Search

CN-121998185-A - Crime prediction method based on space-time point process

CN121998185ACN 121998185 ACN121998185 ACN 121998185ACN-121998185-A

Abstract

The invention belongs to the technical field of data processing and discloses a crime prediction method based on a space-time point process, which comprises the steps of S10, extracting space-time characteristics by respectively carrying out time coding and space embedding on each space-time sequence, realizing space-time characteristic coding of crime events, S20, extracting global relativity of historical crime data by utilizing a human liveness attention mechanism, realizing global dependency capture of the space-time characteristics, S30, calculating dependency relationship among characteristic elements by utilizing a stacked multi-head self attention mechanism, realizing capture of space-time context dependency with finer granularity, S40, taking the obtained space-time characteristics as conditions, guiding a diffusion model to model space-time dynamic evolution of the crime events, and finally generating future crime prediction results. The method can effectively solve the problems of insufficient space-time dependency capture, poor model generalization capability and excessive dependency on parameter assumption in the prior art.

Inventors

  • HE GANG
  • QIU JIAMIN
  • WU YIHONG

Assignees

  • 西南科技大学

Dates

Publication Date
20260508
Application Date
20260122

Claims (8)

  1. 1. A crime prediction method based on a space-time point process is characterized by comprising the following steps: S10, respectively carrying out time coding and space embedding on each time-space sequence in the historical crime data, extracting time-space characteristics, and realizing time-space characteristic coding of crime events; s20, extracting global relativity of historical crime data by using a human liveness attention mechanism to realize global dependency capture of space-time characteristics; S30, calculating the dependency relationship among characteristic elements by using a stacked multi-head self-attention mechanism, and capturing the space-time context dependency with finer granularity; And S40, using the obtained space-time characteristics as conditions, guiding a diffusion model to model the space-time dynamic evolution of the crime event, and finally generating the occurrence time and longitude and latitude of the crime event.
  2. 2. The method for predicting crime based on the space-time point process according to claim 1, wherein the space-time feature coding of the crime event is realized by respectively performing time coding and space embedding on each space-time sequence, extracting the space-time feature, comprising: An input layer for inputting crime event information of a plurality of days in the past; space-time coding, namely, adopting a fixed mode position coding based on a periodic function to input crime time to obtain time characteristics, space embedding, namely, mapping crime space information into an embedded vector by using a linear projection mechanism to obtain space characteristics, and obtaining the space characteristics by adding and fusing the time characteristics and the space characteristics of each event.
  3. 3. The method of claim 2, wherein the crime prediction information for the past days includes time, longitude and latitude information of occurrence of the crime.
  4. 4. The method of claim 1, wherein the human liveness feature is extracted and established according to historical crime data characteristics, and the human liveness feature is integrated into a crime prediction model by using a self-attention mechanism.
  5. 5. The method of claim 4, wherein the human liveness feature is based on the distribution information of the whole historical crime data in time and space respectively, and comprises: Preprocessing historical crime data, segmenting different time and space values in the data, respectively calculating the probability of each segment in global event data, and obtaining the activity of each event in time and space according to the segment of each event; The method comprises the steps of mapping time and space liveness of each event as input features into three feature spaces of query, key and value, calculating similarity scores among input units, and generating a weight matrix, wherein the weight matrix is used for weighting the input features so that global dependency relations of different time periods and different space periods are captured.
  6. 6. A crime prediction method based on a spatiotemporal point process according to claim 1, characterized in that a finer granularity of spatiotemporal context dependent capture is achieved by iterating the query-attention process of multiple self-attention layers using a stacked multi-headed self-attention mechanism.
  7. 7. The method of claim 6, wherein in the stacked plurality of self-attention layers, the multi-headed self-attention mechanism of the first layer is provided with a mask, and the subsequent layers are not masked.
  8. 8. The crime prediction method based on the space-time point process according to claim 1, wherein the space-time characteristics obtained by space-time coding, merging human liveness characteristics and weighting by a stacked multi-head self-attention mechanism are used as conditions for guiding a diffusion model to model the space-time dynamic evolution of a crime event; In the reverse denoising process of the diffusion model, the model receives space-time characteristic conditions, a noisy sequence state and current step information as inputs, predicts noise components which are required to be removed from the inputs in the current step, and finally generates a brand new complete crime space-time sequence at a future time through repeated iteration of the denoising step, so as to obtain the occurrence time and longitude and latitude of a crime event.

