CN113762599A - Short-term crime prediction method based on deep space-time residual error neural network - Google Patents

Short-term crime prediction method based on deep space-time residual error neural network Download PDF

Info

Publication number
CN113762599A
CN113762599A CN202110916670.8A CN202110916670A CN113762599A CN 113762599 A CN113762599 A CN 113762599A CN 202110916670 A CN202110916670 A CN 202110916670A CN 113762599 A CN113762599 A CN 113762599A
Authority
CN
China
Prior art keywords
crime
time
matrix
prediction
space
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110916670.8A
Other languages
Chinese (zh)
Inventor
方路平
李玉
董齐芬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Zhejiang Police College
Original Assignee
Zhejiang University of Technology ZJUT
Zhejiang Police College
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT, Zhejiang Police College filed Critical Zhejiang University of Technology ZJUT
Priority to CN202110916670.8A priority Critical patent/CN113762599A/en
Publication of CN113762599A publication Critical patent/CN113762599A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Alarm Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a short-term crime prediction method based on a deep space-time residual error neural network, which comprises the following steps of: (1) carrying out grid division on the researched city; (2) generating a crime spatio-temporal matrix from historical crime data of a researched city, and preprocessing the data; (3) constructing a short-term crime prediction model based on a deep space-time residual error neural network: and extracting crime time characteristics based on the crime space-time matrix, inputting the crime time characteristics into a crime prediction model for crime prediction to obtain a crime prediction result at the next moment, and continuously optimizing model parameters according to the result of the loss function to obtain a trained crime prediction model. The model structure of the invention is reasonable, and the accuracy of crime prediction is improved by analyzing the time-space correlation of crime data and predicting crimes in fine-grained time and space.

