Research Of Online Learning Algorithm Based On Multi-task And Multi-kernel For Stream Data

Nowadays,in almost every area of society,a large number of stream data of various forms are generated at every moment.These data are produced at extremely high speeds and in large quantities and cannot be stored in total,and therefore,for the increasing stream data,the study of online learning algorithms is very necessary.The characteristics of stream data are complex,and stream data has strong temporal characteristics.Over time,the data stream characteristics may change slightly.Based on the above analysis,this thesis will further study the nonlinear online learning algorithm.The key point of the nonlinear model is to introduce a kernel function.The kernel function method is relatively simple to calculate,but it is generally used for single task learning.For complex data,it may be a combination of multiple tasks.There is a certain correlation between these tasks.Therefore,the multi-task learning method is also a direction worthy of attention.This thesis introduces the multi-kernel method when studying the multi-task online learning framework,which has more practical theoretical value and practical significance for the complex flow data processing.Based on the multi-task online learning framework,this thesis introduces the idea of multi-kernel to get a multi-task and multi-kernel online learning algorithm for streaming data.The algorithm has two innovations: one is to use a forgetting variable to optimize the weight update strategy,and the two is to use a limited threshold method to control the number of kernel functions.Its main work is as follows.The algorithm firstly determines a kernel function combination suitable for the current data features through iterative methods,and then learns multiple related tasks jointly,learning not only the common features of multiple tasks but also learning the inherent characteristics of each task,so that stream data can be processed more quickly.In addition,we introduce the concept of the input window,temporarily storing data that cannot be processed in real time,so that it can be calculated together at the next iteration to ensure the integrity of the data.In order to further optimize the online learning algorithm,this thesis proposes a multi-task and multi-kernel online learning algorithm that combines the characteristics of forgetting.Forgetting characteristics are manifested in two aspects: Forgetting characteristics are manifested in two aspects: on the one hand,a forgetting variable is introduced,and the weight updating strategy is optimized from the perspective of duality,so as to improve the convergence of the algorithm.The weight includes the inherent weight of each task and the weight of all tasks.On the other hand,when the amount of data increases and the model tends to be stable,the number of kernel functions is forgotten through the method defined by the threshold value,thereby reducing the number of basic kernel functions used,making the calculation more simple,thereby improving the model update speed.