388 results about "Cold start" patented technology

The invention discloses a personalized learning resource recommendation method and system based on multi-agent collaboration and a dynamic knowledge graph, and the method comprises the steps: constructing the dynamic knowledge graph, enabling nodes to be associated with teaching resources (videos, test questions, teaching plans, PPT and the like), and enabling edges to represent the logic relation between the resources; collaborative decision is made through four layers of agents: a target determination agent generates a learning target based on student historical learning data and a graph node state; the path planning agent plans a learning path in combination with the target and the learner model; the resource screening agent matches personalized resources from the path nodes; the user portrait intelligent agent updates the learner model in real time; and finally, generating a dynamic recommendation result and feeding back the optimized knowledge graph. Through multi-agent hierarchical collaboration and dynamic interaction with the knowledge graph, the problems of cold start, incomplete resource coverage and path stiffness of a traditional recommendation system are solved, and precise and adaptive learning resource recommendation is realized.

The invention relates to the technical field of intelligent recommendation, and discloses an AI intelligent matching method based on a knowledge graph, and the method comprises the steps: constructing a quaternary knowledge graph containing a time dimension; identifying legal entities and semantic relationships by adopting an entity identification and relationship extraction technology; multi-level semantic features are extracted through a two-layer progressive semantic matching algorithm of a grammar layer, a semantic layer and a reasoning layer; constructing lawyer ability portraits based on a heterogeneous graph neural network and a time sequence perception graph convolution technology; progressive matching calculation is adopted, the optimal matching weight is learned through a multi-layer attention mechanism, and dynamically optimized intelligent matching is achieved. The technical problems of cold start, insufficient semantic understanding ability and poor timeliness processing ability in the existing legal consultation matching system can be solved, and the matching precision and the user satisfaction are improved.

A collaborative transformation matrix learning system extends adaptive compression and encryption architectures through federated, privacy-preserving optimization. Each node analyzes local data distributions to generate anonymized distribution profiles using differential-privacy mechanisms, securely exchanging profiles and validated transformation matrices across a collaborative network. A trust and validation engine verifies mathematical properties and evaluates claimed performance metrics. Validated matrices are integrated into local optimization when trust and performance thresholds are satisfied. The system employs secure multi-party computation, homomorphic encryption, and conflict-resolution logic to ensure integrity of shared insights while preventing exposure of sensitive information. By combining collective learning with local adaptation, the invention accelerates convergence to optimal matrix configurations, mitigates cold-start inefficiencies, and improves compression-encryption efficiency and cryptographic strength across distributed deployments.

The invention discloses a reinforcement learning unmanned ship path control method based on double-track regulation and control random network distillation. The method comprises the operation steps that an unmanned ship builds a path tracking simulation environment and a kinetic model; the unmanned ship builds a core algorithm flexible action evaluation algorithm framework; the unmanned ship deploys a priority experience playback pool based on quality and success guidance; an uncertainty perception and risk perception mechanism is introduced into the unmanned ship; the unmanned ship builds a success rate-based reward attenuation and cold start module, and the unmanned ship calculates a total reward and designs a reward softening mechanism to smooth the total reward; the unmanned ship imports hyper-parameters of all the modules, starts training circulation in a simulation environment, and dynamically adjusts exploration intensity and the like; according to the method, uncertainty and risk indexes are introduced, the exploration intensity of the intelligent agent is controlled, the intelligent agent is prevented from making dangerous actions, and the robustness is improved; a priority experience playback pool based on quality and success guidance is introduced, high-quality samples are better played back, and strategy convergence is accelerated.

The invention discloses a large video model training method and a related device, and relates to the technical field of video recognition, and the method comprises the steps: collecting a training video data frame to obtain an image frame, inputting a preset prompt word, a user question and the image frame into a large image model, and obtaining a thinking chain and a question answer. Performing cold start on the video large model based on the thinking chain and the question answer to enable the video large model to have thinking chain output capability; and combining training video data and questions to generate space and time disordered data and thinking chain data. And inputting the three types of data into the model to obtain corresponding outputs, calculating the accuracy of each output, obtaining space and time accuracy reward values, and training the model through a group relative strategy optimization algorithm in combination with a thinking chain consistency reward value to obtain an inference video large model. According to the method, the trained video large model can have thinking reasoning capability based on thinking chain implementation.

