31 results about "Behavior space" patented technology

The invention discloses a causal chain tracking method and system for cross-space APT attack behaviors, and relates to the technical field of network security. The method comprises the following steps: collecting heterogeneous device logs crossing multiple network domains, extracting original attack data, performing spatial-temporal feature extraction, and generating a device behavior spatial-temporal vector; analyzing the original attack data to obtain attack tactical semantic features, fusing the equipment behavior space-time vector and the attack tactical semantic features, and constructing a three-dimensional association map; and performing attack path reasoning on the three-dimensional association map based on a causal discovery algorithm, identifying a causal dependency relationship, performing data tracking according to the causal dependency relationship, reconstructing an APT attack causal chain, and obtaining an attack logic path. The technical problem that in the prior art, comprehensive and accurate traceability analysis on cross-space APT attacks is difficult is solved, and the technical effects of efficient reconstruction of APT attack causal chains and accurate identification of attack paths are achieved.

The invention relates to an electronic pet multi-modal interaction system based on artificial intelligence. The system comprises a data acquisition module, a multi-modal emotion analysis module, an emotion state modeling module and a dynamic interaction module. The data acquisition module is used for acquiring physiological electric signals, voice data and behavior track information of a user. The multi-modal emotion analysis module adopts a fusion model to carry out correlation analysis on time domain features of physiological signals, semantic features of voice and behavior space features, and generates multi-dimensional emotion feature vectors. The emotional state modeling module performs weighted fusion on the physiological indexes, the voice semantics and the behavior frequency in the multi-dimensional emotional feature vectors based on the LSTM network to generate emotional quantization parameters. And the dynamic interaction module generates an emotion interaction action instruction based on the user interaction intention in combination with the emotion quantization parameter and a preset dynamic interaction mechanism.

The invention discloses an attack trapping and analyzing method and system based on a honeynet, and relates to the technical field of attack analysis, and the method comprises the steps: collecting attack interaction data and an associated attack chain trajectory set; dividing attack behavior units, and constructing a behavior space attitude matrix; constructing a two-dimensional behavior relationship graph, performing interpolation enhancement of sub-time granularity, and raising the dimension of the interpolated behavior node into a three-dimensional semantic space; performing attack path trend analysis on the attack behavior unit, and constructing an attack path bending rule model; and inputting parameters such as the attack chain trajectory set and behavior nodes in the three-dimensional semantic space into the attack path bending rule model, outputting a trapping response strategy, and dynamically adjusting the honeynet environment. According to the invention, the type and position of the honeypot can be automatically adjusted according to the attack behavior change, the structure definition and processing precision of the attack behavior data are effectively improved, and the attack behavior identification result can be conveniently and directly used for trapping strategy optimization.

The invention relates to an implementation method and system based on multi-agent strategy diversification, and belongs to the technical field of multi-agent reinforcement learning. The method comprises the following steps of: firstly, performing strategy updating by adopting a multi-agent near-end strategy optimization framework in an exploration stage, and realizing global value evaluation and local action generation through a centralized training and decentralized execution framework; meanwhile, calculating a variable coefficient of a reward sequence through a sliding window, dynamically detecting a strategy convergence condition, and if the strategy is converged, collecting a current exploration strategy into a historical strategy library; and finally, based on a Stein variational gradient descent algorithm, a strategy particle set is extracted from a historical strategy storage library, the strategy is driven to be optimized to an unexplored behavior space, and local optimum is broken through. According to the method, the diversity, the adaptability and the calculation efficiency of a multi-agent strategy are remarkably improved through self-adaptive convergence detection, a historical strategy storage library framework and a complete framework integrating MAPPO and SVGD.

The invention discloses a code virtualization obfuscation method based on reinforcement learning, and relates to the technical field of code protection. Comprising the steps of 1, selecting an obfuscation method in a behavior space for a key code file in a business application scene by using a reinforcement learning agent to perform code obfuscation on the key code file, 2, evaluating the obfuscation method selected by the reinforcement learning agent from two aspects of obfuscation effect and performance overhead expression, and evaluating the obfuscation effect by using an NCD algorithm. Step 3, evaluating the performance overhead performance of the obfuscation method by directly comparing the running time and the volume of the executable program before and after code obfuscation, and obtaining an evaluation result, and step 3, providing the evaluation result as a feedback reward to the reinforcement learning agent, guiding the reinforcement learning agent to select the obfuscation method in the next iteration, and obtaining the obfuscation result. And circularly selecting the obfuscation method and evaluating the obfuscation method until the obfuscation method meeting the obfuscation intensity and performance overhead performance under the business application scene is obtained, and applying the obfuscation method to the obfuscation key code file.

The invention relates to the technical field of privacy protection, and discloses an AI behavior recognition-based screen information candid shooting prevention monitoring device, which comprises a behavior acquisition module, an image processing module, an image processing module, an image processing module, an image processing module and an image processing module, and is characterized in that the behavior acquisition module is used for acquiring user eye movement, head posture, body orientation and screen sight heat map information; the feature representation module is used for carrying out fusion processing on the collected multi-modal behavior data in a preset time window and constructing behavior representation; the embedding analysis module is used for carrying out nonlinear manifold embedding on the behavior representation; the information potential modeling module is used for constructing a potential information leakage risk field in the low-dimensional behavior space; the behavior intention recognition module is used for evaluating whether the user has a secret photographing intention or not according to the risk field and the embedded representation; and a screen protection module. By constructing a behavior representation mode of nonlinear manifold embedding and carrying out compression expression on high-dimensional multi-modal behavior data, geometric space representation better conforming to the potential intention of a user is obtained, and the recognition accuracy of abnormal behaviors such as candid photography is improved.

The present application relates to the technical field of abnormal behavior recognition, and more particularly to an abnormal behavior monitoring and early warning method based on a space-time graph neural network, medium and equipment, which converts abstract behavior data into a graph structure containing three kinds of nodes and three kinds of edges by extracting the object-space-time ternary relationship contained in the behavior data of the target monitoring object, and quantifies the correlation strength through the edge weight, and carries out deep feature extraction and embedding representation for the object, space and time nodes respectively, converts the abstract graph node into a high-dimensional vector rich in semantic information, and gradually integrates the spatial semantic features and the time dynamic features into the object behavior features, and finally generates a space-time enhanced representation vector that can reflect the individual behavior mode of the object and the influence of the space-time environment, breaking through the limitations of a single object node graph, solving the problems of insufficient behavior space-time context modeling and fragmented processing of spatial features and time features, and realizing accurate identification and early warning of space-time coupled abnormal behavior.

The invention discloses a multi-scale spatio-temporal frequency behavior recognition method based on a point cloud sequence, and relates to the technical field of behavior recognition, and the method comprises the steps: introducing two modules, namely a multi-scale spatio-temporal feature extraction module and a spatio-temporal feature frequency enhancement module; the multi-scale spatio-temporal feature extraction module is used for decomposing an original point cloud sequence in a spatial dimension and a time dimension, and enriching low-level semantic information of the spatial dimension by adopting multi-scale spatio-temporal convolution; and the space-time frequency enhancement module adaptively learns channel information of high-level semantics based on frequency information of a time dimension, so that the discrimination ability of behavior space-time characteristics is improved. Experimental results show that compared with a conventional mainstream point cloud-based method, the method is remarkably improved, and the recognition precision is also superior to that of most methods in other directions.

The invention discloses a multi-outlet neural network adversarial robustness evaluation and defense method and device based on the game theory, and aims to solve the problems of attack-defense mismatching and evaluation limitation in an existing evaluation method. By introducing a game theory thought, attack and defense confrontation of a multi-exit neural network is modeled as a double-player zero-sum game, and behavior spaces of attackers and defenders are defined as a combined set of network exits. Firstly, an approximate revenue matrix is constructed through an attack experiment of a data set and is used for quantifying the revenue relation of an attacker and a defender under different exit combinations; and based on the income matrix, calculating an optimal strategy of an attacker, and selecting an exit combination by maximizing expected income to generate a confrontation sample. And the defender minimizes the maximum loss possibly caused by the attacker by solving an optimal strategy, finally selects an outlet combination according to the probability for reasoning, and takes an average value output by each outlet as a final prediction result.

The invention relates to the field of artificial intelligence, robot technology and aided navigation, in particular to a multi-robot collaborative navigation system based on intention reasoning, which constructs a Riemannian manifold multi-modal feature space based on a differential geometry theory, maps a voice instruction, a gesture operation and an emotional state of a user to a unified feature space, and provides a multi-robot collaborative navigation system based on intention reasoning. In combination with historical behavior data of a user, the user intention is accurately understood, the system adopts a mapping theory in differential geometry to establish a mapping relationship between an intention space and a behavior space, smooth transition from an abstract intention to a specific behavior is realized, and an optimal navigation path is calculated on a joint manifold formed by intention-environment-emotion through a geodesic optimization algorithm. The system can dynamically adjust the navigation strategy according to the emotional state of the user and provide personalized navigation experience, the multi-robot collaborative navigation unit decomposes the navigation strategy into a plurality of subtasks, evaluates the ability index of each robot, optimizes the distribution scheme, realizes the collaborative work of multiple robots, and improves the intention understanding accuracy.

The application belongs to the technical field of data analysis, and specifically discloses a user behavior analysis method, system and program product based on big data. Behavior data of a user in each time period is collected based on big data technology and processed to construct a behavior data matrix by using quantitative behavior parameters. The behavior data matrix is then converted and processed to obtain a standard data matrix. Based on the standard data matrix, convergence calculation and behavior space matrix construction analysis are performed to determine the highlight index of each behavior type. Finally, several behavior types with the highest highlight index are selected as highlight behavior types to construct a user behavior portrait. The application can efficiently and accurately analyze the highlight behavior types of a user based on the behavior data of the user in each time period, thereby constructing a high-matching-degree behavior portrait for the user. This can improve the user experience by accurately classifying and providing targeted services for the user in the future.

The present invention relates to the field of space manipulator control technology, and more specifically to a homotopy-based meta-reinforcement training method for grasping and placing tasks for space manipulators. The method comprises: obtaining a meta-task behavior space, a meta-task state space, and a meta-task reward function; constructing a homotopy task sequence for meta-reinforcement training; obtaining updated SAC neural network parameters for each homotopy task, obtaining updated target meta-neural network parameters, and an updated target neural network model for the Meta-SAC algorithm; and updating the manipulator's current state to its next state. Compared to traditional reinforcement learning and transfer learning methods, the method achieves a higher success rate and faster convergence speed, and exhibits superior performance in the field of space manipulator control.

The invention belongs to the technical field of Internet platform security, and discloses an operation analysis system for an industrial Internet platform, which comprises a login analysis module used for acquiring login data, extracting login behavior characteristics and generating login credibility; the operation verification module is used for monitoring an equipment operation request and judging compliance, and if the equipment operation request does not meet the rule, behavior consistency analysis is carried out; the behavior feature analysis module is used for acquiring a user operation behavior space and generating a behavior consistency degree; and the security response module is used for executing corresponding security disposal measures according to the behavior consistency and the login credibility. A traditional binary judgment model is replaced by quantizing the login credibility, multi-level safety response is achieved in combination with operation behavior analysis, the problem of login error interception in the industrial environment is solved, the defect that an existing system is insufficient in operation behavior monitoring is overcome, and balance between safety protection and production efficiency is achieved.

A behavior center simulation model construction method based on a meta-model comprises the steps that a micro-behavior meta-model is constructed, the micro-behavior meta-model comprises control class meta-model elements and function class meta-model elements, and the micro-behavior meta-model is constructed based on the function class meta-model elements and the control class meta-model elements in the micro-behavior meta-model; a microscopic behavior model comprising three behavior levels is constructed, and a simulation model with the microscopic behavior model as the center is constructed. According to the method, behaviors in a functional component are decoupled from an entity model, a meta-model technology is applied, full-behavior space modeling is carried out in a low-code form, the behavior modeling efficiency is improved, meanwhile, functions and control of the behaviors are decoupled in a behavior structure, modeling of behaviors of any granularity and combinations of behaviors of different granularities can be supported technically, and the behavior modeling efficiency is improved. The reusability of different levels of behaviors in the simulation model is solved, and the development efficiency of the simulation model is also improved.

This invention relates to a file loading method, apparatus, electronic device, and storage medium, belonging to the field of file loading technology. The method includes: monitoring user historical browsing behavior and setting different reporting methods and timings; obtaining user historical browsing trajectories by monitoring user historical browsing behavior, cleaning and analyzing the embedded data, classifying user historical browsing behavior, and statistically confirming user actual needs; mapping user actual needs to a state space and behavior space in reinforcement learning, predicting user behavior, and outputting the user's demand weights for different resources; generating an initial resource loading sequence based on the user's historical browsing trajectory, combining the predicted user behavior with the weights under different user behaviors, and performing multiple iterations through a user behavior prediction module to obtain a resource loading order sequence. The method, apparatus, electronic device, and storage medium of this invention reduce file loading time.

The invention provides an intelligent research journey recommendation method and system based on collaborative filtering and LSTM fusion, and the method comprises the steps: taking a multi-modal behavior time sequence as the input of a pre-constructed research behavior space-time attention network, and obtaining a time sequence preference implicit vector; calculating the matching degree of the time sequence preference implicit vector and the dynamic constraint factor, and correcting the time sequence preference implicit vector by using the constraint factor feature vector based on a calculation result to obtain a scene adaptive time sequence preference vector; extracting a time sequence behavior feature vector of the current research journey group, and performing weighted fusion on the time sequence behavior feature vector and the scene adaptive time sequence preference vector to obtain a final time sequence preference vector; the final time sequence preference vector is used for outputting a personalized research and study journey recommendation list. According to the multi-dimensional intelligent recommendation method, the group preference, the personal time sequence behavior, the research education attribute and the real-time resource state can be fused, so that the recommendation accuracy and the user experience are improved.

The invention discloses a large language model recommendation method based on collaborative similarity alignment. The method comprises the following steps: S100, a similarity alignment stage: respectively extracting vector representations of articles from language space embedding generated by a large language model and behavior space embedding generated by a collaborative filtering method; s200, a quality evaluation stage: performing quality evaluation on the behavior space embedding based on a statistical view angle and a collaborative view angle, and screening out a high-quality collaborative signal meeting a pre-auditing condition, and S300, a joint training stage: performing joint optimization on a similarity alignment loss and a standard training target of a language model, and completing end-to-end model training to generate a recommendation result. According to the method, more essential collaborative similarity signals are explored to guide large model learning recommendation knowledge, information loss is avoided, noise data can be accurately screened through the introduced quality evaluation module, only high-quality collaborative signals are reserved, and therefore the accuracy and generalization of recommendation results are improved.

The invention discloses a multi-behavior user intention unwrapping representation learning method based on information bottleneck, which effectively separates pseudo-correlation intentions in multi-behavior interaction by using an information bottleneck principle so as to improve the performance of a recommendation system. Specifically, the method comprises the following steps: firstly, introducing a time-sensitive pseudo-correlation coefficient, dynamically measuring a pseudo-correlation intention proportion of a user in an auxiliary behavior, and using the coefficient to guide a subsequent multi-intention learning task; secondly, pre-training the multi-behavior interaction data of the user and the article by adopting a graph convolutional neural network so as to learn embedded representation under high-order connectivity; then, embedding and mapping the auxiliary behaviors into a target behavior space by utilizing an orthogonal projection technology, so as to separate a real related intention and a pseudo related intention; and finally, after real related intention embedding and pseudo related intention embedding of the user are obtained, multi-intention learning is carried out based on an information bottleneck principle, under the guidance of a pseudo related coefficient, a pseudo related intention in an entanglement auxiliary behavior is deentangled in a targeted manner, and meanwhile, the real related intention is reserved and transmitted to a target behavior as far as possible, so that the user experience is improved. Therefore, the modeling capability and recommendation effect of the recommendation system on the target behavior of the user are improved.

The application provides a multi-element information collaborative decision method based on intention recognition for an autonomous underwater robot in a weak communication and partial observation environment, so as to improve the adaptability and intelligent decision-making ability of an AUV group in a complex dynamic environment. Specifically, the application provides an intention recognition method based on decision consensus. Firstly, a decision consensus among AUVs is constructed through fuzzy reasoning, and key features such as a packet loss rate, a distance, and a number of partners are mapped to a behavior space, so that a preliminary intention inference is efficiently realized with a minimum amount of data exchange. Subsequently, a residual convolutional neural network is used to further strengthen feature extraction, and precise intention prediction is realized. Finally, by completing global state information and optimizing decision input, a dynamic perception-intention reasoning-collaborative decision closed loop is constructed, and more efficient task planning and cooperation efficiency of the AUV group are realized.

The present disclosure relates to the field of animal behavior and pattern recognition technology, and particularly relates to an animal behavior abnormal pattern recognition method, system, device and medium. The present disclosure constructs a complex behavior pattern expression based on a probability generation model, greatly improving the adaptability and accuracy of the model. The present disclosure uses a Gaussian mixture model or other probability generation model to describe the "normal behavior space", rather than using a rigid fixed threshold. This enables the model to automatically learn and express various normal sub-patterns of animal behavior, such as foraging and resting, and their dynamic changes. This data-driven modeling approach can better adapt to behavior changes under different species, different seasons and different environments, avoiding the high false alarm rate problem caused by the "one-size-fits-all" approach of traditional rule engines, making the abnormality judgment more scientific and accurate.

The application discloses a Q network control method and device and a readable storage medium, and relates to the electronic information field. In the Q network control method provided by the application, a preset behavior is a behavior randomly selected in a behavior space A or an optimal behavior selected in the behavior space A by using a database operator. A plurality of experience samples are extracted from an experience pool, and corresponding expected benefits are calculated based on the database operator. Finally, the parameters and vector indexes of the Q network and the parameters and vector indexes of a target Q network are updated according to the expected benefits. The database operator is an operator realized based on an approximate nearest neighbor search technology. Compared with the previous training control method, the approximate nearest neighbor search technology is introduced in the scheme to obtain the database operator, so that the search range is greatly reduced. Compared with the previous DQN which traverses all the action spaces, the training and decision efficiency is increased, and the scheme can be applied to large-scale discrete behavior scenes.

The application provides an API behavior prediction and security policy management method based on machine learning, relates to the field of computer network security, and comprises the following steps: acquiring API original data in real time, preprocessing the original data to obtain preprocessed original data, acquiring API dimension features in real time based on the preprocessed original data, and fusing to obtain unstructured data; a dynamic characteristic validity verification rule is established to preprocess the unstructured data, and a unified dimension standardization feature tensor is obtained; the joint modeling of API behavior space-time features is performed through the collaborative architecture of a multi-head time attention mechanism and a dynamic graph neural network, API behavior prediction is performed, API dynamic security management and control strategies are generated in real time through a strategy engine, strategy execution is performed, the security strategy execution effect is monitored in real time, and iterative optimization is performed. The application solves the problems of fixed and rigid security management and control strategies of a traditional API gateway, and the difficulty in dynamically adjusting security strategies according to real-time access conditions and behaviors.

A behavioral data generation device according to one aspect of the present invention is provided with an acquisition unit and a generation unit. The acquisition unit acquires behavior description information describing a person's behavior, including action information relating to the person's actions, attribute information relating to the person's attributes, and spatial information relating to a space in which the person acts. The generation unit uses a machine learning model, including a latent diffusion model, to generate behavioral data that has a skeleton structure and that simulates the person's behavior on the basis of the behavior description information.

The invention provides an information recommendation method, system and device and a medium, and belongs to the field of computers.The method comprises the following steps that basic data used for information recommendation are obtained, and a multi-modal knowledge graph is constructed through the basic data; the method comprises the following steps: acquiring an initial state s of a multi-modal knowledge graph selected by a user, calculating an attention weight of an adjacent neighbor of the user by using a hierarchical attention mechanism of a strategy network between the user and an article pair, generating an action a by using a ReLU layer and a TanH layer of the strategy network, and interacting with the multi-modal knowledge graph by an intelligent agent through the action a to obtain a multi-modal knowledge graph. And generating a reward r as a feedback signal, and recommending candidate items to the user by the intelligent agent according to the feedback signal. According to the method, the attention can be focused on the most interested article of the user, and the complexity of interaction between the intelligent agent and the environment in reinforcement learning is reduced, so that the number of reinforcement learning behavior spaces is greatly reduced, and the most effective information is quickly and accurately obtained.

The application discloses a kind of attack trapping and analysis method and system based on honeynet, it is related to attack analysis technical field, including: attack interaction data and associated attack chain trajectory set are collected;Divide attack behavior unit, construct behavior space posture matrix;Two-dimensional behavior relationship atlas is constructed, interpolation enhancement is carried out to sub-time granularity, and the behavior node after interpolation is upgraded to three-dimensional semantic space;Attack path trend analysis is carried out to attack behavior unit, and attack path bending law model is constructed;Attack chain trajectory set, the behavior node in three-dimensional semantic space and the like parameter are input into attack path bending law model, and output is trapping response strategy, and dynamically adjusts honeynet environment.The application can automatically adjust honeypot type and position according to attack behavior change, effectively improve the structural clarity and processing precision of attack behavior data, and facilitate to directly use attack behavior identification result for trapping strategy optimization.

The invention relates to the technical field of abnormal behavior recognition, in particular to an abnormal behavior monitoring and early warning method based on a space-time diagram neural network, a medium and equipment. According to the method, abstract behavior data is converted into a graph structure containing three nodes and three edges, association strength is quantified through edge weights, deep feature extraction and embedded representation are carried out on object, space and time nodes respectively, abstract graph nodes are converted into high-dimensional vectors rich in semantic information, and the high-dimensional vectors are subjected to feature extraction and embedded representation; the spatial semantic feature and the time dynamic feature are fused into the object behavior feature in stages, and the finally generated space-time enhanced representation vector can reflect the individual behavior pattern of the object and the space-time environment influence at the same time, so that the limitation of a single object node map is broken through; the problems of insufficient behavior spatio-temporal context modeling and separation processing of spatial features and time features are solved, and accurate recognition and early warning of spatio-temporal coupling abnormal behaviors are realized.

The present disclosure provides a method, device, electronic device, vehicle, and product for predicting a meeting trajectory. The method includes: in response to detecting an oncoming vehicle, traversing behavior samples in a pre-constructed driving paradigm behavior space, wherein the behavior samples include the vehicle's longitudinal acceleration and lateral offset parameters; each time a behavior sample is traversed, trajectory deduction is performed for the vehicle and the oncoming vehicle, respectively, based on the behavior sample and the current motion parameters of the vehicle and the oncoming vehicle, to obtain a trajectory pair for the vehicle and the oncoming vehicle; quantitative evaluation is performed on the plurality of trajectory pairs obtained after the traversal to determine a target trajectory pair that is suitable for the vehicle; and the vehicle is controlled to travel according to the deduced trajectory of the vehicle in the target trajectory pair. Accordingly, the present disclosure achieves trajectory prediction during meeting by traversing the driving paradigm behavior space and performing trajectory deduction, thereby improving the safety and traffic efficiency of autonomous vehicles during meeting.

The invention provides a timing training anti-cheating system and a timing training anti-cheating method based on a trusted root and a behavior space-time chain, which can fuse time, space and behavior multi-dimensional information, ensure credibility from a data source and realize all-around and high-reliability cheating prevention through intelligent association analysis, and comprises a timing training terminal with a built-in trusted execution environment TEE module; the trusted execution environment TEE module is used for generating environment digital fingerprints and encrypting and signing the collected vehicle state sensing terminal data; the system further comprises a cloud collaborative judgment platform which is used for receiving a signed chain type data packet, and the chain type data packet comprises a plurality of data units which are linked according to a time sequence and signed by TEE; based on a preset physical rule model, performing logic consistency verification on the spatio-temporal trajectory data and the vehicle behavior data in the chained data packet; and judging whether simulation positioning cheating exists through a multi-mode causal analysis module.

The invention provides an intelligent terminal network intrusion behavior identification method and device based on deep learning, and the method comprises the steps: building a dynamic behavior space-time atlas comprising a network address node and an internal operation node through obtaining the original network flow and operation log data of a terminal in a preset time period; in the map, a behavior evolution relationship is represented through a time sequence edge, and co-occurrence association of the network and an operation behavior is represented through an association edge. And inputting the atlas into a pre-trained intrusion behavior prediction deep network, simulating propagation evolution of intrusion on the atlas by using a space-time propagation prediction layer, and outputting intrusion probability distribution of each node and an intrusion propagation path graph among the nodes. And performing intrusion behavior sequence reconstruction based on the result, generating a candidate intrusion behavior sequence, calculating a credibility score, finally screening out a target sequence with the highest score, and generating a response instruction containing a terminal identifier and an intrusion time period. According to the method, hidden intrusion behavior modes can be deeply mined, and a complete attack chain can be reconstructed.

The application relates to the technical field of edge AI computing systems of graph neural networks, and discloses a behavior space-time prediction method and system of an edge AI chip. The method comprises the following steps: extracting customers and shelf entities in real time based on a video stream, and constructing a dynamic space-time node set; generating a sparse heterogeneous space-time high-order correlation structure based on the node set; and performing behavior prediction and loss prevention response with topological perception based on the structure through a topological graph neural network after knowledge distillation. The application is aimed at solving the problems of high memory occupation in traditional unstructured data processing, difficulty of traditional graph structure in capturing multi-body interaction, low behavior discrimination precision of an edge model, and lag of a loss prevention response. Through model quantization pruning, dynamic sparse pruning, and multi-dimensional loss prevention evidence fusion, the application adapts to resource constraints of an edge chip, improves the fine behavior discrimination precision and real-time response speed, and realizes accurate loss prevention and shelf layout optimization in an unmanned sales scene.