Description

Crime prediction method based on space-time point process Technical Field The invention belongs to the technical field of data processing, and particularly relates to a crime prediction method based on a space-time point process. Background Crime is a ubiquitous social problem that has destructive effects on individual well-being, community cohesiveness and economic viability. Therefore, how to accurately predict the spatiotemporal dynamics of criminal events to optimize police deployment and take effective precautions becomes an important point for urban management and public safety research in various countries. Existing crime prediction techniques can be broadly divided into four classes, a statistical-based model, a machine-learning-based model, a deep-learning-based model, and a point-process-based model. These models each have different advantages and application scenarios, but there are still significant limitations in dealing with complex, dynamic urban crime phenomena. The statistical model is used for identifying criminal high-incidence areas by utilizing historical criminal data through technologies such as regression analysis, hot spot drawing and the like. Such models, while relatively simple to calculate, are descriptive in nature, have limited capture capability for dynamic infection effects and complex spatiotemporal associations of criminal events, and predictions tend to lag behind actual changes. The machine learning model predicts the relationship between the training history data and the characteristics of geography, socioeconomic and the like. Although the model can capture partial nonlinear relation, the performance of the model is seriously dependent on the completeness of characteristic engineering, and the generalization capability of the model in different cities or communities is weaker. In addition, the deep learning model based on the convolutional neural network and the cyclic neural network provides stronger capability for capturing the space-time mode. The standard CNN/LSTM architecture is primarily directed to regular, equally spaced grid data or sequence data. However, crime events are essentially asynchronous, irregular point events that occur in continuous time and air, and the direct application of these models requires the discretization of data into coarse grids and periods of time, resulting in a significant loss of information that does not accurately express the precise time, location, and their complex interactions of individual events. The model based on the point process is a theoretical framework for modeling an asynchronous event stream, and the core idea is that a history event can influence the occurrence of a future event, which intuitively fits the 'near repetition' and infection phenomenon of crimes, is theoretically suitable, but has poor flexibility in practice and limited expression capability, usually requires strict parameterization assumption, and is difficult to describe criminal activities in the real world. Despite the significant advances made by the prior art in crime prediction, the prior art still faces some challenges: the capturing of space-time dependence is insufficient, that is, the complex space-time correlation in crime events is difficult to capture by the existing model, and the limitation is particularly prominent in urban complex environments. The generalization capability of the model is poor, namely the generalization capability of a plurality of models (especially a machine learning model which is severely dependent on feature engineering and a model which is trained by specific regional data) is poor in different cities and different social backgrounds. The learned mode and specific region characteristics are difficult to adapt to the spatial heterogeneity and time variability of the crime mode, and the general application value of the method is limited. Over-dependent parameter assumptions for mathematical manageability of theoretical frameworks such as traditional point process models, strong parameter assumptions (e.g., exponential form, fixed range of influence) are often made about the pattern of influence of events (e.g., the spatiotemporal decay kernel). The simplifying assumptions and crimes have deviation in a real city complex evolution mode, so that the model has insufficient flexibility and larger prediction deviation. Disclosure of Invention In order to solve the problems, the invention provides a crime prediction method based on a space-time point process, which combines the space-time point process with a diffusion model, and captures global crime dynamics and finer granularity space-time correlation respectively by utilizing a human liveness attention mechanism and a stacked multi-head self-attention mechanism, thereby improving the accuracy of crime prediction. The method can effectively solve the problems of insufficient space-time dependency capture, poor model generalization capability and excessive dependen