Incremental Learning Algorithms And Their Applications In Time Series Prediction

As a hot research field,Time Series Prediction(TSP)has very high practical application value.Incremental learning as a research branch in machine learning is good at processing streaming data,and time series is a typical streaming data that changes with time,so incremental learning is very suitable for solving problems of TSP.This paper mainly proposes two novel incremental learning algorithms for processing TSP problems.In this paper,incremental learning paradigm is combined with Incremental Support Vector Machine(ISVM),establishing a novel algorithm for TSP,namely Double Incremental Learning(DIL)algorithm.ISVM is developed on the basis of Support Vector Machine(SVM),which is very suitable for processing online continuous learning problems such as time series prediction.In DIL algorithm,ISVM is utilized as the base learner,while incremental learning is implemented by combining the existing base models with the ones generated on the new data.A novel weight update rule is proposed in DIL algorithm,being used to update the weights of the samples in each iteration.Furthermore,a classical method of integrating base models is employed in DIL.The DIL algorithm inherits many advantages of ISVM and incremental learning,so it can achieve desirable prediction effect for TSP.Experimental results on six benchmark TSP datasets verify that DIL possesses preferable predictive performance compared with other existing excellent algorithms.In view of the deficiencies in the DIL algorithm,some improvements are made in this paper,and the improved algorithm is called the dual weights optimization Incremental Learning(w~2IL)algorithm.The w~2IL utilizes Incremental Extreme Learning Machine(IELM)as the base learner.Compared with the DIL algorithm,there are two main improvements in w~2IL.These are also w~2IL's two major innovations,namely,two subalgorithms:Adaptive Samples Weights Initialization(AdaSWI)and Adaptively Weighted Predictors Aggregation(AdaWPA).The AdaSWI subalgorithm initializes the samples'weights adaptively based on the generated base models'prediction errors,and fine-tunes the samples'weights based on the distances from the samples to the training datasets of base models,achieving more appropriate samples weights initialization.While AdaWPA algorithm adaptively adjusts base predictors'weights based on prediction instances and integrates them.Besides,AdaWPA subalgorithm makes use of Fuzzy C-Means(FCM)clustering algorithm for distance measurement,further reducing computational complexity and storage space of the algorithm.Compared to the DIL algorithm and other existing excellent algorithms,the performance of w~2IL has been significantly improved,and numerical experiments on six real-world TSP datasets have further verified this.