Online Semi-Supervised Learning Using Max-Flow Algorithm

Abstract

In a common machine learning methods to classification, one can only make use of a labeled set from training the
classifier. The problem with the labeled instances is those can
be hard, expensive or may be very much time consuming to get
because of the need the help of human annotators. Sometimes
unlabeled data can be easy to get, but there have been some
methods to utilize them. Semi-supervised learning solves these
issues by making use of the huge size of unlabeled data, mixed
with labeled data for creating the better classifiers. Due to the
semi-supervised learning needs minimum human effort as well as
provides greater accuracy. The proposed system working on the
semi supervised learning of online data streams to improve the
classification accuracy. To achieve this, online max flow algorithm
is used, which seeks a maximum feasible flow through a singlesource, single-sink graph. By using labeled and unlabeled data,
graph is learn and updated according to online data stream
arrival. Also for classification, min cut is used. To improve the
classification accuracy and efficiency, system makes use of cosine
similarity and feature selection method respectively. The system
uses KDD dataset for analysis purpose and experimental results
prove that the proposed system achieves higher accuracy and
efficiency