Adaptive recommendation and feedback optimization methods and systems for power business scenarios
By constructing multi-dimensional scene element tuples and dynamic scene state vectors, and combining user feedback to optimize the recommendation system for power business scenarios, the problem of recommendation results being disconnected from user tasks in existing technologies has been solved, achieving efficient closed-loop optimization with real-time response and adaptive learning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING SGITG ACCENTURE INFORMATION TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies cannot deeply analyze and utilize the complex business scenario information contained in user queries, resulting in recommendation results being disconnected from the user's current work tasks, lacking real-time interactive intent perception and feedback response capabilities, and failing to achieve adaptive learning and accuracy improvement.
By receiving the structured intent parsing results, a multi-dimensional scene element tuple is constructed and a dynamic scene state vector is generated. Combined with the user's long-term interest vector and candidate standard features, a hybrid deep recommendation network is used for fusion calculation to generate an initial recommendation list. Based on the interaction feedback sequence, real-time learning and optimization are performed to form a closed-loop feedback mechanism.
It achieves deep representation and precise local perception of power industry scenarios. The recommendation system can respond to user intent in real time, forming an efficient closed loop of scenario analysis, intelligent recommendation and feedback optimization, and improving the accuracy and adaptability of single interaction.
Smart Images
Figure CN122087183A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power information technology, specifically to an adaptive recommendation and feedback optimization method and system for power business scenarios. Background Technology
[0002] In the planning, design, operation, maintenance, and compliance review processes of the power industry, practitioners frequently need to consult a vast number of technical standards, regulations, and specifications. While keyword-based or semantic search technologies can help users find directly relevant standards, when faced with complex and comprehensive business tasks, users often have an implicit need to "not know what to look up." For example, when undertaking a "city distribution network automation upgrade" project, designers not only need to know the direct equipment standards but also a series of related standards such as design specifications, safety procedures, and acceptance requirements.
[0003] Existing technologies cannot deeply analyze and utilize the complex business scenario information contained in user queries, resulting in a disconnect between recommendation results and the user's current specific work tasks, leading to poor scenario adaptability. At the same time, existing methods lack the ability to perceive and respond to users' real-time interactive intentions, and cannot dynamically adjust recommended content based on feedback in a single session. Finally, due to the rigidity of existing recommendation models, they fail to deeply associate user behavior feedback with specific business contexts and form a closed loop, resulting in the system's inability to achieve rapid adaptive learning and accuracy improvement. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, this invention provides an adaptive recommendation and feedback optimization method for power business scenarios, comprising: Receive the structured intent parsing results for the current session; Based on the business scenario parameters in the structured intent parsing results, a multi-dimensional scenario element tuple is extracted and constructed. Based on the multi-dimensional scenario element tuple, the associated information is retrieved from the pre-constructed business scenario-standard association graph. Based on the associated information, a dynamic scenario state vector is generated. The dynamic scene state vector, the user's long-term interest vector obtained from the user profile database, and the candidate standard features related to the current session obtained from the standard feature database are input into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; the interaction feedback sequence generated for the initial recommendation list is obtained. Based on the core intent categories in the interactive feedback sequence and structured intent parsing results, perform real-time feedback learning, dynamically optimize the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generate and output the optimized recommendation list.
[0005] Preferably, based on the business scenario parameters in the structured intent parsing results, a multi-dimensional scenario element tuple is extracted and constructed. Relationship information is retrieved from a pre-constructed business scenario-standard association graph based on the multi-dimensional scenario element tuple. A dynamic scenario state vector is generated based on the relationship information, including: The business scenario parameters are analyzed to identify and extract at least three elements from the core tasks, equipment objects, project stages, and compliance requirements, and a structured multidimensional scenario element tuple is constructed. The elements in the multidimensional element tuple are used as query nodes. Multi-hop traversal and subgraph retrieval are performed in the business scenario-standard association graph. The relevant standard entities, technical term entities and their relationship paths are used as association information. The multidimensional scene element tuples and associated information are input into the scene encoder. The scene encoder then uses an attention mechanism to fuse and compress the input multidimensional scene element tuples and associated information, and outputs a fixed-dimensional dynamic scene state vector.
[0006] Preferably, the dynamic scene state vector, the user's long-term interest vector, and candidate standard features are input into a hybrid deep recommendation network for fusion calculation to generate an initial recommendation list, including: A multi-channel input layer is constructed for a hybrid deep recommendation network. The first channel of the multi-channel input layer is the dynamic scene state vector, the second channel is the user's long-term interest vector, and the third channel is the candidate standard features. Among them, the candidate standard features are multimodal features, including the text semantic vector and metadata of the candidate standards. By using deep cross layers in the network, attention interaction is performed between the dynamic scene state vector and the text semantic vector to calculate the scene-semantic matching score; feature cross is performed between the user's long-term interest vector and metadata to calculate the personalization-attribute matching score. The scene-semantic matching score and the personalization-attribute matching score are weighted and fused to generate the final recommendation score for each candidate criterion; All candidate criteria are sorted in descending order based on the final recommendation score to generate an initial recommendation list.
