responding to a user's collaborative schedule request, the system aggregates the comprehensive situational representation of all participants, calculates the global matching degree of future candidate time windows to avoid conflicts, and generates priority-ranked schedule plans. After the plan is generated, the system automatically generates visualized recommendation reasons by tracing key historical subgraphs in the temporal knowledge graph and combining them with model attention weights. Ultimately, the user-confirmed plan will be synchronized to all relevant terminals and visually prompted using each user's virtual avatar. This achieves a fundamental shift from passive recording to proactive prediction, from simple conflict checking to understanding asynchronous rhythms, and from black-box decision-making to transparent and reliable explanation by simulating deep temporal reasoning in human cognition, significantly improving the collaborative efficiency of schedule management.

[0051] Figure 1 This is a flowchart illustrating a schedule management system method that deeply integrates artificial intelligence in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0052] Based on the same technical concept, and referring to Figure 3, this application embodiment also provides a schedule management system device that deeply integrates artificial intelligence, adopting the following technical solution: The device includes: The virtual avatar module is used to collect schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and to generate and bind a virtual avatar identifier for each user. The data association module is used to dynamically build and maintain a team-level time-series knowledge graph based on schedule-related data. Each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp. The module is updated in real time based on the user's subsequent actions on the schedule. The data analysis module is used to perform multi-scale temporal evolution analysis on target users based on temporal knowledge graphs. It uses graph neural networks to learn the long-term evolutionary representation of users and relationships to capture stable habits and periodic patterns. It uses sequence models and relationship-aware graph attention networks to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target users. The schedule recommendation module is used to respond to users' schedule creation requests. For collaborative requests involving multiple users, it obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, it generates a visual reason for recommending the schedule plan by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights inside the model. The plan confirmation module is used to receive user confirmation information and synchronize the final confirmed schedule plan to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provide visual prompts using the virtual avatars of each user.

[0053] In some embodiments, the virtual avatar module is specifically used to respond to a user's registration or avatar creation instruction in the mobile application subsystem and receive a front-facing avatar image uploaded by the user. The virtual avatar generation unit sends a frontal headshot image to the cloud, uses a face detection algorithm to locate facial feature points, extracts facial feature vectors, and generates a 2D stylized avatar based on a pre-trained style transfer model using a pre-set style template; or, it generates a 3D mesh of the user's face using a 3D face reconstruction model, generates a facial texture map by combining the style prompts selected by the user, and outputs a 3D virtual avatar with customizable clothing, expressions, and accessories through a rendering engine. The generated virtual avatar data is bound to the user's unique system account identifier and stored in a pre-set cloud user profile database; The bound virtual avatar data is synchronized in real time via an encrypted network to the mobile application subsystem and schedule management terminal subsystem associated with the user's account, serving as a unified visual identifier for the corresponding user across the entire system platform.

[0054] In some embodiments, the data association module is specifically used to receive schedule-related data streams from the mobile application subsystem and the schedule management terminal subsystem. The schedule-related data streams include at least new schedule creation, existing schedule modification, schedule status update, and schedule deletion operations and their corresponding time information. The data stream related to the schedule is parsed, the core elements corresponding to each operation are extracted, and they are structured and represented in a four-tuple format. The subject user is the unique identifier of the user who initiates the operation or is the subject of the schedule. The relationship is a semantic tag describing the type of operation or schedule, including creation, modification, completion, and participation. The object is the target to which the operation or schedule is directed, and its type includes other user identifiers, specific task items, and location information. The timestamp is the precise time point or time period of the operation execution or schedule arrangement. The quadruples are stored as new facts in a cloud-based time-series knowledge graph database, and their relationship with historical quadruples is established in the time dimension to form a graph structure that describes the dynamic evolution of the team. Based on the user's subsequent actions on the schedule, the real-time incremental update mechanism of the graph is triggered: when the schedule-related data stream is received, the changed data part is identified, only the changed related quadruples and their associated edges are updated, and the update instructions are pushed to all relevant clients in real time through the subscription message protocol. Through an incremental update mechanism, the time-series knowledge graph continues to evolve and serves as the sole authoritative data source for subsequent multi-scale time-series evolution analysis.