Description

Short-term crime prediction method based on deep space-time residual error neural network
Technical Field
The invention relates to the field of crime prediction, in particular to a short-time crime prediction method based on a deep space-time residual error neural network.
Background
The crime prediction is to apply a scientific method to analyze and research historical crime records and relevant factors possibly influencing crimes, and to make scientific conjectures on the overall conditions, structures, development trends and the like of future crimes in a certain space-time range. Many scholars research crime prediction problems, wherein the crime case quantity prediction means that the crime case quantity of a certain area in a specified time period in the future is predicted according to crime record data and other available data, and guidance can be provided for optimizing and deploying patrol police.
There are currently some efforts to predict the number of criminal cases. For example, the regularity of the crime data in the time distribution is utilized, that is, the number of crimes in a next period of a certain area is predicted according to the number of crimes in the past period of the area, and a commonly used prediction method is a time sequence analysis (such as an ARMA model, a SARIMA model), a statistical regression algorithm, and the like. Or the crime is forecast and analyzed by combining the time and space information, and the following references [1 ]: liumeilin, rare, Huanghong, etc., criminal intelligence prediction analysis [ J ] intelligence journal, 2018,37(9):27-37 ] based on space-time sequence mixing model. The above methods are crime prediction research methods under space-time coarse granularity (i.e. time is usually divided into one month or one year, and the range of the researched area is usually the whole city or a larger area), and the space-time coarse granularity is effective in mastering the long-term tendency of crimes, but has weak pertinence to microscopic patrol commands and the like.
Time and space division under the space-time fine granularity (namely time is divided to days or even hours generally, the area range to be researched is a dense area divided by cities generally) is dense, and the micro patrol command is more convenient and flexible, so that the centralized police force is facilitated. There are few studies on the prediction of crime in short time under the fine granularity of space and time.
Disclosure of Invention
The invention aims to overcome the difficulties of large space-time granularity, insufficient capture of crime space-time characteristics under space-time fine granularity and the like in most of current crime prediction researches, and provides a short-time crime prediction method based on a deep space-time residual error neural network by improving an ST-3DNet algorithm for short-time traffic flow prediction, so that crime prediction is carried out on fine-granularity time and space, and the aim of improving crime prediction accuracy is fulfilled.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a short-time crime prediction method based on a deep space-time residual error neural network comprises the following steps:
(1) and dividing the prediction area into I multiplied by J grid units with the same size according to the longitude and latitude.
(2) Obtaining historical crime data of a prediction area, projecting the historical crime data into each grid unit according to crime places at one-hour intervals, and statistically generating a crime space-time matrix X of the prediction area in each time period ttFor the crime space-time matrix XtIntegral processing is carried out to obtain an integral crime space-time matrix YtThen crime the score with the spatio-temporal matrix YtPerforming normalization processing to generate normalized crime space-time matrix Yt,scaled(ii) a Wherein for the crime spatio-temporal matrix xtIntegral processing is carried out to obtain an integral crime space-time matrix YtThe method comprises the following specific steps:
Figure BDA0003205856120000021
where k is t- (t mod 24).
(3) According to normalized crime spatio-temporal matrix Yt,scaledExtracting time characteristics aiming at the proximity similarity, the daily periodicity and the weekly periodicity, and respectively generating a proximity time matrix YhTime of day matrix YdA periodic time matrix YwAnd constructing three shared modules, a fusion layer for fusing outputs of the three shared modules and an output layer as a crime prediction model. The three shared modules are each made up of stacked 3D convolutional layers, multi-layer residual units, and one "recalibration" (Rc) module. With a time matrix Y of adjacent time constructed at time thTime of day matrix YdAnd period of the weekTime matrix YwCrime spatio-temporal matrix Y for t +1 time period as inputs to three modules respectivelyt+1And as a prediction target of the crime prediction model, training by using the acquired historical crime data of the prediction area to obtain a trained crime prediction model, and performing crime prediction on the to-be-predicted time period by using the trained crime prediction model.
