Cognitive-driven crowd sensing user-task relationship dynamic analysis method

Abstract

How to adaptively adjust a task allocation strategy in a large-scale, multi-position and high-precision sensing environment is the key of providing sustainable trust services by a sensing platform. In order to solve the problem, the invention provides a cognitive-driven crowd sensing user-task relationship dynamic analysis method, which is helpful for a system to sense user behaviors in real time and adaptively realize accurate matching between the user and a sensing task. Firstly, in order to reasonably describe a user participation motivation, a user-task internal relation is mined by using a causal decision technology; secondly, designing a self-alarm module, a self-control module and a self-healing module, and iteratively and circularly constructing a complete user cognitive causal graph; finally, user participation behaviors are dynamically perceived based on a causal graph and fed back and regulated in time, and a causal remodeling process is introduced to filter confounding factors, so that a filtering foam phenomenon is effectively avoided while user participation willingness is improved, and adaptability and accuracy of a crowd sensing task allocation stage are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of mobile crowd intelligence perception, and in particular relates to a cognition-driven crowd intelligence perception user-task relationship dynamic analysis method. Background Art

[0002] Mobile Crowd Sensing (MCS) is an emerging Internet of Things technology. It uses a large number of ordinary users' mobile devices as sensing units, collects data through built-in sensors, and uploads the data to the central service for processing and analysis through the network. This model can effectively complete large-scale and complex perception tasks, and provide data support and intelligent services for many fields such as environmental monitoring, traffic management, public health, social networks, etc. Reasonable task allocation can ensure that the right participants perform tasks at the right time and place, thereby maximizing resource utilization and ensuring the quality of perception data. Therefore, the task allocation mechanism is an indispensable key factor in determining whether the MCS platform can obtain good service quality.

[0003] Most existing works adopt a platform-centric model and design incentive mechanisms from the perspective of game theory to study the task allocation problem. The main process is: users submit bids that reflect their willingness or cost to participate, and then the platform selects a set of users and their payments based on various target constraints to maximize benefits. However, such models usually assume that users are rational and do not comprehensively consider users' abilities, preferences, and reliability, which may lead to failure in the allocation process. This assumption and the fact that each user's preferences can be abstracted into the setting of a single bid parameter may be an oversimplification of complex user behavior. In MCS, user perception is a key factor in the allocation stage, so more and more user-centric models have been proposed. These studies have improved the task acceptance rate and allocation efficiency to a certain extent by giving users more freedom to choose tasks of interest.

[0004] Based on the perception scenario, which is divided into offline or online, the existing task allocation methods are mainly divided into two categories. In offline scenarios, the platform usually assumes that the user is available during the entire perception period, and obtains information such as the user's history, device type, and participation capabilities in advance to plan the user's participation behavior. This static task allocation method tends to be overly dependent on historical interaction data, and the understanding of user participation willingness and behavior patterns is not in-depth enough. At the same time, in most application scenarios, it has weak adaptability to dynamic environmental changes, and the practicality and flexibility of task allocation strategies are limited. Figure 1As shown in (1), by deeply investigating the user background of a, it can be found that the user's mobility pattern is random and the location information is difficult to accurately predict. At the same time, when a is in different communities, the user's task preferences and willingness to participate vary greatly due to the influence of different friendship relationships. Therefore, a single, static task allocation cannot adapt to the real-time changes in user behavior and environment, and lacks an effective feedback mechanism to adjust the task allocation strategy in real time, which will lead to reduced task allocation efficiency.

[0005] Based on the above problems, a large number of dynamic task allocation methods for online scenarios have emerged. It enables the system to make instant allocations based on the user's real-time status and behavior patterns, improving the platform's work efficiency and response speed. Some work incorporates user task feedback into model updates to achieve dynamic iterative modeling of user data, which improves the adaptability of task allocation algorithms to a certain extent. However, feedback loops are not always beneficial, as they may encourage narrower and narrower content exposure over time, which may lead to filter bubbles if not paid attention to. Figure 1 As shown in (2), the platform believes that users prefer to participate in tasks that are repeated frequently in a short period of time, and this cognition is biased. As the feedback loop continues to learn, this bias will be further amplified, eventually leading to the convergence of task allocation results and reduced user participation enthusiasm.

[0006] Causal reasoning is a thinking process that analyzes and explains causal relationships, aiming to understand the causal relationship between different variables or events. Causal reasoning can provide a deeper understanding of the behavior and patterns of complex systems and provide stronger support for decision-making. In addition, the concept of metacognitive control process was proposed by American psychologist John Flavel, who defined it as the ability to monitor, regulate and control individual cognitive activities, and the understanding and management of self-cognitive processes. The metacognitive control process helps individuals to effectively monitor and regulate themselves in complex cognitive tasks, thereby improving cognitive execution ability and task performance. Summary of the invention

[0007] In view of the above existing research methods and existing problems, the present invention proposes a cognitive-driven crowd intelligence perception user-task relationship dynamic analysis method. By using causal decision-making technology to dynamically perceive the changes in user-task relationships, and guiding user behavior based on a feedback loop mechanism. This enables the crowd intelligence system to dynamically capture the user's real preferences, thereby selecting tasks that are easy for users to accept for allocation, and effectively improving the task completion rate and data collection efficiency.

[0008] First, by constructing a cognitive causal graph to infer the causes of the intrinsic connection between users and tasks, it is helpful to select tasks that are easy for users to accept for assignment, thereby improving the task completion rate and data collection efficiency. At the same time, a feedback loop mechanism is introduced to dynamically model user data and realize the adaptive allocation of perception tasks. When facing the filter bubble phenomenon, the concept of cognitive plasticity causal graph is introduced to simulate the human perception and decision-making process. Figure 2 As shown in the figure. Among them, the self-warning module is the premise of the plasticity of the cognitive causal graph. When the filter bubble phenomenon exists in the detection perception platform, it will trigger the reconstruction response of the causal graph at the previous moment. Then the self-control module performs key subgraph detection on a macro level to remove confounding factors, so as to reduce the computational cost when reconstructing the causal graph and ensure the accuracy of the causal graph. Finally, the self-healing module improves the graph link prediction process by introducing counterfactual reasoning technology, and enhances the user-task relationship by using the global graph structure and the causality of the link, further ensuring the cognitive plasticity of the causal graph. The cognitive causal graph is reshaped based on self-warning, self-control and self-healing loop feedback, so as to fully extract and mine the relevant attributes of the user-task relationship in the task allocation stage. The impact of users' real interests on task participation is analyzed to provide accurate task allocation results consistent with user preferences.

[0009] Compared with the prior art, the beneficial effects of the present invention are: (1) For the first time, the concept of cognitive plastic causal graph is proposed. By introducing "self-warning", "self-control" and "self-healing" in metacognitive control, a sustainable cyclic learning process is formed, which effectively ensures the dynamic and flexible mapping of user-task causal relationship in mobile crowd intelligence perception. (2) An active perception adaptive task allocation method is proposed, which mines the user's real task preferences in a fine-grained manner through the causal graph, and introduces a feedback mechanism to perceive the changes in user behavior patterns in real time, and fully understands the dynamic data generation process, thereby effectively improving the flexibility of task allocation. (3) In response to the filter bubble phenomenon caused by the convergence of task lists, a causal reshaping process is introduced to ensure the advancement of cognitive graphs. By learning the key subgraph structure, the user-task causal relationship is enhanced, and the interference of confounding factors in updating the causal graph is filtered out, so as to achieve accurate matching of perceived users and perceived tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is an existing challenge in the task allocation stage;

[0011] Figure 2 It is a cognitive plastic causal diagram;

[0012] Figure 3 It is the process of building a causal graph;

[0013] Figure 4 It is a self-alarm triggering process;

[0014] Figure 5 It is a self-control process;

[0015] Figure 6 It’s a self-healing process. DETAILED DESCRIPTION

[0016] Analyzing user behavior to extract causal relationships is the key to improving user participation in tasks. At the same time, users’ rich social relationships can be used as a supplementary factor to improve the user model. Therefore, we first conduct heterogeneous graph representation learning based on user-task historical interaction records and social relationships, and construct causal graphs by simulating human perception logic. Figure 3 As shown, first, semantic information is obtained based on the relationships and meta-paths derived from the heterogeneous graph pattern. For example, when the target node type is given as "task", semantic information such as "co-participants" can be obtained through the relationship "user-task" and the meta-path "user-task-user" respectively. In order to maintain the integrity of the target node context, all meta-paths starting from the target node type within the specified length limit are taken into account. Therefore, node n consists of three parts: node feature x, label y and meta-path z. However, due to the heterogeneity of the graph, these variables usually have different forms and may exist in different feature spaces. For example, the x and y labels are usually vectors of different dimensional sizes, while z is a set of vectors. In order to solve the above problems, all these variables need to be embedded in the same representation space. For node n i , its eigenvector x i and label vector y i It is encoded by two linear transformations as follows:

[0017] Then, for node n i The j-th meta-path variable is obtained by using the NeighborSetEncoder function j Encode the corresponding neighbor set into a fixed-size vector representation Right now:

[0018] At the same time, in order to avoid introducing confounding factors into the causal graph, which would lead to a decrease in the accuracy of the algorithm. x 、Linear y and NeighborSetEncoder j The encoders need to remain independent of each other to avoid false correlations between fitting variables. Then, based on causal reasoning technology, the causal relationship between the above semantic variables is mined and a cognitive causal graph is constructed. First, the trainable matrix is ​​initialized represents a causal graph, where q is the number of meta-paths, A i,j Indicates h i Is caused by jThe probability of one of the causes being the same. Note that since a variable cannot be its own cause, the diagonal of the matrix A is constrained to be zero, i.e., A i,j = 0. Based on matrix A, the variable h k The structure allocation function f k The definition is as follows:

[0019] At the same time VariableDecoder k It is to reconstruct the variable h based on the obtained causal relationship k The encoding function of EffectEncoder is modeled and learned by using a multilayer perceptron (MLP). ik and VariableDecoder k :

[0020] Building cognitive causal graphs based on historical user data usually involves a feedback loop process, incorporating user feedback into model updates, and accurately controlling user behavior and precisely regulating it by fitting historical interaction data. However, feedback loops are not always beneficial. Because over time, more and more single tasks may be encouraged to be exposed to users, leading to the filter bubble phenomenon. Therefore, the key to improving perceived quality lies in understanding when filter bubbles will occur and how to mitigate the negative effects of filter bubbles.

[0021] Inspired by the theory of metacognitive control, a self-alert module is designed to detect the severity of filter bubbles in the perception platform. When the user's participation enthusiasm and perceived quality decrease significantly, the adaptive reconstruction of the cognitive causal graph in the next stage is triggered. The system dynamically perceives the changing trend of the user's motivation for task participation, eliminates the confounding factors that are irrelevant to the perceived user's participation, and realizes the precise control of the user-task relationship. The process is as follows: Figure 4 As shown. During different time periods when the perception platform releases tasks, the system dynamically monitors the types and quantities of the perception user's task list in real time and makes decisions on whether to control the filter bubble phenomenon. The severity of the filter bubble is mainly judged from the two aspects of diversity and isolation. The related concepts of the indicators are as follows:

[0022] Diversity: It is represented by counting the number of categories in the perception task list.

[0023] Isolation: Isolation is often used in sociological research to measure the degree of isolation between different user groups. Given two user groups a and b, the isolation is calculated as follows:

[0024] The above indicators are calculated through the heuristic rules designed by the crowd intelligence perception system, based on the severity of the filter bubble phenomenon and the classification. Then, the severity is presented to the perception system. When the system detects that the perceived user task list tends to be homogenized and the quality of the perceived data cannot be effectively guaranteed, the self-control and self-healing process of the next stage of cognitive map is triggered. By imitating the human thinking decision-making process, the model is expected to be endowed with the ability of autonomous cognitive learning and metacognitive control and regulation.

[0025] When filter bubbles occur in the detection perception platform, it is necessary to rebuild the causal graph to explore new user participation motivations and increase task diversity. However, it takes a lot of time to re-perform causal reasoning based on heterogeneous graph patterns. This is because only some factors no longer affect user interest behavior, and it is meaningless to globally search and reconstruct the causal graph. Therefore, Figure 5 As shown in the figure, by designing an automatic control module, key subgraph detection is performed based on the original heterogeneous graph. By discovering the key structure in the graph data, that is, the subgraph that plays a leading role in the causal graph, the time cost is effectively reduced while filtering out interference factors.

[0026] First, we design a subgraph detection module. It learns to find key subgraphs {g i}. At the same time, the detected subgraphs are further collected by determining the local neighborhood of the central node. However, due to the arbitrary size of the graph and the rich dependencies, it is difficult to directly sample some nodes and detect key subgraphs. Therefore, we first sample a set of seed nodes. As the central node of the subgraph, the subgraph around the central node is then sampled. In addition, the subgraph sampling task is decomposed into two subtasks, and deep agent and neighbor agent learning strategies are designed to solve them. For each central node v i , by using a 2-layer Graph Neural Networks (GNN) encoder to convert node features into node encodings as initialization, and then detect the surrounding subgraphs v i :

[0027] After obtaining the key subgraph {g i}, merge them into sketch G g =(V g ,E g ), and use the attention mechanism to learn subgraph embedding. Specifically, each subgraph is regarded as a supernode in the sketch, V g The connectivity between the two subgraphs is determined by the number of common nodes between the two subgraphs. i and g j The number of common nodes in is greater than the constant bc When i,j will be added to the edge set of the sketch. Then we learn the subgraph representation from two perspectives: within the subgraph and between subgraphs. For the within-subgraph, we add each subgraph g i Input into the GNN-based encoder δ to obtain the embedding of the nodes in the subgraph. Then g i The nodes in the H(g i ) can be calculated by the following generalized equation: H(g i )=δ(g i )={h i (v j )|v j ∈V(g i )}

[0028] Then, the inter-subgraph attention mechanism is adopted to learn the interactions between subgraphs from their embedding vectors. Specifically, through the multi-head attention mechanism, g j Medium i The attention coefficient α ij , the subgraph represents It can be obtained by guiding the information aggregation through its neighbor subgraph attention:

