Research On Multi-task Multi-view Incremental Learning Algorithm For Data Stream Classification

Data stream classification aims to obtain useful models from large amounts of rapidly generated data,where training data and data to be classified are not obtained at once,but arrive continuously in the form of time-sequential data stream.Unlike the static classification process,combining all the training data to construct a classifier is not allowed in the data stream classification task,and it is not a very rational approach to accumulate data blindly(limited storage space and limited computing power).So there is a very important limitation in data stream classification:the model can only be trained(tuned)using the current or limited phase of training data.At present,for the data stream classification problem in human activity recognition,common algorithms often only consider the use of incremental learning to solve the catastrophic forgetting problem in data stream classification(meaning that when learning a new task or adapting to a new environment,forgetting or losing the previous acquired abilities)while ignoring that sensor-based data streams tend to have task heterogeneity as well as view heterogeneity.In this thesis,we formalize the problem of classification based on data stream on mobile sensors as a multi-task multi-view problem,where multi-task refers to multiple individuals participating in training,and multi-view refers to sensors located on multiple body parts of the individual.On the basis of incremental learning,we make full use of the idea of multi-task multi-view learning,and propose a multi-task multi-view incremental learning algorithm for data stream.Specifically,this thesis proposes two related neural network models,namely the MTMVIS model and the AMTMVIS model.For the data stream classification problem,the MTMVIS model is implemented as follows:First,we use a hierarchical attention mechanism to weight different kinds of sensors and different measurements produced in the same time interval to achieve data alignment and enhance confidence in discriminative sensor data.Next,to study more fine-grained task-view interaction dependencies,we model the task-view relationship with a multi-gate mixture-of experts model,we build a gating network for each task view,and each task-view can be the output of each expert obtains a specific feature representation.Furthermore,since previous multi-task multi-view learning models simply represent the average of all view outputs as the task output,ignoring that different views have different importance for the same task,for this we use another attention layer for each task to construct a view fusion layer for measuring the importance of different views to the same task.Finally,the catastrophic forgetting problem is overcome by restricting the relevant parameters of old stages.In addition,based on the MTMVIS model,the AMTMVIS model proposes a special adaptive output layer for each task,in which each layer of the network in the model is trained to obtain a customized feature representation from the attention layer as the input part of the final output layer,assign different weights to different network layers at different stages.The adaptive output effectively improves the scalability of the model.Finally we demonstrate the superiority of our algorithm through experiments on two different benchmark datasets.