Further, in the step (2), the crime data preprocessing step is as follows:
(2.1) acquiring historical crime data of a prediction area, and dividing according to one-hour time intervals; for each time period t, projecting historical crime data into each grid unit according to longitude and latitude, and statistically generating a crime space-time matrix X of the prediction area in the time period tt：
Figure BDA0003205856120000022
Wherein,
Figure BDA0003205856120000023
representing the number of crimes of the grid cell (i, j) over the time period t;
(2.2) criminal spatio-temporal matrix X generated in the step (2.1)tIntegral processing is carried out to obtain an integral crime space-time matrix Yt。
(2.3) integrating the crime spatio-temporal matrix Y of the step (2.2)tNormalization processing is carried out and a normalized crime space-time matrix Y is generatedt,scaled. The normalization processing formula is as follows:
Figure BDA0003205856120000024
wherein,
Figure BDA0003205856120000031
and
Figure BDA0003205856120000032
respectively representing the integrated value of the crime amount of the grid cell (i, j) and the normalized value thereof over the time period t,
Figure BDA0003205856120000033
and
Figure BDA0003205856120000034
respectively representing the maximum value and the minimum value of the integrated crime number value of the prediction area in the time period t;
further, the step (3) is based on a crime spatio-temporal matrix YtExtracting time characteristics aiming at the proximity similarity, the daily periodicity and the weekly periodicity, and respectively generating a proximity time matrix YhTime of day matrix YdPeriodic time matrix TwSpecifically, the following are shown:
Figure BDA0003205856120000035
Figure BDA0003205856120000036
Figure BDA0003205856120000037
wherein lw、ld、lwThe positive integers indicate the lengths of the proximity similarity, the day periodicity and the week periodicity in the time period T, T-24 indicates the period of the crime data day cycle time characteristic, d-1 indicates a fixed time day, and w-7 indicates a fixed time week.
Further, the fusion layer fusion result in the step (3) is represented as YfThe method specifically comprises the following steps:
Figure BDA0003205856120000038
wherein: degree is the Hadamard product of the matrix, Wfh，Wfd，WfwRespectively representing the weight matrixes learned from three characteristics of adjacent similarity, day periodicity and week periodicity,
Figure BDA0003205856120000039
and respectively representing time matrixes output by three sharing modules of the proximity similarity, the day periodicity and the week periodicity.
Further, the output layer in the step (3) is represented as:
Figure BDA00032058561200000310
f is the function of activation and,
Figure BDA00032058561200000311
the crime prediction result is the t +1 time period.
Further, the loss function trained by using the acquired historical crime data of the prediction area in the step (3) is represented as:
Figure BDA00032058561200000312
where theta is a learnable parameter in the model,
Figure BDA00032058561200000313
for crime prediction results of the model at time r +1, Yt+1The true value at time t +1 in the historical crime data,
Figure BDA0003205856120000041
is a matrix
Figure BDA0003205856120000042
L2 norm.
Compared with the prior art, the invention has the advantages and effects that:
(1) the invention is a crime forecasting method for capturing the crime space-time characteristics better under the space-time fine granularity; (2) the invention constructs three shared modules, and a fusion layer and an output layer for fusing the outputs of the three shared modules as a crime prediction model, so that the model is more suitable for the field of short-term crime prediction, the crime prediction accuracy is improved in fine-grained time and space, and reference and guidance are provided for police deployment and daily prevention and control of each grid area.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
fig. 2 is a crime data point processing analysis diagram, in which a is original crime data and b is processed crime data;
FIG. 3 is a schematic diagram of temporal feature extraction;
fig. 4 is a diagram of a crime prediction model.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
Referring to fig. 1 to 4, a method for predicting a crime in a short time based on a deep space-time residual neural network includes the following steps:
(1) dividing the urban grids: assuming that the longitude range and the latitude range of the predicted region are [ longitude _ start and longitude _ end ] and [ latitude _ start and latitude _ end ], the city under study is divided into I × J grid cells of the same size according to the following formula:
x＝(longtitude_end-longtitude_start)/I