[0007] Preferably, the interaction feedback sequence generated for the initial recommendation list is obtained, including: Monitor and record the exposure, clicks, ignores, dwell time, and final adoption behavior of each candidate criterion in the initial recommendation list within the current session; Encode behaviors in chronological order into behavior sequences with timestamps and intensity weights to construct interaction feedback sequences; The interaction feedback sequence is associated with and stored in relation to the business scenario parameters and core intent types for the current session.
[0008] Preferably, based on the core intent categories in the interactive feedback sequence and structured intent parsing results, real-time feedback learning is performed. Based on the real-time feedback learning results, the order and / or content of the initial recommendation list are dynamically optimized, and an optimized recommendation list is generated and output, including: Based on the core intent category, the corresponding on-the-spot learning algorithm is selected from the preset policy mapping relationship; whereby the policy mapping relationship defines the correspondence between different core intent categories and different on-the-spot learning algorithms; The interactive feedback sequence is input into the selected real-time learning algorithm to generate real-time adjusted weights for each item in the initial recommendation list; The original scores of each item in the initial recommendation list are recalculated by adjusting the weights in real time, and / or the order of the list is adjusted to generate an optimized recommendation list.
[0009] Preferably, after dynamically optimizing the order and / or content of the initial recommendation list based on the learning results of real-time feedback, and generating and outputting the optimized recommendation list, the method further includes: Store the interaction feedback sequence and its corresponding scene context in the feedback knowledge base; When the preset model update conditions are met, long-term feedback learning is performed based on historical data stored in the feedback knowledge base, which consists of interactive feedback sequences and corresponding scene contexts, to generate model update parameters. The model update parameters are used to iteratively update the hybrid deep recommendation network and the user's long-term interest vector; The iteratively updated hybrid deep recommendation network and user long-term interest vectors are used to perform fusion calculations to generate a new initial recommendation list.
[0010] Preferably, the preset model update conditions include at least one of the following: reaching a preset time period, the feedback data accumulated in the feedback knowledge base reaching a preset threshold, and monitoring that the average user satisfaction index of the optimized recommendation list is continuously lower than a preset threshold. Long-term feedback learning is performed based on historical data stored in the feedback knowledge base, consisting of interaction feedback sequences and corresponding scene contexts, to generate model update parameters, including: An incremental learning algorithm is used to calculate the parameter gradient of the hybrid deep recommendation network and the update amount of the user's long-term interest vector as model update parameters, using historical data as training samples. The incremental learning algorithm includes minimizing the loss function and regularization constraints that incorporate a knowledge distillation mechanism.
[0011] Based on the same inventive concept, this invention also provides an adaptive recommendation and feedback optimization system for power business scenarios, comprising: The intent parsing receiving module is used to receive the structured intent parsing results for the current session; The state vector generation module is used to extract and construct multi-dimensional scene element tuples based on business scene parameters in the structured intent parsing results, retrieve association information from the pre-built business scene-standard association graph based on the multi-dimensional scene element tuples, and generate dynamic scene state vectors based on the association information. The initial list generation module is used to input the dynamic scene state vector, the user's long-term interest vector obtained from the user profile library, and the candidate standard features related to the current session obtained from the standard feature library into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; and to obtain the interaction feedback sequence generated for the initial recommendation list. The optimized list generation module is used to perform real-time feedback learning based on the core intent categories in the interactive feedback sequence and structured intent parsing results. Based on the real-time feedback learning results, it dynamically optimizes the order and / or content of the initial recommendation list, and generates and outputs the optimized recommendation list.
[0012] Preferably, the state vector generation module is specifically used for: The business scenario parameters are analyzed to identify and extract at least three elements from the core tasks, equipment objects, project stages, and compliance requirements, and a structured multidimensional scenario element tuple is constructed. The elements in the multidimensional element tuple are used as query nodes. Multi-hop traversal and subgraph retrieval are performed in the business scenario-standard association graph. The relevant standard entities, technical term entities and their relationship paths are used as association information. The multidimensional scene element tuples and associated information are input into the scene encoder. The scene encoder then uses an attention mechanism to fuse and compress the input multidimensional scene element tuples and associated information, and outputs a fixed-dimensional dynamic scene state vector.
[0013] Preferably, the initial list generation module is specifically used for: A multi-channel input layer is constructed for a hybrid deep recommendation network. The first channel of the multi-channel input layer is the dynamic scene state vector, the second channel is the user's long-term interest vector, and the third channel is the candidate standard features. Among them, the candidate standard features are multimodal features, including the text semantic vector and metadata of the candidate standards. By using deep cross layers in the network, attention interaction is performed between the dynamic scene state vector and the text semantic vector to calculate the scene-semantic matching score; feature cross is performed between the user's long-term interest vector and metadata to calculate the personalization-attribute matching score. The scene-semantic matching score and the personalization-attribute matching score are weighted and fused to generate the final recommendation score for each candidate criterion; All candidate criteria are sorted in descending order based on the final recommendation score to generate an initial recommendation list.
