Taking Analytics Applications from Labs to the Real World: Transfer Learning in Practice
Traditional supervised learning models’ performances degrade if “nature” of test samples differ from that of training samples. For example, a classifier built to discriminate between “books” with positive, negative and neutral reviews when applied to discriminate between “kitchen products” into the same set categories, its performance drops. This relates to one of the fundamental probably approximately correct (PAC) assumptions that the training and test samples come from the same distribution. Consequently it leads to the practical implication that supervised models need to be provided with (enough number of) training samples from the domain where it is expected to be applied. This leads to laborious, tedious and ongoing labeling exercise limiting scalability and fast deployment of supervised algorithms.
In real life analytics applications, building models from scratch for every new domain hinders large-scale adoption of supervised statistical learning based analytics applications. Transfer learning techniques allow domains, tasks, and distributions used in training and testing to be different, thus reducing the requirement for labelled data. However, brute force techniques suffer from the problem of negative transfer, and we need to judge when and how much to transfer.
Pan, Sinno Jialin, and Qiang Yang. “A survey on transfer learning.” IEEE Transactions on knowledge and data engineering 22.10 (2010): 1345-1359. [online at https://www.cse.ust.hk/~qyang/Docs/2009/tkde_transfer_learning.pdf]
Bhatt HS, Dandapat S, Balaji P, Roy S.. “SODA: Service Oriented Domain Adaptation Architecture for Microblog Categorization.” [online at https://aclweb.org/anthology/N/N16/N16-3016.pdf]
Bhatt HS, Semwal D, Roy S. An Iterative Similarity based Adaptation Technique for Cross Domain Text Classification. CoNLL 2015. 2015 Jul 30:52. [online at: http://aclweb.org/anthology/K/K15/K15-1006.pdf]
In the first half of this talk, I will provide a brief overview of Transfer Learning techniques touching upon theory, applications, and systems. In the second half, I will talk about a real-life example how Transfer Learning can be effectively used for a social media analytics product going over resultant benefits.
Shourya Roy is currently a Senior Scientist and Research Manager at Xerox Research, India where he leads the “Text and Graph Analytics” group. In this role, Shourya is leading a group of researchers working on large scale text and graph analytics problems in domains such as outsourcing and customer care, healthcare and education. As a part of this role he also looks after research and business opportunities in customer care domain for Xerox in South-East Asia and Australia. Shourya’s research interest spans Text and Data Mining, Natural Language Processing, Machine Learning, and Human Computation. Over the years, Shourya’s work has led to about 50 patent disclosures and over 50 publications in premier journals and conferences such as ACL, AAAI, SIGKDD, SIGMOD, VLDB, WWW. He has taken up different professional roles including program committee member in top ranked conferences, editor in reputed journals, reviewer of journal and conference papers, advisor to students etc. He has been associated with several workshops in renowned text, data and web mining conferences – notably, the series of “Noisy Text Analytics”(AND) workshops which he co-initiated in 2007. This year he is co-organizing two workshops viz. Network Data Analytics (NDA) with SIGMOD 2016 and Health Data Management and Mining (HDMM) with ICDE 2016.