y＝(latitude_end-latitude_start)/J
wherein x and y are the length and width, respectively, of each of said grid cells, each of said grid cells being represented by an index (i, j);
(2) crime data preprocessing: obtaining historical crime data in a preset time range of the prediction area, and dividing the preset time range into a plurality of time periods according to one-hour time intervals, wherein each time period is represented as t. For each time period t, projecting historical crime data into each grid unit according to longitude and latitude, and statistically generating a crime space-time matrix X of the prediction area in the time period ttFor the crime space-time matrix XtIntegral processing is carried out to obtain an integral crime space-time matrix YtThen crime the score with the spatio-temporal matrix YtPerforming normalization processing to generate normalized crime space-time matrix Yt,scaled(ii) a The method specifically comprises the following substeps:
(2.1) for each time period t, projecting historical crime data into each grid unit according to longitude and latitude, and statistically generating a crime space-time matrix X of the prediction area in the time period tt：
Figure BDA0003205856120000051
Wherein,
Figure BDA0003205856120000052
representing the number of crimes of the grid cell (i, j) over the time period t;
(2.1) criminal spatio-temporal matrix X generated in the step (2.1)tIntegral processing is carried out to obtain an integral crime space-time matrix Yt. The integral processing formula is as follows:
Figure BDA0003205856120000053
where k is t- (t mod 24). As shown in fig. 2 before and after the processing, it can be seen that crime data is more regular after the point processing.
(2.3) integrating the crime spatio-temporal matrix Y of the step (2.2)tNormalization processing is carried out and a normalized crime space-time matrix Y is generatedt,scaled. The normalization processing formula is as follows:
Figure BDA0003205856120000054
wherein,
Figure BDA0003205856120000055
and
Figure BDA0003205856120000056
respectively representing the integrated value of the crime amount of the grid cell (i, j) and the normalized value thereof over the time period t,
Figure BDA0003205856120000057
and
Figure BDA0003205856120000058
respectively representing the maximum value and the minimum value of the integrated crime number value of the prediction area in the time period t.
(3) Constructing an improved ST-3DNet model: according to normalized crime spatio-temporal matrix Yt,scaledExtracting time characteristics aiming at the proximity similarity, the daily periodicity and the weekly periodicity, and respectively generating a proximity time matrix YhTime of day matrix YdA periodic time matrix YwAnd constructs three shared modules, a fusion layer for fusing outputs of the three shared modules, and an output layer as a crime prediction model, as shown in fig. 4. Each of the modules consists of a stack of 3D convolutional layers, a multi-layer residual unit, and a "recalibration" (Rc) module. With a time matrix Y of adjacent time constructed at time thTime of day matrix YdAnd a cycle time matrix YwCrime spatio-temporal matrix Y for t +1 time period as inputs to three modules respectivelyt+1And as a prediction target of the crime prediction model, training by using the acquired historical crime data of the prediction area to obtain a trained crime prediction model, and performing crime prediction on the to-be-predicted time period by using the trained crime prediction model.
Further, the improved ST-3DNet model constructed in the step (3) has the following steps:
(3.1) As shown in FIG. 3, the time characteristics are extracted according to the proximity similarity, the day periodicity and the week periodicity by using the following formula, and a proximity time matrix Y is generated respectivelyhTime of day matrix YdA periodic time matrix Yw。
Figure BDA0003205856120000061
Figure BDA0003205856120000062
Figure BDA0003205856120000063
Wherein lh、ld、lwRespectively showing the lengths of the proximity similarity, the day periodicity and the week periodicity at the time T, wherein T shows the period of the time characteristics of the day period of the crime data, d shows one day of fixed time, and w shows one week of fixed time;
(3.2) as shown in the figure (4), the three time characteristics of the proximity similarity, the day periodicity and the week periodicity are utilized, so that the long-time dependent learning is increased. And aiming at three characteristics of proximity similarity, day periodicity and week periodicity, three shared modules are respectively constructed. And constructing stacked 3D convolutional layers in each module, and preliminarily extracting the spatial features of the crime data. Next, multi-layer residual units are added, each layer of residual unit comprising two 2D convolutional layers and two Relu activation functions. Finally, a "recalibration" (Rc) module is added to identify and quantify each of the grid cells; the output characteristics of the three modules are fused by using a method based on weight matrix fusion, and the fusion result is expressed as YfThe method specifically comprises the following steps:
Figure BDA0003205856120000064
where DEG is the Hadamard product of the matrix, Wfh，Wfd，WfwRespectively representing the weight matrixes learned from three characteristics of adjacent similarity, day periodicity and week periodicity,
Figure BDA0003205856120000065
respectively representing time matrixes output by three sharing modules of proximity similarity, daily periodicity and weekly periodicity; fused output YfInput into the final activation layer, the formula is:
Figure BDA0003205856120000066
f is an activation function, and the prediction result at the moment of t +1 is obtained
Figure BDA0003205856120000067
(3.3) training by using the acquired historical crime data of the prediction area, wherein the loss function of model training is expressed as:
Figure BDA0003205856120000068
where theta is a learnable parameter in the model,
Figure BDA0003205856120000069
for the prediction of the model at time t +1, Yt+1Is the true value at time t +1,
Figure BDA00032058561200000610
is a matrix
Figure BDA00032058561200000611
L2 norm.
And (3.4) obtaining the trained crime prediction model and carrying out crime prediction on the time period to be predicted by using the trained crime prediction model.
Example (b): the data in the actual experiment are predicted as follows:
(1) selecting experimental data
In this embodiment, taking crime data from 1 month and 1 day in 2015 to 30 months in 2015, from 7 months and 1 day in 2015 to 31 months in 2015, and from 1 month and 1 day in 2016 to 6 months and 30 days in 2016 as an example, the crime data in los angeles city is divided according to a time granularity of 1 hour, the spatial granularity is the size of grids divided into 16 × 16 in los angeles city, the number of crimes in each time period in each grid is counted, and a matrix is generated in each space-time granularity. And (3) performing model parameter training by using the crime data of the first 24 weeks as training data, and performing model performance verification by using the crime data of the last 2 weeks as test data.
(2) Parameter determination
The experiment is realized based on the tenserflow environment, and the construction of the whole experiment model framework is completed by using the keras. The overall experimental settings were as follows: three layers of Conv3D and two layers of residual error units are arranged for each shared module, the number of convolution kernels of Conv3D and residual error units is 64, wherein the convolution kernel size of the first layer of 3D convolution is l x 3(l represents the length of the input time characteristic), the convolution kernels of the other 3D convolution layers are 3 x 3, the convolution kernel size of the residual error units is 3 x 3, and the activation function is ReLU; selecting optimal parameters through experimental comparison, and comparing the similarity (l)h) Daily periodicity (l)d) Periodic (l)w) Set as 6, 2 and 2 respectively, and input into the crime prediction model of the invention to predict the crime amount at the next moment. In the comparison experiment, Closense and Period in the model ST-3DNet are respectively set to be 6 and 2, and the Closense and Period are input into the ST-3DNet network to predict the crime amount at the next moment.
(3) Results of the experiment
In the experiment, crime prediction was carried out using three sets of data of 1/2015 to 6/2015 for 30 days, 7/1/2015 to 2015 for 12/31 days, and 2016 for 1/2016 to 2016 for 6/2016 for 30 days, respectively, meanwhile, the ST-3DNet method of the original crime data and the crime prediction method of the invention are compared with the ST-3DNet method after the crime data preprocessing (integral processing and normalization processing) and the crime prediction method of the invention, the result statistical analysis is shown in Table 1, the result shows that the crime prediction method of the invention is lower than the RMSE of the ST-3DNet method under the original crime data, the ST-3DNet method and the crime prediction method of the present invention after the inventive crime data preprocessing are lower than the RMSE under the original crime data, the RMSE of the crime prediction method is the lowest, which shows that the prediction capability of the crime prediction model is greatly improved compared with that of a reference model.
Table 1: data comparison results
Figure BDA0003205856120000071
Figure BDA0003205856120000081
Figure BDA0003205856120000082
It should be understood that the above examples are only for clarity of illustration and are not intended to limit the embodiments. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. This need not be, nor should all embodiments be exhaustive. And obvious variations or modifications of the invention may be made without departing from the scope of the invention.