The invention discloses a cross-domain AI knowledge aggregation method based on collaborative filtering. The method comprises the steps that S1, multi-source heterogeneous AI knowledge data and user behavior data are collected and preprocessed; s2, constructing a double-tower cross-domain embedded network, and outputting a cross-domain semantic fusion sequence; s3, modeling and analyzing user preferences through the improved Bi-GRU network, and generating user behavior preference vectors; s4, performing semantic diffusion and neighborhood reasoning on the cold start user, and complementing interest features; s5, adopting a double-tower recall structure and an XGBoost model to sort and generate a cross-domain recommendation list; s6, constructing a context rule base to execute context adaptability judgment, and generating a matched knowledge aggregation recommendation list; and S7, performing incremental learning according to user feedback information, and dynamically updating the double-tower cross-domain embedded network and the improved Bi-GRU network. According to the method, the knowledge matching precision, the cold start adaptability and the scene adaptation capability of cross-domain recommendation content are improved.

The invention discloses an electric vehicle charging load generation method and system. The method comprises the steps that a one-dimensional time sequence signal is converted into a heterogeneous feature token containing local, global and guiding information through mixed token processing; constructing a decomposition type diffusion model based on the token, and directly generating three core components of trend, seasonality and residual error of the time sequence; optimizing a model training process by adopting a composite loss function and a warm start mechanism so as to solve the cold start problem of model training; and performing high-quality fusion reconstruction on each component through a weighted overlap-add method, and outputting a final charging load curve. According to the method, highly real and diversified charging load data can be generated, and powerful data support is provided for power grid planning and dispatching in a structure interpretable mode.

The invention belongs to the technical field of knowledge maps, and discloses a next interest point recommendation method based on a cross-regional city space knowledge map. According to the method, on the basis of the geographic space knowledge graph and the user preference knowledge graph, a cross-regional relationship modeling mechanism is introduced, the spatial proximity and the reachability between the regions are comprehensively considered, and more comprehensive description of the cross-regional behavior of the user is realized. According to the method, a recommendation framework combining a geographic module and a sequence module is designed, spatial dependence between interest points can be captured, a high-order sequence mode in user sign-in behaviors can be mined, and recommendation accuracy and diversity are improved. Geographic representation and sequence representation are fused through a consistency learning framework, the robustness and generalization ability of the model can be enhanced, and the model can still keep stable performance in sparse data and cold start scenes. According to the method, the defects of insufficient region boundary perception, recommendation result centralization, poor cross-region prediction adaptability and the like of an existing method are overcome.

The invention discloses a graph database query acceleration method and system. The method comprises the following steps: constructing panoramic query observation data through non-intrusive log interception and structured processing; the query statement is analyzed into an abstract syntax tree, semantic normalization processing is carried out, and a unique hash value is generated; predicting a hotspot query template and high-frequency parameters thereof based on a popularity value score model, and asynchronously preloading a result to a cache; dynamically calculating adaptive survival time for each cache item according to the data change frequency and the access popularity; and constructing a multi-level cache architecture comprising a local cache, a distributed cache and a database built-in cache, and performing intelligent routing and collaborative backfilling based on a hash value. According to the method, intelligent and transparent acceleration of graph database query is realized, the query performance is remarkably improved, the database load is reduced, and the problems of redundancy, high cold start delay and difficulty in balancing cache consistency caused by grammar difference in traditional cache are solved.

The invention discloses a private cloud platform-based micro-service application remote main and standby disaster recovery rapid switching method, relates to the technical field of computers, and is used for solving the problems of long remote main and standby disaster recovery switching time, cold start storm, insufficient capacity evaluation and discharge strategies and complex back switching. The switching process is decomposed into controllable steps by combining real-time state feedback through a clear sequential arrangement and batch reconnection strategy of a data surface and a control surface, so that the switching time is remarkably shortened, the target of minute-level continuous availability is met, and meanwhile, the switching efficiency is greatly improved through heat set construction and standby side quasi real-time preheating before switching. Core hotspot data and indexes are loaded to the standby side in advance, it is ensured that the standby side has the high cache hit rate and index availability during master raising, and therefore the performance bottleneck caused by access of a large amount of cold data is avoided.

The invention discloses an AI person and post matching intelligent recommendation method based on a graph neural network, and relates to the technical field of data processing, and the method comprises the following steps: S1, generating a multi-view heterogeneous graph; s2, outputting a community membership matrix and a community-level statistical feature vector; s3, generating an initial feature vector; s4, constructing an improved HetGNN model for non-cold start nodes, and outputting a fusion feature vector; s5, outputting a decoupling feature vector; s6, constructing a reciprocity preference matching mechanism, and generating a corrected matching score; and S7, outputting a final man-post matching result. According to the method, the limitations of single feature representation, poor cold start recommendation effect and insufficient privacy protection in a traditional man-post matching method are overcome, and an efficient, safe and credible solution is provided for intelligent configuration and decision support of human resources.

The invention belongs to the technical field of crowd sensing, and discloses a trust-based crowd sensing distributed privacy protection method and system, and the method comprises the steps: constructing a detailed reputation evaluation system, covering the calculation of a reputation value and the setting of a reputation threshold value of a participating user, and integrating a game model in an evaluation stage, the user is guided to avoid malicious behaviors through an income incentive mechanism so as to obtain higher income; a dobby machine model is introduced to dynamically balance the decision-making process of'exploring 'new users and'utilizing' high-reputation users, so that the cold start problem of insufficient reputation of the new users is effectively solved; and finally, verifying the security of the scheme through theoretical derivation of the security attribute of the scheme, verifying the validity of the scheme in the aspects of data credibility, privacy protection effect and the like in combination with experimental analysis, and providing support for credible operation and privacy protection of the crowd sensing system.

The invention provides a virtual user simulation method and system for dialogue testing, and the method comprises the steps: obtaining a plurality of pieces of effective multi-round dialogue data from a to-be-tested system, and carrying out the multi-label labeling, so as to construct a basic sample data set containing user portrait labels, the key sample data set further comprises user emotion tags; constructing a group optimization strategy model, performing cold start training based on the basic sample data set, and performing intensive training by using a multi-dimensional composite reward function based on the key sample data set to obtain a dialogue generation model with an emotion cognition-verbal skill expression decoupling architecture; and taking a historical round of dialogue of the to-be-tested system aiming at the current interaction as input of a dialogue generation model, and generating a next round of user reply consistent with the user emotion and the user portrait. According to the invention, deep, efficient and automatic adversarial testing is carried out on the task-oriented dialogue system, so that system defects which are difficult to find by a traditional testing method are effectively exposed.

The invention relates to the technical field of industrial equipment predictive maintenance and fault diagnosis, in particular to a multi-motor coupling vibration intelligent diagnosis method and system. According to the technical scheme, the method comprises the following steps: generating a coupled vibration simulation data set with an accurate fault tag through a parameterized fault simulation module based on collected vibration data of multiple motors in a normal operation state and physical parameters of motor equipment; according to the invention, cold start of the diagnosis system under the condition of no historical fault sample is realized; a deep reinforcement learning agent is used for processing a space-time coupling signal, and a layered reward mechanism is combined, so that the decoupling diagnosis precision and the early warning capability are remarkably improved; through the online self-adaptive learning module, the system can be finely adjusted in real time according to field feedback to quickly adapt to the state of individual equipment; and the simulation-model evolution module continuously optimizes the simulator and the diagnosis model by using the accumulated real data to form a self-iterative closed loop.

The invention discloses an intention recognition model reinforcement learning training method and system controlled by an operating system, and the method comprises the steps: constructing intention and slot labels for natural language commands controlled by the operating system, and constructing a training data set according to the natural language commands and the corresponding intention and slot labels; a LoRA technology is adopted to supervise and finely adjust the pre-trained intention recognition model based on the training data set, and a cold start intention recognition model is obtained; and carrying out reinforcement learning training on the cold-start intention recognition model by adopting a reinforcement learning method introducing group relative strategy optimization, so as to obtain the intention recognition model which can be finally used for end-side deployment. The method aims at solving the problems that in the prior art, coverage of an intention and slot label system is insufficient, data quality requirements are high but acquisition is difficult, model customization cost is too high, and user feedback driving optimization is lacked, and the speed and precision of operating system intention recognition on end side equipment are improved.

A text sequence recommendation method and system based on a large language model is disclosed, belonging to the technical field of recommendation algorithms. The method includes: a data preprocessing stage, a large language model pre-training stage, a sequence model fine-tuning stage and a matching stage. According to this disclosure, a large language model is introduced into a text sequence recommendation task, so that text can be better modeled by utilizing rich pre-training corpus of the large language model; meanwhile, sequence modeling is performed on the text, the capability of sequence recommendations modeling in a large model is activated, an ID-based recommendation paradigm in a traditional recommendation algorithm is eliminated, and recommendation task learning processing is better performed in a cold start scenario and a knowledge transfer scenario; and finally, a recommendation result is finally optimized by a sequence model.

The invention relates to the technical field of deep learning, in particular to a forage grass yield prediction model construction method based on deep learning, which comprises the steps of constructing a multi-source data fusion module, establishing a cold start mechanism, designing a multi-modal deep learning prediction network, integrating a physical constraint mechanism and constructing a management decision support system. A seasonal attribution analysis function is realized; in the prior art, a simple data superposition or static weighted fusion scheme is generally adopted, and inherent defects of deficiency, different scales and heterogeneity of multi-source data are difficult to process, so that the fusion feature quality is poor; according to the method, firstly, a data blank is accurately filled through an intelligent algorithm based on space-time continuity, then heterogeneous data is unified to a standard grid by using a multi-scale pyramid engine, and finally, deep fusion is performed through an attention mechanism for dynamically calculating importance of each data source; the integrity, the consistency and the information density of the input data are remarkably improved, and a solid and reliable data foundation is laid for subsequent accurate prediction.

The invention discloses an end-end self-control-based transformer area light storage and charging coordinated regulation and control system and an autonomous method. The system constructs an end-end-transformer three-level autonomous architecture: an end-level autonomous unit is responsible for local millisecond-level reflection control; the edge autonomous nodes realize intra-zone collaboration through federated learning, and physical security constraints are embedded into AI decisions by using digital twin bodies; the court-level federation learning center realizes knowledge sharing under cross-court privacy protection through secure multi-party computing, the method further integrates transfer learning to solve the cold start problem of a new court, and a three-time scale regulation and control mechanism is adopted to cope with disturbance of different rates; the autonomous capability, the response speed, the safety level and the intelligent degree of the system are improved.

The invention relates to a photovoltaic source domain selection method based on a statistical index and a maximum and minimum regret value method. The method comprises the following steps of S1, preprocessing photovoltaic power and meteorological data, deleting redundant zero values, performing normalization, and constructing a standardized data set; s2, based on three types of statistical indexes of MMD, CORAL and cosine similarity, primarily selecting a candidate source domain with high compatibility; s3, carrying out detailed selection through a maximum and minimum regret value method, and determining an optimal source domain; s4, constructing a deep learning model, and realizing migration prediction through source domain pre-training and target domain fine tuning; and S5, verifying the performance in multiple dimensions. According to the method, the cold start problem of the newly-built photovoltaic power station can be effectively solved, MSE predicted by Step6 is reduced from 0.6861 to 0.3179, R is increased from 67.74% to 85.05%, and the method is superior to a traditional baseline method. The method can improve the photovoltaic prediction precision in a data scarce scene, is suitable for a photovoltaic power prediction task of a newly-built power station, and facilitates power grid dispatching and operation and maintenance decision making.

The invention relates to the technical field of management systems, and particularly discloses a three-dimensional industrial design material AI intelligent recommendation management system. Comprising a dynamic knowledge graph module, a multi-source perception input module, a neural symbol fusion engine, a situation self-adaptive output module, a self-evolution optimization module, a recommendation rule updating module, a cold start processing module and a multi-target decision module. The neural symbol fusion engine realizes symbol logical reasoning through Transform, the cold start processing module generates a user portrait through gradient embedding and generates virtual user features through cross-map feature propagation, the situation adaptive output module dynamically switches rendering modes according to a design stage, and the self-evolution optimization module performs meta-learning and concept drift detection; through multi-technology fusion and module cooperation, intelligence, precision and high efficiency of three-dimensional industrial design material recommendation are realized, the design efficiency is remarkably improved, and the project landing risk is reduced.

The invention discloses a code repository understanding enhancement method based on Monte Carlo tree search-driven agent reinforcement learning, which comprises the following steps: (1) collecting open source code repository data, preprocessing and filtering, and designing five special code repository exploration tools; (2) constructing a code repository question and answer agent based on the designed code repository exploration tool, and realizing a multi-round thinking-action-observation interaction mode; (3) using a Monte Carlo tree search algorithm to guide the intelligent agent to generate diversified exploration tracks in the code warehouse data, and calculating track rewards through process rewards and correctness rewards; (4) performing reinforcement learning training on the intelligent agent by adopting a GRPO algorithm to realize cold start learning; and (5) in the application process, inputting the question into the trained agent, and finding out codes or text fragments related to question answering in the warehouse. According to the method, the performance of the code warehouse question and answer task is improved, and the training efficiency is improved compared with general agent intensified training.

The invention relates to the technical field of natural language processing and artificial intelligence, and particularly discloses a generation and retrieval collaborative optimization method and system for a ReGenSearch task. The system comprises a reflection module, a query rewriting module, a retrieval and generation collaborative optimization module, a reinforcement learning and reward mechanism module and an answer optimization and generation module. The method comprises the following steps: analyzing problems and expanding backgrounds through a reflection module, generating an optimized retrieval query through a query rewriting module, dynamically selecting a generation or retrieval strategy by a retrieval and generation collaborative optimization module, and judging whether iteration is carried out by combining a self-commenting mechanism; a joint strategy is optimized through a DAPO algorithm, and feedback is evaluated through a layered reward mechanism; and finally, comparing the candidate answers by using a DPO algorithm, and selecting an optimal reasoning path for outputting. The method does not need cold start or fine tuning in the specific field, can directly run based on the pre-trained large language model, adapts to multi-step reasoning and long thinking chain tasks, realizes deep fusion of generation and retrieval, and remarkably improves the accuracy, stability and generalization ability of complex questions and answers.

The invention discloses a large language model alignment method based on multi-user preference dynamic balance, and the method comprises the steps: collecting user interaction data, preference comparison data and public policy question and answer data in a livelihood consultation scene, and dividing the data into structured preference data and unstructured data; preprocessing the divided data to generate a user feature vector and a response feature vector; constructing a user-response bipartite graph, performing multi-hop preference propagation through a graph neural network, capturing potential association among users, and outputting a user preference embedding vector; embedding a vector selection expert path according to user preference through a dynamic gating function, and generating a personalized response; and through an optimization-free embedding aggregation strategy, similar users are retrieved based on a graph structure, and the embedding of the similar users is weighted and aggregated, so that rapid cold start of new users is realized. The precision and efficiency of livelihood consultation can be remarkably improved, and the method is particularly suitable for smart city scenes with high-frequency policy updating and frequent user flow.

This application provides a model training method, an information recommendation method, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. The method includes: during the training of a first prediction model, when a sample collection opportunity is determined, acquiring first sample data of a first recommendation scenario within the current time period and second sample data of a second recommendation scenario within the current time period; merging the first and second sample data to obtain merged sample data; determining a first prediction conversion rate of the first recommendation scenario, and sampling the merged sample data based on the first prediction conversion rate to obtain new sample data; adding the new sample data to the training dataset of the first prediction model to obtain an updated training dataset; and training the first prediction model based on the updated training dataset. This application can improve the accuracy of conversion rate prediction during cold starts.

The present application relates to a kind of based on retrieval enhancement generation's toxicity text acquisition method and system, first by obtaining target text data through crawling platform, and using cold start mode artificial construction small-scale initial data set, inject knowledge base as basic data. The semantic retrieval is carried out to each batch target text, obtain the semantic similarity from knowledge base k text, combine these texts by thought chain reasoning using multiple large language model reasoning, generate toxicity label, and mark the target text to be marked in current batch. The target text after marking is injected into the knowledge base, and the data is continuously collected and labeled in an iterative manner, and the performance of the retriever is continuously optimized through incremental learning mechanism during the iteration process to achieve toxicity text acquisition. The present application can efficiently accumulate large-scale, fine-grained, high-consistency toxicity text labeling corpus, significantly alleviate the problems of high labeling cost, inconsistent quality and poor scalability in traditional toxicity text data collection.

