Research Of Online Multiclass Learning With "Bandit" Feedback Under A Confidence-weighted Approach

One prominent feature of information epoch is the explosively expanding data,which makes it hard to process and extract meaningful information by dumping the whole dataset into a super computer or workstation.Nowadays,more and more large-scale parallel-computing infrastructures,like Hive,HBase,Hadoop,and Spark,are eye-catchy in industries and deployed in practical production while making profits.Meanwhile,schemes based on data streams gain quite a lot attention.Online learning methods based on data streams take advantages in many aspects,as only one traverse over the entire dataset,low requirements for local computers' computing capacity,and calibration to its predicting strategy along the way process samples.Traditional batch learning algorithms,like Na?ve Bayes,Decision Trees,SVM,and Neural Networks,require many traverses on the entire dataset.It is hard to transfer to the knowledge acquired from one specific scenario be to other scenarios when adopting online learning strategies.All of these imply the time,space and effectiveness of learning strategies of learning process.In General,we cannot apply knowledge learnt from one dataset on another dataset with different spatial temporal properties.Online machine learning algorithms,represented the first order learning method Perceptron at the beginning,simply verify its current learning strategy on every instance and update the weights of all features without difference to calibrate learning strategy.However,the whole learning process is quick but converges slowly.Afterwards,another first order learning method Passive-Aggressive proposes another metric scheme based on the margin and is more resistant to noises.Second order online learning algorithms like online learning methods based on confidence,keeps memory of confidence of different weights in the learning process.We use this confidence information to decide the update scale of weights when calibrating learning strategy.Furthermore,multiclass classifier learning with partial feedback is close to people's daily experience.For example,in the practice of recommendation system,the back-end system recommend commodities to customers,but it only acquires the like or dislike information which make the system hard to have a better understanding of users' preferences.First,this thesis discusses and studies how to fulfill streaming tasks on the big data processing infrastructure Spark with Spark Streaming library.A windowed stream is be processed very quickly,thus the delay is seconds even milliseconds.Second,this thesis studies and combines the second order online algorithms to speed up online learning and solve the learning problem of online multiclass classifier with bandit feedback.Third,we evaluate our algorithm on several synthetic and non-synthetic datasets.The empirical results show that our algorithm outperforms in most cases.