Time Series Clustering Based On Mixture Model

Clustering is one of the important tools of data mining.It helps people more intuitively recognize the similarities and differences between time series,so that they can implement different strategies for different things and improve work efficiency.Data mining involves many data types such as space,text,and time series,among which time series is the most common.Economic data is typical time series data,which contains an important feature-time attributes.The attribute of time usually plays the key role of factor in data mining,which cannot be ignored.So,it increased the difficulty of data mining.The size of the data set has grown rapidly over time.Traditional clustering methods cannot effectively deal with time-type highdimensional data.The model-based clustering methods are more adaptable and interpretable to time series by collects prior information and sets an appropriate hypothesis model as the basis of analysis.It has important research value in exploratory technologies such as sequence matching and anomaly detection.This paper considers the GARCH model that describes the characteristics of volatility in time series.Presumed that there are different volatility states in the model.These states are constantly switched under the influence of some external factors,resulting in an MS-GARCH model with variable parameter structure.On the basis,the time series is composed of a mixture of multiple models.Essentially,we use the difference of model parameters to distinguish the volatility persistence of data.Then,the data with similar volatility are grouped into the one category.In terms of calculation,this article is based on Bayesian theory,combining prior information and likelihood function to get posterior probability function.Using MCMC method for numerical simulation.The Gibbs sampling is used for model parameters with known conjugate priors,and the griddy-Gibbs sampling is used for model parameters with complex posterior probability functions,which the process is iterated until it converges to the target distribution.This method solves the path dependence problem of heteroscedasticity model.In summary,using the finite mixture model can fully extract the fluctuation structure of the data,and obtain effective clustering conclusions.