Key object classification method and device based on multivariate asynchronous sequence data
By adopting the classification method of multiple asynchronous sequence data in the prediction of reinfringe risks, the problems of timeliness and correlation processing of asynchronous data are solved, and the accuracy and practicality of reinfringe risks are improved.
Patent Information
- Application Number
- CN202510024113.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-07
- Publication Date
- 2025-05-23
AI Technical Summary
In the prior art, when predicting and preventing re-investigation risks, it is difficult to effectively deal with the timeliness and correlations of asynchronous data, and there are challenges in model construction and data integration.
A key object classification method based on multivariate asynchronous sequence data is adopted. By obtaining static data and asynchronous spatiotemporal data, static embedding features and asynchronous time factor features are extracted, feature fusion and time-dependent operations are performed, and finally classification is used by a classifier.
Effectively process the timeliness and relevance of asynchronous data, identify key factors that affect the risk of reoffending, and improve the accuracy and practicality of predicting reoffending risks.
Smart Images
Figure CN120030407A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of classification and recognition technology, and in particular, relates to a key object classification method and device based on multivariate asynchronous sequence data. Background Art
[0002] Predicting and preventing recidivism risk is an important part of social security. Traditional methods usually rely on static data and use manual scales to conduct questionnaire surveys and assessments, which often fail to fully reflect the real situation. For example, usually only fixed factors such as past records, family background and social support are considered, but there is a lack of comprehensive assessment of the dynamic changes of the subject after release. This makes it impossible to predict recidivism risk in a timely and effective manner in practical applications.
[0003] Against the backdrop of the rapid development of data science and machine learning, the application of asynchronous sequence data has gradually matured. Information such as the psychological characteristics, behavioral patterns, and social environment of the subjects of concern usually exists in the form of time series. These data not only contain the past history of the subjects of concern, but also include relevant information on their placement after release from prison, including psychological crisis assessment results, history of mental illness, addictive behavior, social interaction, educational background, employment status, economic factors, health data, personal credit and other multi-dimensional information. These data are often asynchronous, that is, the sampling of each time series data is not uniform. By analyzing these asynchronous sequence data, researchers can capture the changing trends and potential risks of the subjects of concern during the placement stage after release from prison, thereby improving the accuracy of recidivism risk prediction and providing a basis for early intervention for the placement department.
[0004] In addition, the diversity of asynchronous data provides rich information for the model, which helps to explore complex potential relationships and reveal the key factors that affect recidivism. This characteristic of multivariate asynchronous sequence data not only provides a new perspective for crime risk assessment, but also makes it possible to establish a more complex and sophisticated risk assessment model. By introducing machine learning algorithms, researchers can use these asynchronous data for deep learning and build a dynamic risk assessment model that can adapt to a changing environment, thereby effectively improving the ability to predict recidivism risk.
[0005] However, existing studies have attempted to combine multiple data sources for risk assessment, but they still face challenges in model construction and data integration. First, how to effectively handle the timeliness and relevance of asynchronous data is an important issue. The collection time of different data sources may be inconsistent, which makes data fusion and analysis difficult; secondly, how to select appropriate features to improve model performance is also a difficult problem that needs to be solved urgently. Summary of the invention
[0006] The purpose of this application is to provide a key object classification method and device based on multivariate asynchronous sequence data, to improve the accuracy and practicality of recidivism risk identification, to effectively process the timeliness and relevance of asynchronous data, to conduct a comprehensive analysis of behaviors at different time nodes, and thus to identify the key factors affecting recidivism risk.
[0007] In order to achieve the above purpose, the technical solution of this application is as follows:
[0008] A key object classification method based on multivariate asynchronous sequence data, comprising:
[0009] Obtain static data of key objects, perform feature extraction through quantization and embedding operators, and obtain static embedded features;
[0010] Obtain asynchronous spatiotemporal data of key objects, and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data;
[0011] The static embedding features, asynchronous spatiotemporal data and asynchronous time factor features are fused to obtain fused features, and the fused features are evenly divided into blocks without overlap along the time dimension;
[0012] Each feature block in the fusion feature is operated by a time-dependent operator to obtain the time-dependent features corresponding to each feature block;
[0013] For the time-dependent features of the fused features, feature mapping is performed along the time dimension to obtain mapping features, and the mapping features are classified using a classifier to obtain classification results.
[0014] Furthermore, the asynchronous time factor feature of the asynchronous moment in the asynchronous spatiotemporal data is extracted, and the calculation formula is as follows:
[0015]
[0016] in, For the moment The asynchronous time factor characteristic, ∈ is a fixed small amount to prevent data overflow, α 0 and α 1 are learnable weight parameters.
[0017] Furthermore, the static embedding feature is represented as E, and the asynchronous spatiotemporal data is represented as X raw :
[0018]
[0019] Where N represents the number of types of asynchronous spatiotemporal data, T represents the time length, Indicates that at time When the asynchronous spatiotemporal data is detected
[0020] The fusion feature is represented by X, and the fusion formula is:
[0021]
[0022] The present application also proposes a key object classification device based on multivariate asynchronous sequence data, comprising:
[0023] The static data processing module is used to obtain the static data of key objects, extract features through quantization processing and embedding operators, and obtain static embedded features;
[0024] Asynchronous spatiotemporal data processing module, used to obtain asynchronous spatiotemporal data of key objects and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data;
[0025] The fusion and blocking module is used to fuse the static embedding features, asynchronous spatiotemporal data and asynchronous time factor features to obtain fused features, and to block the fused features evenly and without overlap along the time dimension;
[0026] The time-dependent feature extraction module is used to operate each feature block in the fusion feature through the time-dependent operator to obtain the time-dependent feature corresponding to each feature block;
[0027] The classification module is used to perform feature mapping along the time dimension on the time-dependent features of the fusion features to obtain mapping features, and classify the mapping features using a classifier to obtain classification results.
[0028] Furthermore, the asynchronous spatiotemporal data processing module extracts the asynchronous time factor characteristics of the asynchronous moments in the asynchronous spatiotemporal data, and the calculation formula is as follows:
[0029]
[0030] in, For the moment The asynchronous time factor characteristic, ∈ is a fixed small amount to prevent data overflow, α 0 and α 1 are learnable weight parameters.
[0031] Furthermore, the static embedding feature is represented as E, and the asynchronous spatiotemporal data is represented as X raw :
[0032]
[0033] Where N represents the number of types of asynchronous spatiotemporal data, T represents the time length, Indicates that at time When the asynchronous spatiotemporal data is detected
[0034] The fusion feature is represented by X, and the fusion formula is:
[0035]
[0036] This application proposes a key object classification method and device based on multivariate asynchronous sequence data. By integrating data from multiple sources and using advanced machine learning technology, it can dynamically monitor and evaluate the risk of recidivism, and has obvious advantages in classification accuracy and reasoning time. At the same time, it provides strong support for practice, provides scientific decision-making support for the management and prevention of recidivism, and contributes to the realization of social security and stability. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a flow chart of the key object classification method based on multivariate asynchronous sequence data in this application.
[0038] Figure 2 This is a schematic diagram of the feature fusion blocks for this application. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0040] An embodiment of the present application, such as Figure 1 As shown, a key object classification method based on multivariate asynchronous sequence data is provided, comprising:
[0041] Step S1: Obtain static data of key objects, perform feature extraction through quantization processing and embedding operators, and obtain static embedded features.
[0042] This embodiment is described by taking prisoners as key targets as an example, and is not limited to a specific group of key targets.
[0043] First, obtain the static data of key targets, for example, including: gender, age, marital status, age, family support, number of recidivism, rehabilitation plan, points deducted for violations and disciplinary violations, and in-prison vocational training (academic advancement) and other static data.
[0044] These static data are then quantized and expressed as mathematical feature vectors e, and the quantization method is shown in Table 1:
[0045] Table 1
[0046]
[0047] Finally, by embedding the operator f eThe mathematical feature vector is converted into a continuous embedding vector to represent the static embedding feature E of the key object, and its mathematical expression is:
[0048] E=f E (e)∈R d ,
[0049] Where d represents the static embedding feature dimension, and the embedding operator f e It is a fully connected neural network.
[0050] Step S2: Acquire asynchronous spatiotemporal data of key objects, and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data.
[0051] This embodiment obtains asynchronous spatiotemporal data of key subjects, for example, including: number of recidivism, time served, parole and other related data, as well as relevant information on the post-prison placement stage, including psychological crisis assessment results, history of mental illness, addictive behavior, social interaction, educational background, employment status, economic factors, health data, personal credit, etc. These data are often collected at different time points and are therefore asynchronous.
[0052] Asynchronous spatiotemporal data using formula Indicates, where N represents the number of types of asynchronous spatiotemporal data, that is, the spatial dimension, and T represents the time length, such as within 2 years, T=2. Indicates that at time When the asynchronous spatiotemporal data is detected For example, with In terms of cumulative crimes, the key target committed 2 crimes on November 23, 2003; It means that the psychological crisis assessment result of the key target on January 3, 2004 was 0.8.
