Abstract for the Co-training Seminar on Friday

[Amit.Mandvikar@asu.edu]

hello folks,

For tommorrow's AIDB seminar I am presenting a seminar on Co-training, a very 
interesting topic. The seminar will be from 3-4pm tommorrow Friday, the 29th. I 
am attaching a text file with the abstract for the seminar. I would like to 
apologize for being this late. 
I hope everyone finds time to come for this seminar and put forth their views 
and opinions on Co-training. 

sincerely, 
Amit Mandvikar.

					Abstract

Classification Learning is a predictive process in which new applications and systems are automated to predict outputs without any or as less intervention as possible from human experts. The job of a learning algorithm is to relieve the experts from making those decisions and to automate the process of making decisions. Learning algorithms tend to make the simpler decisions themselves and then refer only the tougher (complex) decisions (cases where the learning algorithm is not completely sure about the classes) to the human experts.

Classification algorithms are usually supervised while clustering algorithms are unsupervised. Co-training is a combination of both of supervised and unsupervised algorithms. Co-training leads to a model wherein the clustering properties of the data are used to increase the prediction accuracy for the classification process. Co-training incrementally first estimates labels for the unlabeled training data and then appends that to the labeled set. In doing this it reduces the necessity for extensive expert involvement in the labeling process. Co-train also reduces overfitting to a specific training set by using multiple classifiers and then combining their outputs later.  

The seminar on Co-training will include the following major topics: -
·	Introduction to Co-training.
·	Why do we use Co-training?
·	Where can we apply co-training?
·	Description of the algorithms.
·	Comparison with EM.
·	Experiments and results.

The talk will focus on how Co-training manages to outperform the Expectation-Maximization algorithm and under what circumstances.