The application relates to a cold start object recommendation method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring content features and initial preference features of a target object in a cold start state; performing multi-layer perception processing on the content features based on a meta-mapping network to determine a feature mapping relationship matched with the target object; the meta-mapping network is obtained based on interaction data of a recommendation object in a simulated cold start state; performing mapping transformation on the initial preference features according to the feature mapping relationship to obtain updated preference features close to a target preference feature space distribution; the target preference feature space distribution is used for representing a space distribution of preference features corresponding to the recommendation object; and performing recommendation analysis on the target object based on the updated preference features to obtain a recommendation result for the target object. The method can improve the recommendation accuracy of the target object in the cold start stage.

The embodiment of the invention provides a strategy determination method and device. The method comprises the following steps: acquiring real-time index data and a real-time prediction load of a Serverless function; determining a plurality of target optimization functions according to the real-time index data and the real-time prediction load; the multiple target optimization functions comprise a first target function and a second target function, the first target function is used for minimizing the cold start delay of the function, and the second target function is used for minimizing the resource cost of the function; the candidate solution of each target optimization function indicates a combined strategy of a resource allocation strategy and a cold start strategy; a plurality of target optimization functions are processed by utilizing a multi-target evolutionary algorithm, and a target combination strategy is screened out from numerous candidate solutions by simulating the principle of survival of victims and inferior and continuous optimization of populations in a biological evolution process, so that the resource cost of the functions is reduced on the basis of ensuring cold start low delay of the functions; and the resource allocation efficiency is improved.

The invention discloses a virtual space-time environment automatic completion method and system based on data density dynamic perception and a medium, and relates to the field of computer data processing. The method comprises the following steps: monitoring a data activeness index (such as spatio-temporal information entropy) of a spatio-temporal grid in real time in a distributed index database of a virtual spatio-temporal environment; when the index meets a preset data sparseness condition, generating a completion request; analyzing the space-time attribute tag of the target grid and acquiring associated multi-modal reference data from an external heterogeneous data source; carrying out feature recombination on the reference data by utilizing a generative neural network model, and generating discrete spatio-temporal data slices containing spatio-temporal anchor points and scene content attributes (such as environment illumination and material mapping); and finally, injecting the slices into the database to fill data holes. According to the method, a passive trigger mechanism and a structured reconstruction technology of unstructured data are introduced, so that the computing power cost of a large-scale virtual environment is effectively reduced, the problems of data sparseness and cold start of a long-tail region are solved, and the historical authenticity and time-space consistency of a virtual scene are ensured.

The invention discloses a table data analysis large model training and application method based on agent interaction reinforcement learning. Collecting multi-source data, generating a reference reply text, and converting to obtain multi-source word vector data; screening the multi-source word vector data through multiple times of repeated reasoning processing of a plurality of strong models; extracting data and inputting the data into a cold start model with preset weight parameters for training processing of supervised fine tuning of all parameters; inputting multi-source word vector data into the trained cold start model, and training by adopting a reinforcement learning method; and applying the cold start model trained by the reinforcement learning method to answer processing of text questions in an actual scene. According to the method, through reasonable multi-source data setting and matching, training process and algorithm design, the actual application effect of the large language model in the field of two-dimensional table data analysis is improved.

Embodiments of the present specification provide a data processing method and system, the method comprising: a reasoning framework, in response to a cold start instruction sent by a cloud platform, cold starting on a target graphics processor, sending an opening hijacking command to a hijacking module, and during the process of loading first model weights of a first reasoning model, initiating a video memory application to the target graphics processor; the hijacking module, in response to the opening hijacking command, hijacking the video memory application, and redirecting the video memory application to a first virtual address in a fixed virtual address space; the reasoning framework, in a case where it is determined that the first model weights are completed loading, sending an ending hijacking command to the hijacking module, and based on the first model weights in the first virtual address, constructing a first reasoning execution graph to execute a first reasoning task using the first reasoning execution graph. By hijacking the CUDA video memory allocation and redirecting it to a pre-reserved fixed address, the address of the model weights is ensured to be persistent and stable, so that the CUDA Graph continues to be effective in multiple loadings and hot switching.