Interpretative Data Analysis And Its Application In Crime Pattern Mining And Forecasting

With the development of the big data era,the numerical machine learning algorithms are more and more popular in realms of image recognizing,audio and video processing.As for the interpretative work,it becomes disadvantaged in the research boom.For some complex problems,the rules and explanations are complicated and difficult to be straightened out,where the machine learning algorithm shows its advantages.In the initial upsurge of machine learning,the rule-based algorithms and the interpretability research have been pushed to the opposite side of the machine learning.There used to be a statement that the big data need no causal relationship.However,in our opinion,the correlation in big data is important,but the deep causal relationship is the core factor of the development in big data.With the increase of artificial intelligent application demand,the age of intelligent data analysis will come when the same data will need to make various decision due to different interpretations.In this thesis,the interpretative data analysis is proposed which improves the numerical algorithms on the basis of interpretation.In order to make exploratory research on the interpretive data analysis,this thesis tries to explain the whole process from three steps,data input,model processing and output mining.In the step of data input,the study of data matrix is divided into row problems and column problems.As a representative problem of data rows,imbalanced data problem is selected to integrate the interpretative data analysis method.In column problems,an interpretative attempt is made on Principal Components Analysis and its meliorations.For data input,the interpretative approaches can be applied in a wide range.For the rest steps,a specific application domain is essential.The crime data analysis is chosen for two reasons,one is the significance to the security of people's life and property,the other is the application requirement of interpretation.In the step of model improvement,two specific recurrent neural network structures are extended into interpretative method.In the step of output mining,interpretative improvements are appended to traditional pattern mining approaches,so as to get elaborate crime patterns.The main contributions of this article include:1.Interpretative Imbalanced Data Analysis Method: In this thesis,Generative Adversarial Minority Enlargement(GAME)method is proposed on the basis of local linear perspective,which integrates the Generative Adversarial(GA)game thinking and the explanation into over-sampling method.Unlike the GAN,the GAME utilizes the GA game method to adjust the classification line rather than to generate synthetic samples.The GAN learns the distribution of the real data,while the GAME introduces both the majority and the minority samples for training.From the experimental results,the classification effect of the input data processed by our model is better.2.Interpretative Principal Components Analysis Method: This thesis proposes an interpretative method on PCA and its improved algorithms.Different from the original approaches,our method selects the principal components mainly on the interpretation instead of the numerical calculation.The calculation results are treated as an adjusting factor.Through the test of different classifiers,the interpretative principal components by our models achieve better performance than the existing methods.3.Interpretative Locally Connected Recurrent Neural Networks: In this thesis,an interweaved time series technique is stated to handle multi time series of various time intervals without resampling.To put it into practice,we design a local connected RNN structure,and carry it out on LSTM and GRU cells.Through the test of the real world crime data,the quantitative prediction of the crime events in our model is more accurate than that in the comparison time series model and the RNNs without explanation structure.4.Interpretative Knotted-Clues Technique For Crime Patterns Mining: This thesis studies the near-repeat effect of crime incidents.On the basis of interpretation,a knotted-clue technique is proposed to analyze the fine-grained crime patterns among different districts and various crime types,which provides advices for mining and combating crime organization,police deployment and cooperation,and reasonable decision making support.