Claims (6)

1. A short-time crime prediction method based on a deep space-time residual error neural network is characterized by comprising the following steps:
(1) and dividing the prediction area into I multiplied by J grid units with the same size according to the longitude and latitude.
(2) Obtaining historical crime data of a prediction area, projecting the historical crime data into each grid unit according to crime places at one-hour intervals, and statistically generating a crime space-time matrix X of the prediction area in each time period ttFor the crime space-time matrix XtIntegral processing is carried out to obtain an integral crime space-time matrix YtThen crime the score spatio-temporal matrix rtPerforming normalization processing to generate normalized crime space-time matrix Yt，scaled(ii) a Wherein for the crime space-time matrix XtIntegral processing is carried out to obtain an integral crime space-time matrix YtThe method comprises the following specific steps:
Figure FDA0003205856110000011
where k is t- (t mod 24).
(3) According to normalized crime spatio-temporal matrix Yt，scaledExtracting time characteristics aiming at the proximity similarity, the daily periodicity and the weekly periodicity, and respectively generating a proximity time matrix YhTime of day matrix YdA periodic time matrix YwAnd constructing three shared modules, a fusion layer for fusing outputs of the three shared modules and an output layer as a crime prediction model. The three shared modules are each made up of stacked 3D convolutional layers, multi-layer residual units, and one "recalibration" (Rc) module. With a time matrix Y of adjacent time constructed at time thTime of day matrix YdAnd a cycle time matrix YwCrime spatio-temporal matrix Y for t +1 time period as inputs to three modules respectivelyt+1And as a prediction target of the crime prediction model, training by using the acquired historical crime data of the prediction area to obtain a trained crime prediction model, and performing crime prediction on the to-be-predicted time period by using the trained crime prediction model.
2. The method of claim 1, wherein in step (2), the crime data is preprocessed by:
(2.1) acquiring historical crime data of a prediction area, and dividing according to one-hour time intervals; for each time period t, projecting historical crime data into each grid unit according to longitude and latitude, and statistically generating a crime space-time matrix X of the prediction area in the time period tt：
Figure FDA0003205856110000012
Wherein,
Figure FDA0003205856110000021
representing the number of crimes of the grid cell (i, j) over the time period t;
(2.2) criminal spatio-temporal matrix X generated in the step (2.1)tIntegral processing is carried out to obtain an integral crime space-time matrix Yt。
(2.3) integrating the crime spatio-temporal matrix Y of the step (2.2)tNormalization processing is carried out and a normalized crime space-time matrix Y is generatedt，scaled. The normalization processing formula is as follows:
Figure FDA0003205856110000022
wherein,
Figure FDA0003205856110000023
and
Figure FDA0003205856110000024
respectively representing the integrated value of the crime amount of the grid cell (i, j) and the normalized value thereof over the time period t,
Figure FDA0003205856110000025
and
Figure FDA0003205856110000026
respectively representing the maximum value and the minimum value of the integrated crime number value of the prediction area in the time period t.
3. The method of claim 1, wherein the crime prediction in step (3) is based on a crime spatiotemporal matrix YtExtracting time characteristics aiming at the proximity similarity, the daily periodicity and the weekly periodicity, and respectively generating a proximity time matrix YhTime of day matrix YdA periodic time matrix YwSpecifically, the following are shown:
Figure FDA0003205856110000027
Figure FDA0003205856110000028
Figure FDA0003205856110000029
wherein lh、ld、lwThe positive integers indicate the lengths of the proximity similarity, the day periodicity and the week periodicity in the time period T, T-24 indicates the period of the crime data day cycle time characteristic, d-1 indicates a fixed time day, and w-7 indicates a fixed time week.
4. The crime prediction method of claim 1, wherein the fusion layer fusion result in step (3) is represented as YfThe method specifically comprises the following steps:
Figure FDA00032058561100000210
wherein:
Figure FDA00032058561100000211
is the Hadamard product of the matrix, Wfh，Wfd，WfwRespectively representing the weight matrixes learned from three characteristics of adjacent similarity, day periodicity and week periodicity,
Figure FDA00032058561100000212
and respectively representing time matrixes output by three sharing modules of the proximity similarity, the day periodicity and the week periodicity.
5. The crime prediction method of claim 1, wherein the output layer in step (3) is expressed as:
Figure FDA0003205856110000031
f is the function of activation and,
Figure FDA0003205856110000032
the crime prediction result is the t +1 time period.
6. The crime prediction method of claim 1, wherein the loss function trained in step (3) using the acquired historical crime data of the prediction area is expressed as:
Figure FDA0003205856110000033
where theta is a learnable parameter in the model,
Figure FDA0003205856110000034
for crime prediction results of the model at time t +1, Yt+1The true value at time t +1 in the historical crime data,
Figure FDA0003205856110000035
is a matrix
Figure FDA0003205856110000036
L2 norm.
CN202110916670.8A 2021-08-11 2021-08-11 Short-term crime prediction method based on deep space-time residual error neural network Pending CN113762599A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110916670.8A CN113762599A (en) 2021-08-11 2021-08-11 Short-term crime prediction method based on deep space-time residual error neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110916670.8A CN113762599A (en) 2021-08-11 2021-08-11 Short-term crime prediction method based on deep space-time residual error neural network