[0014] Preferably, the initial list generation module is specifically used for: Monitor and record the exposure, clicks, ignores, dwell time, and final adoption behavior of each candidate criterion in the initial recommendation list within the current session; Encode behaviors in chronological order into behavior sequences with timestamps and intensity weights to construct interaction feedback sequences; The interaction feedback sequence is associated with and stored in relation to the business scenario parameters and core intent types for the current session.
[0015] Preferably, the optimized list generation module is specifically used for: Based on the core intent category, the corresponding on-the-spot learning algorithm is selected from the preset policy mapping relationship; whereby the policy mapping relationship defines the correspondence between different core intent categories and different on-the-spot learning algorithms; The interactive feedback sequence is input into the selected real-time learning algorithm to generate real-time adjusted weights for each item in the initial recommendation list; The original scores of each item in the initial recommendation list are recalculated by adjusting the weights in real time, and / or the order of the list is adjusted to generate an optimized recommendation list.
[0016] Preferably, the system also includes a module update module for: Store the interaction feedback sequence and its corresponding scene context in the feedback knowledge base; When the preset model update conditions are met, long-term feedback learning is performed based on historical data stored in the feedback knowledge base, which consists of interactive feedback sequences and corresponding scene contexts, to generate model update parameters. The model update parameters are used to iteratively update the hybrid deep recommendation network and the user's long-term interest vector; The iteratively updated hybrid deep recommendation network and user long-term interest vectors are used to perform fusion calculations to generate a new initial recommendation list.
[0017] Preferably, the preset model update conditions include at least one of the following: reaching a preset time period, the feedback data accumulated in the feedback knowledge base reaching a preset threshold, and monitoring that the average user satisfaction index of the optimized recommendation list is continuously lower than a preset threshold. The module update module is specifically used for: An incremental learning algorithm is used to calculate the parameter gradient of the hybrid deep recommendation network and the update amount of the user's long-term interest vector as model update parameters, using historical data as training samples. The incremental learning algorithm includes minimizing the loss function and regularization constraints that incorporate a knowledge distillation mechanism.
[0018] Based on the same inventive concept, the present invention also provides an electronic device, comprising: at least one processor and a memory; wherein the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, an adaptive recommendation and feedback optimization method for power business scenarios, as described above, is implemented.
[0019] Based on the same inventive concept, the present invention also provides a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the adaptive recommendation and feedback optimization method for power business scenarios as described above.
[0020] Compared with the closest existing technology, the present invention has the following beneficial effects: This invention provides an adaptive recommendation and feedback optimization method for power business scenarios, comprising: receiving structured intent parsing results for the current session; extracting and constructing multi-dimensional scenario element tuples based on business scenario parameters in the structured intent parsing results, retrieving association information from a pre-constructed business scenario-standard association graph based on the multi-dimensional scenario element tuples, and generating a dynamic scenario state vector based on the association information; inputting the dynamic scenario state vector, user long-term interest vector obtained from a user profile database, and candidate standard features related to the current session obtained from a standard feature database into a hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; obtaining an interaction feedback sequence generated for the initial recommendation list; performing real-time feedback learning based on the interaction feedback sequence and the core intent categories in the structured intent parsing results, dynamically optimizing the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generating and outputting an optimized recommendation list. This invention constructs a process of dynamic scene perception and real-time intent optimization, extracting multi-dimensional elements from business scene parameters and generating dynamic scene state vectors. This achieves a deep and structured representation of power industry scenarios, and enhances the local accuracy of scene perception and understanding capabilities of the recommendation system. By learning by combining real-time user interaction feedback sequences with core intent types, the order and content of the current recommendation list are dynamically optimized, enabling the recommendation results to respond to user intent in real time. The entire process forms an efficient closed loop from scene analysis, intelligent recommendation, feedback collection to real-time optimization, improving the accuracy of single interactions and scene adaptability. Attached Figure Description
[0021] Figure 1 This invention provides a flowchart illustrating an adaptive recommendation and feedback optimization method for power business scenarios. Figure 2 The present invention provides a structural diagram of an adaptive recommendation and feedback optimization system for power business scenarios; Figure 3 A schematic diagram of the electronic device provided by the present invention. Detailed Implementation
[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0023] Example 1: This invention provides an adaptive recommendation and feedback optimization method for power business scenarios. Specifically, Figure 1 The flowchart of the adaptive recommendation and feedback optimization method for power business scenarios provided in this embodiment of the invention is shown in the figure, and includes the following steps: S101: Receive the structured intent parsing result for the current session; S102: Based on the business scenario parameters in the structured intent parsing results, extract and construct multi-dimensional scenario element tuples, retrieve the associated information from the pre-constructed business scenario-standard association graph based on the multi-dimensional scenario element tuples, and generate a dynamic scenario state vector based on the associated information. S103: Input the dynamic scene state vector, the user's long-term interest vector obtained from the user profile database, and the candidate standard features related to the current session obtained from the standard feature database into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; obtain the interaction feedback sequence generated for the initial recommendation list; S104: Based on the core intent categories in the interactive feedback sequence and structured intent parsing results, perform real-time feedback learning, dynamically optimize the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generate and output the optimized recommendation list.