[0029] In order to obtain a discriminant graph representation containing more information, the bidirectional mutual information maximization mechanism is used to further enhance the representation of the subgraph from the global perception and label perception levels. Global-aware information maximization aims to encourage subgraph representations to capture global properties of the entire heterogeneous graph. Its goal is to make local subgraph representations and the global graph representation z G Maximizing the mutual information between them. Label-aware information maximization aims to encourage subgraph representations to be different in graph samples. Its goal is to maximize the subgraph-label pair on a given graph G. Finally, the learned enhanced subgraph representation is used as the graph representation of the original heterogeneous graph and as the input of the self-healing module in the next stage. The graph structure data processed in the self-control stage not only contains the key causal information that reflects the user's true preferences, but also effectively filters out noisy information and low-quality perception users, further ensuring the data reliability of the perception platform.

[0030] The key subgraph {g} processed in the self-control stage still contains confounding factors, which leads to a decrease in the accuracy of user-task causal relationship recognition. For example, user a is attracted by the high-reward incentive mechanism and has rich historical task interaction records, but uploads low-quality perception data to damage the interests of the platform. At the same time, because the model captures a's task preferences based on facts, it will trust users and cause more tasks to be assigned to a, which cannot effectively guarantee the reliability of platform data. Therefore, by designing a self-healing module, such as Figure 6 As shown in the figure, false correlation factors are eliminated based on counterfactual reasoning. For example, by observing the willingness of a to participate when the task reward is reduced, the cause of the connection between users and tasks is further analyzed. And learning to predict the missing causal links on the graph is expected to further improve the embedding representation learning process of the graph structure.

[0031] First, since the existence of a link between a given user-task node pair and the counterfactual treatment is unknown, the present invention uses the most recently observed contextual results as a substitute. By finding the nearest neighbor node of the node pair, each observed node pair is treated in the opposite way, and the result of the nearest neighbor is used as the counterfactual link. Specifically, The corresponding counterfactual link (n a ,n b )for:

[0032] Simultaneous counterfactual processing matrix T CF and the counterfactual adjacency matrix B CF Defined as:

[0033] Then, link prediction is performed based on the fact processing matrix T and the adjacency matrix B. CF As a supplementary input. And use the counterfactual result B CF As the training target of counterfactual data samples. The model consists of two trainable components: graph encoder f and link decoder g. The graph encoder generates the embedding vector of the node from the graph data, and the link decoder projects the representation vector of the node pair into the link prediction strategy. The graph encoder f selects the GCN model, and each layer is defined as:

[0034] At the same time, the output of the last layer of the encoder is expressed as As the H-dimensional representation vector of the node, X represents the feature matrix of the graph.

[0035] The link decoder g for predicting whether there is a link between a pair of nodes is implemented based on a multilayer perceptron and is defined as:

[0036] Finally, the user's true preference is inferred based on the reshaped cognitive causal graph, and the perception task that meets the user's expectations is assigned to the perception user.

[0037] Although the above describes the specific implementation mode of the present invention in conjunction with the drawings, it is not intended to limit the scope of protection of the present invention. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and should be included in the scope of protection of the present invention.

Claims

1. A cognitive-driven crowd-sensing user-task relationship dynamic analysis method, characterized in that: It includes cognitive causal diagram construction, self-warning module, self-control module and self-healing module.

2. The method for dynamic analysis of user-task relationships based on cognitive-driven crowd intelligence perception according to claim 1 is characterized in that: Based on the user's historical task records and social relationships, a cognitive causal graph is constructed to mine user preferences, and combined with existing demand information and other contextual features, adaptive decisions are made on task allocation results that suit the user's preferences.

3. The method for dynamic analysis of user-task relationships based on cognitive-driven crowd intelligence perception according to claim 1 is characterized in that: By designing a self-alert module, the severity of filter bubbles in the perception platform is detected, and the changing trend of user motivation for task participation is dynamically perceived. When the user's participation enthusiasm and perceived quality decrease significantly, the adaptive reconstruction of the cognitive causal graph is triggered.

4. The method for dynamic analysis of user-task relationships based on cognitive-driven crowd intelligence perception according to claim 1 is characterized in that: By designing an automatic control module to detect key subgraphs based on the original causal graph, the subgraphs that play a dominant role in the causal graph in the graph data are mined, and confounding factors that are irrelevant to the causes of perceived user participation are eliminated, effectively reducing the time overhead of graph reconstruction.

5. The method for dynamic analysis of user-task relationships based on cognitive-driven crowd intelligence perception according to claim 1, characterized in that: By designing a self-healing module to eliminate false correlation factors based on counterfactual reasoning, we further analyze the causes of user-task connections and learn to predict missing causal links on the graph to achieve accurate matching of user-task relationships.

Images