[0055] In some embodiments, the data analysis module is specifically used to extract all historical interaction quadruples of the target user within a long historical time window from the time-series knowledge graph, and fuse them to construct a unified dense knowledge graph, in which all historically appearing entities are retained as nodes, and the relationships between entities across timestamps are fused into the edges of the graph. A relational graph convolutional network is used as a semantic aggregator to perform multi-round message passing and feature aggregation on a unified dense knowledge graph, and to learn the embedding representation of target users and various relationships at a long-term scale as a long-term evolutionary representation. From the temporal knowledge graph, a series of continuous knowledge graph snapshots of the target user within a short historical time window are extracted. Based on the occurrence pattern of the target entity, the knowledge graph snapshot sequence is naturally evolved into slices, which are divided into explicit slices containing the target entity and implicit slices not containing the target entity. For each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies; and a gated recurrent unit is used to capture the temporal dependencies between different snapshots within the slice to generate the local evolutionary representation of the corresponding slice. By introducing a cross-slice attention mechanism with location embedding, the local evolutionary representations of all explicit slices are adaptively weighted and aggregated. The attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, thereby generating a short-term evolutionary representation of the target user. By using learnable weighted parameters, the long-term evolutionary representation and the short-term evolutionary representation are weighted and summed, and then passed through a nonlinear transformation layer to obtain the comprehensive situational representation of the target user.

[0056] In some embodiments, the data analysis module is specifically used to extract the k-hop neighbor entities and their relationships for each knowledge graph snapshot within an explicit slice, with the target user entity as the central node, and construct a locally connected subgraph, where k is a natural number greater than 0. A relation-aware graph attention network is used to encode the locally connected subgraphs. By calculating the attention weights of the target user node and its neighboring nodes under specific relationships, the features and relational semantic information of the neighboring nodes are adaptively aggregated to generate the structural embedding representation of the target user under the corresponding snapshot. The structure embedding representation sequence corresponding to multiple snapshots arranged in chronological order within an explicit slice is input into a gated recurrent unit; the hidden state of the previous snapshot and the structure embedding representation of the current snapshot are used as input to update its hidden state, thus modeling the temporal evolution dependency across snapshots within the slice. The hidden state output by the gated recurrent unit after processing the last snapshot in the explicit slice is used as the local evolution representation of the corresponding explicit slice.

[0057] In some embodiments, the data analysis module is specifically used to assign a learnable position index to all explicit slices arranged in chronological order and generate corresponding position embedding vectors to encode the order information of each slice in the overall time series; The local evolutionary representation of each explicit slice is added to the corresponding position embedding vector to generate the enhanced slice representation of the corresponding slice; Based on the current schedule query to be predicted, a query vector is generated; the semantic similarity between the query vector and the augmented slice representation of each explicit slice is calculated as an initial score to measure the relevance of the corresponding slice to the current prediction target topic. The enhanced slice representation of each explicit slice is used as the key and value, the query vector is used as the query input, the scaled dot product attention calculation is performed, the initial score is normalized, and the adaptive attention weights corresponding to each explicit slice are generated. Based on the adaptive attention weights, the enhanced slice representations of all explicit slices are weighted and summed to obtain the aggregated representation vector, which serves as the short-term evolutionary representation of the target user.

[0058] In some embodiments, the schedule recommendation module is specifically used to respond to a schedule creation request initiated by the mobile application subsystem, and the natural language processing unit parses the request content to extract the target time range, event type, and set of participating users associated with identifiers; For each user in the participating user set, obtain the corresponding comprehensive situational representation to form a multi-user situational representation set; Based on a multi-user situational representation set, a global matching score is calculated for each candidate time window within a target time range. The calculation process comprehensively evaluates the individual matching degree of each user in the corresponding window and introduces a penalty factor to avoid time conflicts with each user's existing confirmed schedule. Select at least one candidate time window with the highest global matching score, combine it with the corresponding event type and participants, generate a structured schedule plan, and sort it by priority from high to low according to the global matching score; For each generated schedule plan, based on the attention weights used when generating the schedule plan, locate several historical quadruplets from the time-series knowledge graph that contribute more than a preset threshold to the prediction results, and construct a key historical subgraph with these quadruplets as the core. By aligning key historical subgraphs with the elements of corresponding schedule plans, natural language is generated to describe the correlation between historical patterns and current decisions, serving as a visual recommendation for the corresponding schedule plans.

[0059] This application also discloses a control device.

[0060] Specifically, the control device includes a memory and a processor, the memory storing a computer program that can be loaded by the processor and executed as described above for a deeply integrated artificial intelligence-based schedule management system.

[0061] This application also discloses a computer-readable storage medium.

[0062] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the aforementioned schedule management system method that deeply integrates artificial intelligence. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A schedule management system method deeply integrated with artificial intelligence, characterized in that, include: The system collects schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and generates and binds a virtual avatar for each user. Based on schedule-related data, a team-level time-series knowledge graph is dynamically constructed and maintained, where each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp, and is updated in real time based on the user's subsequent actions on the schedule. Based on the aforementioned temporal knowledge graph, multi-scale temporal evolution analysis is performed on the target user. Graph neural networks are used to learn the long-term evolutionary representation of users and relationships to capture stable habits and cyclical patterns. Sequence models and relationship-aware graph attention networks are used to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target user. In response to a user's schedule creation request, for collaborative requests involving multiple users, the system obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights within the model, the system generates a visual reason for recommending the schedule plan. Receive user confirmation information and synchronize the final confirmed schedule to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provide visual prompts using each user's virtual avatar as an identifier.

2. The method for a schedule management system deeply integrated with artificial intelligence according to claim 1, characterized in that, The process involves collecting schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and generating and binding a virtual avatar identifier for each user, including: Respond to user registration or avatar creation instructions in the mobile application subsystem and receive front-facing avatar images uploaded by users; The virtual avatar generation unit sends a frontal headshot image to the cloud, uses a face detection algorithm to locate facial feature points, extracts facial feature vectors, and generates a 2D stylized avatar based on a pre-trained style transfer model using a pre-set style template; or, it generates a 3D mesh of the user's face using a 3D face reconstruction model, generates a facial texture map by combining the style prompts selected by the user, and outputs a 3D virtual avatar with customizable clothing, expressions, and accessories through a rendering engine. The generated virtual avatar data is bound to the user's unique system account identifier and stored in a pre-set cloud user profile database; The bound virtual avatar data is synchronized in real time via an encrypted network to the mobile application subsystem and schedule management terminal subsystem associated with the user's account, serving as a unified visual identifier for the corresponding user across the entire system platform.

3. The method for a schedule management system deeply integrated with artificial intelligence according to claim 1, characterized in that, The process of dynamically constructing and maintaining a team-level time-series knowledge graph based on schedule-related data includes: Receive schedule-related data streams from the mobile application subsystem and the schedule management terminal subsystem. The schedule-related data streams include at least new schedule creation, existing schedule modification, schedule status update and schedule deletion operations and their corresponding time information. The data stream related to the schedule is parsed, the core elements corresponding to each operation are extracted, and they are structured and represented in a four-tuple format. The subject user is the unique identifier of the user who initiates the operation or is the subject of the schedule. The relationship is a semantic tag describing the type of operation or schedule, including creation, modification, completion, and participation. The object is the target to which the operation or schedule is directed, and its type includes other user identifiers, specific task items, and location information. The timestamp is the precise time point or time period of the operation execution or schedule arrangement. The quadruples are stored as new facts in a cloud-based time-series knowledge graph database, and their relationship with historical quadruples is established in the time dimension to form a graph structure that describes the dynamic evolution of the team. Based on the user's subsequent actions on the schedule, the real-time incremental update mechanism of the graph is triggered: when the schedule-related data stream is received, the changed data part is identified, only the changed related quadruples and their associated edges are updated, and the update instructions are pushed to all relevant clients in real time through the subscription message protocol. Through an incremental update mechanism, the time-series knowledge graph continuously evolves and serves as the sole authoritative data source for subsequent multi-scale time-series evolution analysis.

4. The method for a schedule management system deeply integrated with artificial intelligence according to claim 1, characterized in that, Based on the time-series knowledge graph, multi-scale time-series evolution analysis is performed on the target user, and graph neural networks are used to learn the long-term evolutionary representation of users and relationships to capture stable habits and periodic patterns. Sequence models and relationship-aware graph attention networks are used to learn their short-term evolutionary representation to capture recent changes and contextual influences. The long-term evolutionary representation and the short-term evolutionary representation are adaptively weighted and fused to generate a comprehensive situational representation of the target user, including: From the time-series knowledge graph, extract all historical interaction quadruples of the target user within a long historical time window, and merge them to construct a unified dense knowledge graph, in which all historically appearing entities are retained as nodes, and the relationships between entities across timestamps are merged into the edges of the graph; A relational graph convolutional network is used as a semantic aggregator to perform multi-round message passing and feature aggregation on the unified dense knowledge graph, and to learn the embedding representation of target users and various relationships at a long-term scale as a long-term evolutionary representation. From the time-series knowledge graph, a series of continuous knowledge graph snapshots of the target user within a short historical time window are extracted, and the knowledge graph snapshot sequence is naturally evolved and sliced ​​based on the occurrence pattern of the target entity, dividing it into explicit slices containing the target entity and implicit slices not containing the target entity. For each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies; and a gated recurrent unit is used to capture the temporal dependencies between different snapshots within the slice to generate the local evolutionary representation of the corresponding slice. By introducing a cross-slice attention mechanism with location embedding, the local evolutionary representations of all explicit slices are adaptively weighted and aggregated, wherein the attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, thereby generating a short-term evolutionary representation of the target user. By using learnable weighted parameters, the long-term evolutionary representation and the short-term evolutionary representation are weighted and summed, and then passed through a nonlinear transformation layer to obtain the comprehensive situational representation of the target user.

5. The method for a schedule management system deeply integrated with artificial intelligence according to claim 4, characterized in that, For each explicit slice, a relation-aware graph attention network is used to encode the subgraph structure in each snapshot within the slice and extract structural dependencies. Furthermore, a gated recurrent unit is used to capture the temporal dependencies between different snapshots within a slice, generating a local evolutionary representation of the corresponding slice, including: For each knowledge graph snapshot within an explicit slice, with the target user entity as the central node, extract its k-hop neighbor entities and their relationships, and construct a locally connected subgraph, where k is a natural number greater than 0; A relation-aware graph attention network is used to encode the local connected subgraph; by calculating the attention weights of the target user node and its neighboring nodes under a specific relationship, the features and relational semantic information of the neighboring nodes are adaptively aggregated to generate a structural embedding representation of the target user under the corresponding snapshot. The structure embedding representation sequence corresponding to multiple snapshots arranged in chronological order within an explicit slice is input into a gated recurrent unit; the hidden state of the previous snapshot and the structure embedding representation of the current snapshot are used as input to update its hidden state, thus modeling the temporal evolution dependency across snapshots within the slice. The hidden state output by the gated loop unit after processing the last snapshot in the explicit slice is used as the local evolution representation of the corresponding explicit slice.

6. The method for a schedule management system deeply integrated with artificial intelligence according to claim 4, characterized in that, The method involves introducing a cross-slice attention mechanism with location embedding to adaptively weight and aggregate the local evolutionary representations of all explicit slices. The attention weights are dynamically calculated based on the topic relevance of each slice to the current prediction target, generating a short-term evolutionary representation of the target user, including: Assign a learnable position index to all explicit slices arranged in chronological order and generate corresponding position embedding vectors to encode the order information of each slice in the overall time series; The local evolution representation of each explicit slice is added to the corresponding position embedding vector to generate the enhanced slice representation of the corresponding slice; Based on the current schedule query to be predicted, a query vector is generated; the semantic similarity between the query vector and the enhanced slice representation of each explicit slice is calculated as an initial score to measure the relevance of the corresponding slice to the current prediction target topic. The enhanced slice representation of each explicit slice is used as the key and value, the query vector is used as the query input, the scaled dot product attention calculation is performed, the initial score is normalized, and the adaptive attention weights corresponding to each explicit slice are generated. Based on the adaptive attention weights, the enhanced slice representations of all explicit slices are weighted and summed to obtain the aggregated representation vector, which serves as the short-term evolutionary representation of the target user.

7. The method for a schedule management system deeply integrated with artificial intelligence according to claim 1, characterized in that, The system responds to user schedule creation requests. For collaborative requests involving multiple users, it obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule scheme. Based on the scheduling scheme, by analyzing the key historical subgraphs in the time-series knowledge graph and the internal attention weights of the model, a visualized recommendation rationale for the scheduling scheme is generated, including: In response to a schedule creation request initiated by the mobile application subsystem, the natural language processing unit parses the request content and extracts the target time range, event type, and set of participating users associated with identifiers; For each user in the participating user set, the corresponding comprehensive situational representation is obtained to form a multi-user situational representation set; Based on the multi-user situational representation set, for multiple candidate time windows within the target time range, a global matching score for each candidate time window is calculated; wherein, the calculation process comprehensively evaluates the individual matching degree of each user in the corresponding window, and introduces a penalty factor to avoid time conflicts with each user's existing confirmed schedule. Select at least one candidate time window with the highest global matching score, combine it with the corresponding event type and participants, generate a structured schedule, and sort it by priority from high to low according to the global matching score; For each generated schedule plan, based on the attention weight used when generating the schedule plan, several historical quadruplets that contribute more than a preset threshold to the prediction result are located from the time-series knowledge graph, and key historical subgraphs are constructed with these quadruplets as the core. Alignment analysis is performed between the key historical subgraphs and the elements of the corresponding schedule plan to generate natural language describing the correlation between historical patterns and current decisions, serving as a visual recommendation reason for the corresponding schedule plan.

8. A schedule management system device deeply integrated with artificial intelligence, characterized in that, The device includes: The virtual avatar module is used to collect schedule-related data from multiple users through the mobile application subsystem and the schedule management terminal subsystem, and to generate and bind a virtual avatar identifier for each user. The data association module is used to dynamically build and maintain a team-level time-series knowledge graph based on schedule-related data. Each fact is represented as a quadruple, which includes: subject user, relationship, object, and timestamp. The module is updated in real time based on the user's subsequent actions on the schedule. The data analysis module is used to perform multi-scale temporal evolution analysis on target users based on the temporal knowledge graph, and to learn the long-term evolutionary representation of users and relationships using graph neural networks to capture stable habits and periodic patterns. It also uses sequence models and relationship-aware graph attention networks to learn the short-term evolutionary representation to capture recent changes and contextual influences. The long-term and short-term evolutionary representations are adaptively weighted and fused to generate a comprehensive situational representation of the target users. The schedule recommendation module is used to respond to users' schedule creation requests. For collaborative requests involving multiple users, it obtains the comprehensive situational representation of all participating users, calculates the global matching degree of future candidate time windows, automatically avoids conflicts, and generates at least one priority-ranked schedule plan. Based on the schedule plan, by analyzing the key historical subgraphs in the time-series knowledge graph and the attention weights inside the model, it generates a visual reason for recommending the schedule plan. The plan confirmation module is used to receive user confirmation information and synchronize the final confirmed schedule plan to the mobile application subsystem and schedule management terminal subsystem of all relevant users, and provide visual prompts using the virtual avatars of each user.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images