Publications (1)

Publication Number Publication Date
CN113762599A true CN113762599A (en) 2021-12-07

Family

ID=78788937

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110916670.8A Pending CN113762599A (en) 2021-08-11 2021-08-11 Short-term crime prediction method based on deep space-time residual error neural network

Country Status (1)

Country Link
CN (1) CN113762599A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114418244A (en) * 2022-03-29 2022-04-29 北京零点远景网络科技有限公司 Case prediction analysis method and device, electronic equipment and storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111598059A (en) * 2020-06-28 2020-08-28 中国人民公安大学 Crime prediction model training method and device and electronic equipment
CN111861028A (en) * 2020-07-29 2020-10-30 浙江工业大学 Method for predicting crime number based on spatio-temporal data fusion

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111598059A (en) * 2020-06-28 2020-08-28 中国人民公安大学 Crime prediction model training method and device and electronic equipment
CN111861028A (en) * 2020-07-29 2020-10-30 浙江工业大学 Method for predicting crime number based on spatio-temporal data fusion

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BAO WANG等: "Deep Learning for Real-Time Crime Forecasting and Its Ternarization", CHINESE ANNALS OF MATHEMATICS, SERIES B, pages 949 - 966 *
SHENGNAN GUO等: ""Deep Spatial–Temporal 3D Convolutional Neural Networks for Traffic Data Forecasting", IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, pages 3913 - 3926 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114418244A (en) * 2022-03-29 2022-04-29 北京零点远景网络科技有限公司 Case prediction analysis method and device, electronic equipment and storage medium
CN114418244B (en) * 2022-03-29 2022-07-08 北京零点远景网络科技有限公司 Case prediction analysis method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
Liu et al. Attentive crowd flow machines
Wang et al. Deep learning for real-time crime forecasting and its ternarization
CN109754113B (en) Load prediction method based on dynamic time warping and long-and-short time memory
Yao et al. Modeling spatial-temporal dynamics for traffic prediction
Fan et al. A novel deep generative modeling-based data augmentation strategy for improving short-term building energy predictions
CN112532439B (en) Network flow prediction method based on attention multi-component space-time cross-domain neural network model
CN111161535A (en) Attention mechanism-based graph neural network traffic flow prediction method and system
Esquivel et al. Spatio-temporal prediction of Baltimore crime events using CLSTM neural networks
CN110009046B (en) Urban community safety prediction method based on big data
CN115148019A (en) Early warning method and system based on holiday congestion prediction algorithm
CN111598059A (en) Crime prediction model training method and device and electronic equipment
Massaoudi et al. Performance evaluation of deep recurrent neural networks architectures: Application to PV power forecasting
CN114692984A (en) Traffic prediction method based on multi-step coupling graph convolution network
CN109145743A (en) A kind of image-recognizing method and device based on deep learning
CN118211494B (en) Wind speed prediction hybrid model construction method and system based on correlation matrix
CN117612369A (en) Short-time traffic flow prediction method, system, storage medium, equipment and terminal
CN113762599A (en) Short-term crime prediction method based on deep space-time residual error neural network
Xiong et al. DCAST: a spatiotemporal model with DenseNet and GRU based on attention mechanism
CN110766231A (en) Crime prediction method and system based on multi-head neural network
CN117933531A (en) Distributed photovoltaic power generation power prediction system and method
Gao et al. Interpretable deep learning for hourly solar radiation prediction: A real measured data case study in Tokyo
CN117472511A (en) Container resource monitoring method, device, computer equipment and storage medium
CN117133129A (en) Traffic speed prediction method based on multi-component attention-seeking neural network
CN116630113A (en) SCC-RCE prediction unit-based prediction network and gridding PM2.5 concentration prediction method
CN116646917A (en) Campus multi-step probability power load prediction method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination