Research On Time Series Prediction In Complex Scenarios

Time series data is a collection of observed values that are observed in chronological order.It is a kind of important data object in sequence data and widely exists in our daily life and scientific research field.The characteristics of time series data include large data volume,high dimension and constant update.In addition,time series data have the characteristics of numerical value and continuity.It is generally believed that the key information in time series exists in the overall change rather than a specific value.The increasingly extensive use of time series data has led to a lot of research and development attempts in the field of data mining.Classification and prediction is a classic problem in the field of data mining.However,due to the complexity of time series data,the classification and prediction of time series data has become a special challenge in classification research in the past decades.Many researchers have conducted extensive and in-depth research on this issue.Although great progress has been made in the study of time series classification problems,there are still some problems to be solved in the study of specific algorithms in some complex scenarios.First of all,when the discriminating subsequences in the time series are phase offset,how to find out these subsequences accurately and efficiently for accurate classification?Secondly,when there are multiple class labels for a time series,how to make use of the dependency among multi-label and establish an effective multi-label classification algorithm for time series?Thirdly,how to implement an effective multi-variable time series classification algorithm when a time series instance has multiple variables simultaneously;Finally,in the recommendation system,how to combine the temporal information in the user rating sequence to establish the recommendation prediction algorithm.This paper makes an in-depth study on how to establish effective time series prediction algorithms based on the above four complex environments.The main contributions are as follows:(1)An algorithm named Regularized Random shapelet Forest(RRSF)is designed.Due to the inherent characteristics of shapelet algorithm,the process of RRSF algorithm to find the discriminant subsequences which are phase offset.In addition,the randomly selected strategy in this paper accelerated the discovery process of shapelet,guaranteed the prediction accuracy of the algorithm through the ensemble learning method,and greatly reduced the redundant shapelet in the random forest by punishing similar shapelet further improves the accuracy and interpretability of the algorithm.(2)An algorithm named ReliefF based Stacking(RFS)algorithm was designed.By adding the label attributes into the original attribute space and selecting the attributes.The algorithm not only qualitatively utilizes the dependencies between tags,but also quantitatively calculates the dependencies.Then the time series symbolization method is adopted to transform the data to generate a multi-label classification model suitable for the time series data.This is not only a deep extension of the dependency utilization method between time series data labels,but also a new application of multi-label attribute selection method.(3)A multivariable time series classification algorithm is designed.This algorithm transforms the multivariable time series data into the univariable time series data,and adds the trend characteristics between different variables on the basis of retaining the statistical characteristics,effectively utilizes the interrelationship between the variables,and solves the classification problem of the multivariable time series.In addition,the prediction problem of dropout of MOOC is regarded as a practical application scenario,the validity of the algorithm is verified in the real MOOC data,and an early prediction attempt is made.(4)A time-series collaborative ranking algorithm based on the local low-rank assumption of rating matrix is designed and implemented.This algorithm combines the temporal information in the rating sequence and the content of ranking learning technology.We assume that the scoring matrix has local low rank,and then choose to use the list ranking function to optimize its matrix factorization model.The above research results have realized targeted time series prediction algorithms in a variety of complex scenarios,showing the high efficiency of each algorithm in the prediction process,improving the interpretability of time series prediction methods.This paper also tries to solve some practical application problems,verifying the practicability of the algorithms in this paper.