An Optimization Study Of Non-parametric Algorithms With Unbiased Divergence In Event Sequence Pattern Mining

With the advancement of human science and technology,the spatio-temporal event data that documents the occurrence of events is becoming more and more abundant.The researches on spatio-temporal sequential-pattern mining based on event types has been widely used in many fields.Since the prior information in many fields is unknown,the accuracy of pattern matching will be affected by the parametric model.In addition,the anomalous patterns of events represented by the anomalous data in the original data set usually occur within a certain spatio-temporal range.The detection method of the spatio-temporal anomaly interval that considers the correlation between data attributes and applies to the sequence of length changes is an important component of discovering the sequential patterns of spatio-temporal events.Aiming at the two problems of anomaly detection and trigger pattern matching in event sequential-pattern mining,in this paper,we propose an non-parametric triggering based on unbiased divergence algorithm(UDNT).The algorithm is mainly divided into two parts,the anomaly detection phase: the unbiased KL divergence sub-algorithm(UKLD)firstly defines the anomaly interval of the spatio-temporal time series,and after the delay is embedded,the detection interval and the remaining interval are estimated as Gaussian distribution.By accumulating and accelerating the parameter estimation process of Gaussian distribution,the unbiased KL divergence can be used to calculate the difference level between intervals,and then use the non-maximum value suppression method to obtain the spatio-temporal anomaly interval.Trigger pattern matching phase: The non-parametric triggering pattern sub-algorithm(NPTP)firstly based on the multivariate Hawkes model,uses non-parametric conditional strength functions to define the trigger relationship,then iteratively calculates the conditional probability of the trigger relationship,and finally obtains the significant event type trigger relationship by rank selection and take the mean value of the median probability as its significance value.Combining the two sub-algorithms to deal with the problem of spatio-temporal event sequential patterns,firstly extracting important abnormal parts from the original data and then performing the subsequent pattern mining process can be suitable for a variety of application scenarios.To verify the effectiveness of the algorithm,firstly select the general urban event synthesis dataset and the real event dataset from CHICAGO DATA PORTAL as the simulation dataset,and then perform simulation verification on Matlab.Secondly,in synthesis dataset,compare the accuracy of the sub-algorithm UKLD with RKDE and HOT SAX when the detection sequence length is a fixed value and a value interval.Finally,in the real event dataset,compare the performances of mining algorithms NPTP,CSTP and SSTS when using the data after the abnormality has been extracted by UKLD to trigger the relationship and compare the calculation accuracy of NPTP under different discretization levels.The simulation results show that for different lengths or arbitrary value intervals,the average accuracy of UKLD has been greatly improved,indicating that the UKLD algorithm can detect anomalies more accurately.For the three discretization levels,the mining performance of NPTP is better than that of CSTP and SSTS,and the best discretization level calculated by the D function makes NPTP achieve the highest accuracy.The UDNT algorithm proposed in this paper can guarantee a higher accuracy of event triggering relationship extraction when being applied to the mining of spatio-temporal event sequential patterns.