[0024] This invention constructs a process of dynamic scene perception and real-time intent optimization, extracting multi-dimensional elements from business scene parameters and generating dynamic scene state vectors. This achieves a deep and structured representation of power industry scenarios, and enhances the local accuracy of scene perception and understanding capabilities of the recommendation system. By learning by combining real-time user interaction feedback sequences with core intent types, the order and content of the current recommendation list are dynamically optimized, enabling the recommendation results to respond to user intent in real time. The entire process forms an efficient closed loop from scene analysis, intelligent recommendation, feedback collection to real-time optimization, improving the accuracy of single interactions and scene adaptability.
[0025] Preferably, based on the business scenario parameters in the structured intent parsing results, a multi-dimensional scenario element tuple is extracted and constructed. Relationship information is retrieved from a pre-constructed business scenario-standard association graph based on the multi-dimensional scenario element tuple. A dynamic scenario state vector is generated based on the relationship information, including: The business scenario parameters are analyzed to identify and extract at least three elements from the core tasks, equipment objects, project stages, and compliance requirements, and a structured multidimensional scenario element tuple is constructed. The elements in the multidimensional element tuple are used as query nodes. Multi-hop traversal and subgraph retrieval are performed in the business scenario-standard association graph. The relevant standard entities, technical term entities and their relationship paths are used as association information. The multidimensional scene element tuples and associated information are input into the scene encoder. The scene encoder then uses an attention mechanism to fuse and compress the input multidimensional scene element tuples and associated information, and outputs a fixed-dimensional dynamic scene state vector.
[0026] Preferably, the dynamic scene state vector, the user's long-term interest vector, and candidate standard features are input into a hybrid deep recommendation network for fusion calculation to generate an initial recommendation list, including: A multi-channel input layer is constructed for a hybrid deep recommendation network. The first channel of the multi-channel input layer is the dynamic scene state vector, the second channel is the user's long-term interest vector, and the third channel is the candidate standard features. Among them, the candidate standard features are multimodal features, including the text semantic vector and metadata of the candidate standards. By using deep cross layers in the network, attention interaction is performed between the dynamic scene state vector and the text semantic vector to calculate the scene-semantic matching score; feature cross is performed between the user's long-term interest vector and metadata to calculate the personalization-attribute matching score. The scene-semantic matching score and the personalization-attribute matching score are weighted and fused to generate the final recommendation score for each candidate criterion; All candidate criteria are sorted in descending order based on the final recommendation score to generate an initial recommendation list.
[0027] Preferably, the interaction feedback sequence generated for the initial recommendation list is obtained, including: Monitor and record the exposure, clicks, ignores, dwell time, and final adoption behavior of each candidate criterion in the initial recommendation list within the current session; Encode behaviors in chronological order into behavior sequences with timestamps and intensity weights to construct interaction feedback sequences; The interaction feedback sequence is associated with and stored in relation to the business scenario parameters and core intent types for the current session.
[0028] Preferably, based on the core intent categories in the interactive feedback sequence and structured intent parsing results, real-time feedback learning is performed. Based on the real-time feedback learning results, the order and / or content of the initial recommendation list are dynamically optimized, and an optimized recommendation list is generated and output, including: Based on the core intent category, the corresponding on-the-spot learning algorithm is selected from the preset policy mapping relationship; whereby the policy mapping relationship defines the correspondence between different core intent categories and different on-the-spot learning algorithms; The interactive feedback sequence is input into the selected real-time learning algorithm to generate real-time adjusted weights for each item in the initial recommendation list; The original scores of each item in the initial recommendation list are recalculated by adjusting the weights in real time, and / or the order of the list is adjusted to generate an optimized recommendation list.
[0029] Preferably, after dynamically optimizing the order and / or content of the initial recommendation list based on the learning results of real-time feedback, and generating and outputting the optimized recommendation list, the method further includes: Store the interaction feedback sequence and its corresponding scene context in the feedback knowledge base; When the preset model update conditions are met, long-term feedback learning is performed based on historical data stored in the feedback knowledge base, which consists of interactive feedback sequences and corresponding scene contexts, to generate model update parameters. The model update parameters are used to iteratively update the hybrid deep recommendation network and the user's long-term interest vector; The iteratively updated hybrid deep recommendation network and user long-term interest vectors are used to perform fusion calculations to generate a new initial recommendation list.
[0030] Preferably, the preset model update conditions include at least one of the following: reaching a preset time period, the feedback data accumulated in the feedback knowledge base reaching a preset threshold, and monitoring that the average user satisfaction index of the optimized recommendation list is continuously lower than a preset threshold. Long-term feedback learning is performed based on historical data stored in the feedback knowledge base, consisting of interaction feedback sequences and corresponding scene contexts, to generate model update parameters, including: An incremental learning algorithm is used to calculate the parameter gradient of the hybrid deep recommendation network and the update amount of the user's long-term interest vector as model update parameters, using historical data as training samples. The incremental learning algorithm includes minimizing the loss function and regularization constraints that incorporate a knowledge distillation mechanism.
[0031] The adaptive recommendation and feedback optimization method for power business scenarios provided by this invention will be described in detail below, taking the "wind farm access system design" scenario as an example.