[0053] This embodiment uses a continuous time operation operator to extract each asynchronous moment Asynchronous time factor characteristics
[0054]
[0055] Among them, ∈ is a fixed small amount to prevent data overflow, α 0 and α 1 is a learnable weight parameter, which is learned through a neural network.
[0056] Step S3: Fuse the static embedding features, asynchronous spatiotemporal data and asynchronous time factor features to obtain fused features, and divide the fused features into blocks evenly and without overlap along the time dimension.
[0057] In this embodiment, the fusion method is summation, and the fusion formula is:
[0058]
[0059] Among them, X represents the fusion feature. This process enables the features extracted by the model to contain time period information and image information of key targets, which is conducive to the model's classification and discrimination.
[0060] Then, the fused features are divided into blocks evenly and without overlap along the time dimension, which can be expressed as:
[0061] X={X 1 ,…,X P};
[0062] Among them, P is the number of blocks, Represents a feature block in the fused feature.
[0063] like Figure 2 As shown, it is an example of block division. In this example, the number of blocks P is 3, which are divided into three feature blocks.
[0064] Step S4: Perform a time-dependent operator operation on each feature block in the fused feature to obtain a time-dependent feature corresponding to each feature block.
[0065] In this embodiment, for each feature block X in the fusion feature k , through the time-dependent operator operation, the time-dependent feature Z corresponding to each feature block is obtained k , the formula is as follows:
[0066] Z k =f ttcn (X k )
[0067] Among them, f ttcn represents the time-dependent operator, which is specifically expressed by k Perform full connection operations along the time dimension, and then perform activation operations, namely:
[0068]
[0069] Among them, σ is a statistical activation function, such as the sum of each dimension, w i and b are the full connection operation parameters. This process can unify the asynchronous time dimensions into one dimension.
[0070] Step S5: perform feature mapping along the time dimension on the time-dependent features of the fused features to obtain mapping features, and classify the mapping features using a classifier to obtain classification results.
[0071] This embodiment uses Transformer to extract the mapping feature Xtrans , its formula is expressed as:
[0072] X trans =f transformer (Z 1 ,…,Z P ).
[0073] The time-dependent feature combination of each feature block is the time-dependent feature Z of the fusion feature = {Z 1 ,…,Z p}, perform Transformer feature mapping along the time dimension to obtain the mapping features.
[0074] Then, for the mapping feature X trans , use a classifier to classify, for example, get three categories, representing high, medium and low recidivism levels. For example, use softmax to classify, perform probability normalization, and output the classification results of important targets based on the normalization results. This process can effectively process asynchronous data and extract powerful and effective classification features with attention through Transformer operations.
[0075] The technical solution of the present application has obvious advantages in processing asynchronous spatiotemporal data and time series problems through more sophisticated feature extraction, time-dependent modeling and efficient feature fusion methods. When it comes to the classification of key objects in real scenarios, the use of a direct asynchronous modeling method has not been studied in previous technologies.
[0076] In order to verify the effectiveness of the method proposed in this application, the method proposed in this application is compared with the existing synchronization method, and the performance difference is represented by the recognition rate on the test set and the reasoning time in the actual test process. The data set used is simulated data, and the experimental data includes 2000 training samples and 500 test data. The results are shown in Table 1.
[0077] Table 1
[0078] method Recognition rate (number of correctly identified samples / number of test samples) Reasoning time This application method 75.32％ 15ms Synchronous methods 69.52％ 30ms
[0079] It can be seen from Table 1 that the asynchronous method of the present application has obvious advantages in classification accuracy, indicating that the method proposed in the present application is suitable for asynchronous data classification. At the same time, the method proposed in the present application has obvious advantages in reasoning time. This is because when the data sampling time is inconsistent, the synchronous method needs to add a lot of 0 points to each running sample in order to achieve the purpose of synchronization. This will cause two main results. First, all samples will inevitably become longer, especially when the number of samples is limited. Second, a large amount of noise is introduced into the running sample data, which is not conducive to the classification of samples.
[0080] In another embodiment, the present application also proposes a key object classification device based on multivariate asynchronous sequence data, comprising:
[0081] The static data processing module is used to obtain the static data of key objects, extract features through quantization processing and embedding operators, and obtain static embedded features;
[0082] Asynchronous spatiotemporal data processing module, used to obtain asynchronous spatiotemporal data of key objects and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data;
[0083] The fusion and blocking module is used to fuse the static embedding features, asynchronous spatiotemporal data and asynchronous time factor features to obtain fused features, and to block the fused features evenly and without overlap along the time dimension;
[0084] The time-dependent feature extraction module is used to operate each feature block in the fusion feature through the time-dependent operator to obtain the time-dependent feature corresponding to each feature block;
[0085] The classification module is used to perform feature mapping along the time dimension on the time-dependent features of the fusion features to obtain mapping features, and classify the mapping features using a classifier to obtain classification results.