[0032] 1. Reception and Parsing: The user inputs a natural language request: "What are the specific requirements for reactive power regulation and power factor in the design of a wind farm grid connection system?". The upstream intent parsing module (which can be an independent parser based on a large language model) analyzes the request and generates a structured intent parsing result, such as: {"Core Intent Category": "Comparative Analysis", "Business Scenario Parameters", "New Energy Power Station Grid Connection Design", ...}. The system receives this result.
[0033] 2. Scene awareness and vectorization: Extracting elements: Based on "business scenario parameters" (new energy power station access design), the system analyzes and constructs multi-dimensional scenario element tuples, such as: (core task: "system access design"; equipment object: "wind farm / gathering station"; project stage: "preliminary design"; compliance requirements: "meet the power system safety and stability guidelines").
[0034] Graph Retrieval: Using this tuple as the query key, search within the pre-built business scenario-standard association graph to obtain related information, such as partial subgraphs related to the standards "GB / T19963 Technical Regulations for Wind Farm Access to Power Systems" and "NB / T32004 Technical Specifications for Reactive Power Compensation of Photovoltaic Power Stations" and their related clauses and technical terms.
[0035] Generate vector: Input the feature tuples and associated subgraphs into the scene encoder (e.g., based on graph attention network) to generate a fixed-dimensional dynamic scene state vector that comprehensively represents the current complex scene.
[0036] 3. Initial Recommendation Generation: Feature Acquisition: The system reads the user's long-term interest vector from the user profile database (e.g., the user has historically paid more attention to standards related to "relay protection" and "power quality"). At the same time, it obtains a batch of relevant candidate standard features from the standard feature database. Each feature contains a text semantic vector (encoded from the standard content) and metadata (e.g., {standard number: "GB / T19963", type: "national standard", year: 2011}).
[0037] Network computation: Input dynamic scene state vectors, user long-term interest vectors, and candidate standard features into a hybrid deep recommendation network.
[0038] Internal network: Calculate the scene-semantic matching score: The attention mechanism measures the relevance between the dynamic scene state vector and the text semantic vector of each candidate standard.
[0039] Calculate the personalized-attribute matching score: Feature cross measures the degree of matching between the user's long-term interest vector and the metadata of each candidate standard (such as the user's preference for "national standard").
[0040] Fusion and ranking: The semantic matching score and the personalized attribute matching score mentioned above are weighted and fused to obtain the final recommendation score for each standard. The scores are then sorted in descending order to generate an initial recommendation list, such as: ["GB / T19963 Article 6.3.2", "NB / T32004 Article 5.1", "DL / T1196..."].
[0041] 4. Feedback Collection: The list is presented to the user. The system monitors user behavior: The user clicked on "GB / T19963 Article 6.3.2" and stayed for 120 seconds, ignoring "DL / T1196...", and ultimately adopted (downloaded) the first two items. These timestamped and intensity-based behaviors are encoded into an interactive feedback sequence and stored in association with the "business scenario parameters" and "core intent category" (comparative analysis) of this session.
[0042] 5. Real-time feedback optimization: Algorithm selection: Based on the "core intent category" (contrastive analysis), the system selects an immediate learning algorithm (such as a specific reordering algorithm) from the policy mapping table that focuses on improving differentiation and juxtaposition.
[0043] Dynamic optimization: The interaction feedback sequence (clicks, ignore actions) is input into the algorithm. The algorithm interprets that the user pays more attention to the first two items in this "comparative analysis," and then generates and adjusts the weights accordingly. This may involve shifting the third criterion, which is less relevant to the current scenario, to a lower position, or maintaining the prominence of the first two items. Finally, an optimized recommendation list is generated and output.
[0044] 6. Long-term feedback evolution: All data from this session (feedback sequence, scene context, intent category) is stored in the feedback knowledge base.
[0045] When the model update conditions are met (e.g., every Sunday morning), the system initiates long-term feedback learning.
[0046] Incremental learning algorithms (such as regularization that includes knowledge distillation) are used to calculate model update parameters (gradients and update amounts) using historical data from the knowledge base.
[0047] These parameters are used to fine-tune and update the weights of the hybrid deep recommendation network and the user's long-term interest vector. For example, the network's ability to correlate "new energy access design" scenarios with reactive power regulation clauses can be enhanced, or the "reactive power compensation" related dimensions can be strengthened in the user's interest vector.
[0048] The updated model and vectors will be used to process all subsequent user requests, resulting in an iterative improvement in the system's overall recommendation capabilities.
[0049] Existing technology and its shortcomings: (1) Recommendation based on keyword matching and content similarity Method description: Extract keywords from user queries or historical documents, perform full-text matching or TF-IDF similarity calculations in a standard library, and return the criteria with high similarity as recommendations.
[0050] Technical defects: Context-awareness: Relying solely on lexical matching, it fails to understand the business context behind the query (e.g., "design" and "operation" have different standard requirements for the same device), resulting in recommendation results that are disconnected from the context.
[0051] Knowledge fragmentation: Treating standards as isolated documents makes it impossible to recommend standards that are logically related but have dissimilar textual descriptions (such as "Transformer Selection Standard" and the subsequent "Installation, Construction and Acceptance Specification").
[0052] Static and fixed: The recommendation logic is fixed and cannot be dynamically adjusted based on user feedback and the shift in business trends.
[0053] (2) Recommendation based on collaborative filtering and historical behavior Method description: Analyze the user-standard interaction matrix (such as downloads and browsing) to discover similarities between user groups or standards, and make recommendations such as "other standards liked by similar users" or "people who have viewed this standard have also viewed".
[0054] Technical defects: The cold start problem is severe: for newly released standards, new users, or niche professional fields, effective recommendations cannot be made due to a lack of interaction data.
[0055] Lack of domain logic: The recommendation relies entirely on historical behavior statistics, which may reinforce popular standards but ignore key but less common technical specifications, and cannot explain the business rationale behind the recommendation.
[0056] Context confusion: It is impossible to distinguish the different needs of the same user in different project stages (contexts), and it is easy to recommend duplicate or irrelevant content.
[0057] (3) Recommendations based on simple rules or expert systems Method description: Rules are written by domain experts, such as "If 'new energy access' is involved, then standards 'A, B, and C' are recommended".
[0058] Technical defects: High maintenance costs: the rules are difficult to cover massive and dynamically changing business scenarios and standard systems, and have poor scalability and maintainability.
[0059] Poor flexibility: It cannot handle complex and compound query intents, and rules are prone to conflict.
[0060] Unable to evolve on its own: Rule updates rely on manual intervention and cannot automatically learn and optimize from user feedback.
[0061] (4) Lack of a closed-loop feedback mechanism Method description: Most systems only use user clicks, ratings, etc. as simple statistical indicators, or for periodic, manual model retraining.
[0062] Technical defects: Low feedback utilization: Abundant implicit behavioral data (browsing duration, query corrections, workflow context) has not been deeply mined and quantified.
[0063] The optimization is severely lagging: the batch and offline model update method results in slow system response and feedback, making it impossible to achieve real-time adaptation.
[0064] Decoupled from business scenarios: The feedback data is not strongly correlated with the "business scenarios" that generate the recommendations, resulting in inaccurate optimization directions.
[0065] This invention has the following significant beneficial effects: (1) Scene depth perception and precise adaptation: By deconstructing the fuzzy business scene parameters into multi-dimensional element tuples and using the domain knowledge graph to generate dynamic scene state vectors, the system has a deep and structured scene understanding capability, thereby achieving precise matching between recommendation results and complex business workflows.
[0066] (2) Intent-driven real-time interaction optimization: By combining the user's real-time interaction feedback sequence with its core intent category for immediate learning, the recommendation strategy (such as reordering and content supplementation) can be dynamically adjusted within a single session, enabling the system to have interactive intelligence of "learning while using and responding instantly", which significantly improves the satisfaction of a single query.
[0067] (3) Closed-loop feedback and system self-evolution: A two-layer feedback closed loop of "immediate learning to optimize experience and long-term learning to evolve model" was constructed. The system can not only optimize the current result, but also store the feedback data in association with the business context, and use this to drive the incremental iterative update of the core model, realizing the transformation of the recommendation system from "static tool" to "living intelligence", and the long-term recommendation effect continues to improve.
[0068] (4) Knowledge fusion and interpretability enhancement: The method deeply integrates multiple knowledge sources such as domain knowledge graph, multimodal standard features, and long-term user interests, and the optimization decision (such as the intention-based selection algorithm) has a clear logical basis, which enhances the interpretability and credibility of the recommendation results.
[0069] Example 2: Based on the same inventive concept, this invention also provides an adaptive recommendation and feedback optimization system for power business scenarios, the system structure of which is as follows: Figure 2 As shown, the system includes: Intent parsing receiving module 201 is used to receive the structured intent parsing results for the current session; The state vector generation module 202 is used to extract and construct a multi-dimensional scene element tuple based on the business scene parameters in the structured intent parsing result, retrieve the association information from the pre-built business scene-standard association graph based on the multi-dimensional scene element tuple, and generate a dynamic scene state vector based on the association information. The initial list generation module 203 is used to input the dynamic scene state vector, the user long-term interest vector obtained from the user profile library, and the candidate standard features related to the current session obtained from the standard feature library into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; and to obtain the interaction feedback sequence generated for the initial recommendation list. The optimized list generation module 204 is used to perform real-time feedback learning based on the core intent categories in the interactive feedback sequence and structured intent parsing results, dynamically optimize the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generate and output the optimized recommendation list.
[0070] Preferably, the state vector generation module is specifically used for: The business scenario parameters are analyzed to identify and extract at least three elements from the core tasks, equipment objects, project stages, and compliance requirements, and a structured multidimensional scenario element tuple is constructed. The elements in the multidimensional element tuple are used as query nodes. Multi-hop traversal and subgraph retrieval are performed in the business scenario-standard association graph. The relevant standard entities, technical term entities and their relationship paths are used as association information. The multidimensional scene element tuples and associated information are input into the scene encoder. The scene encoder then uses an attention mechanism to fuse and compress the input multidimensional scene element tuples and associated information, and outputs a fixed-dimensional dynamic scene state vector.
[0071] Preferably, the initial list generation module is specifically used for: A multi-channel input layer is constructed for a hybrid deep recommendation network. The first channel of the multi-channel input layer is the dynamic scene state vector, the second channel is the user's long-term interest vector, and the third channel is the candidate standard features. Among them, the candidate standard features are multimodal features, including the text semantic vector and metadata of the candidate standards. By using deep cross layers in the network, attention interaction is performed between the dynamic scene state vector and the text semantic vector to calculate the scene-semantic matching score; feature cross is performed between the user's long-term interest vector and metadata to calculate the personalization-attribute matching score. The scene-semantic matching score and the personalization-attribute matching score are weighted and fused to generate the final recommendation score for each candidate criterion; All candidate criteria are sorted in descending order based on the final recommendation score to generate an initial recommendation list.
[0072] Preferably, the initial list generation module is specifically used for: Monitor and record the exposure, clicks, ignores, dwell time, and final adoption behavior of each candidate criterion in the initial recommendation list within the current session; Encode behaviors in chronological order into behavior sequences with timestamps and intensity weights to construct interaction feedback sequences; The interaction feedback sequence is associated with and stored in relation to the business scenario parameters and core intent types for the current session.
[0073] Preferably, the optimized list generation module is specifically used for: Based on the core intent category, the corresponding on-the-spot learning algorithm is selected from the preset policy mapping relationship; whereby the policy mapping relationship defines the correspondence between different core intent categories and different on-the-spot learning algorithms; The interactive feedback sequence is input into the selected real-time learning algorithm to generate real-time adjusted weights for each item in the initial recommendation list; The original scores of each item in the initial recommendation list are recalculated by adjusting the weights in real time, and / or the order of the list is adjusted to generate an optimized recommendation list.
[0074] Preferably, the system also includes a module update module for: Store the interaction feedback sequence and its corresponding scene context in the feedback knowledge base; When the preset model update conditions are met, long-term feedback learning is performed based on historical data stored in the feedback knowledge base, which consists of interactive feedback sequences and corresponding scene contexts, to generate model update parameters. The model update parameters are used to iteratively update the hybrid deep recommendation network and the user's long-term interest vector; The iteratively updated hybrid deep recommendation network and user long-term interest vectors are used to perform fusion calculations to generate a new initial recommendation list.
[0075] Preferably, the preset model update conditions include at least one of the following: reaching a preset time period, the feedback data accumulated in the feedback knowledge base reaching a preset threshold, and monitoring that the average user satisfaction index of the optimized recommendation list is continuously lower than a preset threshold. The module update module is specifically used for: An incremental learning algorithm is used to calculate the parameter gradient of the hybrid deep recommendation network and the update amount of the user's long-term interest vector as model update parameters, using historical data as training samples. The incremental learning algorithm includes minimizing the loss function and regularization constraints that incorporate a knowledge distillation mechanism.
[0076] Example 3: Based on the same inventive concept, such as Figure 3 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.
[0077] The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in a readable storage medium to implement the corresponding method flow or corresponding function, so as to realize the steps of an adaptive recommendation and feedback optimization method for power business scenarios in the above embodiments.
[0078] Example 4: Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the readable storage medium here can include both the built-in storage medium within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the storage medium to implement the steps of the adaptive recommendation and feedback optimization method for power business scenarios described in the above embodiments.
[0079] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0080] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0081] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0082] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.
Claims
1. An adaptive recommendation and feedback optimization method for power business scenarios, characterized in that, include: Receive the structured intent parsing results for the current session; Based on the business scenario parameters in the structured intent parsing results, a multi-dimensional scenario element tuple is extracted and constructed. Based on the multi-dimensional scenario element tuple, the association information is retrieved from the pre-constructed business scenario-standard association graph. Based on the association information, a dynamic scenario state vector is generated. The dynamic scene state vector, the user long-term interest vector obtained from the user profile database, and the candidate standard features related to the current session obtained from the standard feature database are input into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list. Obtain the interaction feedback sequence generated in response to the initial recommendation list; Based on the interactive feedback sequence and the core intent categories in the structured intent parsing results, immediate feedback learning is performed. Based on the immediate feedback learning results, the order and / or content of the initial recommendation list are dynamically optimized, and an optimized recommendation list is generated and output.
2. The method according to claim 1, characterized in that, Based on the business scenario parameters in the structured intent parsing results, a multi-dimensional scenario element tuple is extracted and constructed. Relationship information is retrieved from a pre-constructed business scenario-standard association graph based on the multi-dimensional scenario element tuple. A dynamic scenario state vector is generated based on the relationship information, including: The business scenario parameters are parsed to identify and extract at least three elements from the core tasks, device objects, project stages, and compliance requirements, and a structured multidimensional scenario element tuple is constructed. The elements in the multidimensional element tuple are used as query nodes. Multi-hop traversal and subgraph retrieval are performed in the business scenario-standard association graph. The relevant standard entities, technical term entities and their relationship paths are used as association information. The multidimensional scene element tuple and the associated information are input into the scene encoder. The scene encoder fuses and compresses the input multidimensional scene element tuple and the associated information based on an attention mechanism, and outputs the dynamic scene state vector with a fixed dimension.
3. The method according to claim 1, characterized in that, The process of inputting the dynamic scene state vector, the user's long-term interest vector, and the candidate standard features into a hybrid deep recommendation network for fusion calculation to generate an initial recommendation list includes: A multi-channel input layer is constructed for the hybrid deep recommendation network. The first channel of the multi-channel input layer is the dynamic scene state vector, the second channel is the user's long-term interest vector, and the third channel is the candidate standard features. The candidate standard features are multimodal features, including the text semantic vector and metadata of the candidate standards. Through a deep cross layer of the network, attention interaction is performed between the dynamic scene state vector and the text semantic vector to calculate the scene-semantic matching score; feature cross is performed between the user's long-term interest vector and the metadata to calculate the personalization-attribute matching score. The scene-semantic matching score and the personalization-attribute matching score are weighted and fused to generate the final recommendation score for each candidate criterion; All candidate criteria are sorted in descending order based on the final recommendation score to generate the initial recommendation list.
4. The method according to claim 1, characterized in that, The step of obtaining the interaction feedback sequence generated for the initial recommendation list includes: Monitor and record the exposure, clicks, ignores, dwell time, and final adoption behavior of each candidate criterion in the initial recommendation list within the current session; The behaviors are encoded in chronological order into a sequence of behaviors with timestamps and intensity weights to construct the interaction feedback sequence; The interaction feedback sequence is associated with and stored in relation to the business scenario parameters and the core intent type for the current session.
5. The method according to claim 1, characterized in that, The process involves performing real-time feedback learning based on the core intent categories in the interactive feedback sequence and the structured intent parsing results, dynamically optimizing the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generating and outputting an optimized recommendation list, including: Based on the core intent category, a corresponding on-the-spot learning algorithm is selected from a preset strategy mapping relationship; wherein, the strategy mapping relationship defines the correspondence between different core intent categories and different on-the-spot learning algorithms; The interactive feedback sequence is input into the selected real-time learning algorithm to generate real-time adjusted weights for each item in the initial recommendation list; The original scores of each item in the initial recommendation list are recalculated using the real-time weight adjustment, and / or the order of the list is adjusted to generate the optimized recommendation list.
6. The method according to claim 1, characterized in that, After dynamically optimizing the order and / or content of the initial recommendation list based on the learning results of real-time feedback, and generating and outputting the optimized recommendation list, the method further includes: The interactive feedback sequence and the corresponding scene context are stored in the feedback knowledge base; When the preset model update conditions are met, long-term feedback learning is performed based on the historical data stored in the feedback knowledge base, which consists of the interaction feedback sequence and the corresponding scene context, to generate model update parameters. The hybrid deep recommendation network and the user's long-term interest vector are iteratively updated using the model update parameters. The iteratively updated hybrid deep recommendation network and the user's long-term interest vector are used to perform fusion calculations to generate a new initial recommendation list.
7. The method according to claim 6, characterized in that, The preset model update conditions include at least one of the following: reaching a preset time period, the feedback data accumulated in the feedback knowledge base reaching a preset threshold, and the average user satisfaction index of the optimized recommendation list being continuously lower than a preset threshold. The step of generating model update parameters by performing long-term feedback learning based on historical data stored in the feedback knowledge base, consisting of the interaction feedback sequences and corresponding scene contexts, includes: An incremental learning algorithm is used to calculate the parameter gradient of the hybrid deep recommendation network and the update amount of the user's long-term interest vector, using the historical data as training samples, as the model update parameters. The incremental learning algorithm includes minimizing a loss function and regularization constraints that incorporate a knowledge distillation mechanism.
8. An adaptive recommendation and feedback optimization system for power business scenarios, characterized in that, include: The intent parsing receiving module is used to receive the structured intent parsing results for the current session; The state vector generation module is used to extract and construct a multi-dimensional scene element tuple based on the business scenario parameters in the structured intent parsing result, retrieve the association information from the pre-constructed business scenario-standard association graph based on the multi-dimensional scene element tuple, and generate a dynamic scene state vector based on the association information. The initial list generation module is used to input the dynamic scene state vector, the user long-term interest vector obtained from the user profile library, and the candidate standard features related to the current session obtained from the standard feature library into the hybrid deep recommendation network for fusion calculation to generate an initial recommendation list; and to obtain the interaction feedback sequence generated for the initial recommendation list. The optimized list generation module is used to perform real-time feedback learning based on the core intent categories in the interactive feedback sequence and the structured intent parsing results, dynamically optimize the order and / or content of the initial recommendation list based on the real-time feedback learning results, and generate and output the optimized recommendation list.
9. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the adaptive recommendation and feedback optimization method for power business scenarios as described in any one of claims 1 to 7 is implemented.
10. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements the adaptive recommendation and feedback optimization method for power business scenarios as described in any one of claims 1 to 7.