[0086] The present application provides a key object classification device based on multi-element asynchronous sequence data corresponding to the above-mentioned key object classification method based on multi-element asynchronous sequence data.
[0087] In a specific embodiment, the asynchronous spatiotemporal data processing module extracts the asynchronous time factor feature of the asynchronous moment in the asynchronous spatiotemporal data, and the calculation formula is as follows:
[0088]
[0089] in, For the moment The asynchronous time factor characteristic, ∈ is a fixed small amount to prevent data overflow, α 0 and α 1 are learnable weight parameters.
[0090] In a specific embodiment, the static embedding feature is represented by E, and the asynchronous spatiotemporal data is represented by X raw :
[0091]
[0092] Where N represents the number of types of asynchronous spatiotemporal data, T represents the time length, Indicates that at time When the asynchronous spatiotemporal data is detected
[0093] The fusion feature is represented by X, and the fusion formula is:
[0094]
[0095] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A key object classification method based on multivariate asynchronous sequence data, characterized in that: The key object classification method based on multivariate asynchronous sequence data includes: Obtain static data of key objects, perform feature extraction through quantization and embedding operators, and obtain static embedded features; Obtain asynchronous spatiotemporal data of key objects, and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data; The static embedding features, asynchronous spatiotemporal data and asynchronous time factor features are fused to obtain fused features, and the fused features are evenly divided into blocks without overlap along the time dimension; Each feature block in the fusion feature is operated by a time-dependent operator to obtain the time-dependent features corresponding to each feature block; For the time-dependent features of the fused features, feature mapping is performed along the time dimension to obtain mapping features, and the mapping features are classified using a classifier to obtain classification results.
2. The key object classification method based on multivariate asynchronous sequence data according to claim 1 is characterized in that: The calculation formula for extracting the asynchronous time factor feature of the asynchronous moment in the asynchronous spatiotemporal data is as follows: ； in, For the moment The asynchronous time factor characteristics, It is a fixed small amount to prevent data overflow. and are learnable weight parameters.
3. The key object classification method based on multivariate asynchronous sequence data according to claim 2 is characterized in that: The static embedding feature is expressed as , the asynchronous spatiotemporal data is represented as : ； in Indicates the number of types of asynchronous spatiotemporal data, Indicates the length of time, Indicates that at time When the asynchronous spatiotemporal data is detected ; The fusion feature is expressed as , the fusion formula is: 。 4. A key object classification device based on multivariate asynchronous sequence data, characterized in that: The key object classification device based on multivariate asynchronous sequence data includes: The static data processing module is used to obtain the static data of key objects, extract features through quantization processing and embedding operators, and obtain static embedded features; Asynchronous spatiotemporal data processing module, used to obtain asynchronous spatiotemporal data of key objects and extract asynchronous time factor features of asynchronous moments in the asynchronous spatiotemporal data; The fusion and blocking module is used to fuse the static embedding features, asynchronous spatiotemporal data and asynchronous time factor features to obtain fused features, and to block the fused features evenly and without overlap along the time dimension; The time-dependent feature extraction module is used to operate each feature block in the fusion feature through the time-dependent operator to obtain the time-dependent feature corresponding to each feature block; The classification module is used to perform feature mapping along the time dimension on the time-dependent features of the fusion features to obtain mapping features, and classify the mapping features using a classifier to obtain classification results.
5. The key object classification device based on multivariate asynchronous sequence data according to claim 4 is characterized in that: The asynchronous spatiotemporal data processing module extracts the asynchronous time factor characteristics of the asynchronous moments in the asynchronous spatiotemporal data, and the calculation formula is as follows: ； in, For the moment The asynchronous time factor characteristics, It is a fixed small amount to prevent data overflow. and are learnable weight parameters.
6. The key object classification device based on multivariate asynchronous sequence data according to claim 5, characterized in that: The static embedding feature is expressed as , the asynchronous spatiotemporal data is represented as : ； in Indicates the number of types of asynchronous spatiotemporal data, Indicates the length of time, Indicates that at time When the asynchronous spatiotemporal data is detected ; The fusion feature is expressed as , the fusion formula is: 。