| Back to Talks |
|
Seminar 2 : Statistical Relation Learning via Collective Matrix Factorization Speaker : Nitish Gupta Date : 06:00 pm, Monday, October 27th, 2014 Venue : KD101 Abstract: With recent advances in recommendation systems and relational learning, knowledge about collaborative filtering(CF) has become indispensable. Matrix Factorization techniques give the state-of-the-art results in recommendation systems in the absence of metadata about entities. Matrix Factorization techniques go beyond recommendation systems and are effective in learning relations from knowledge bases and related binary databases and in matrix completition tasks. Collective Matrix Factorization is an extension of matrix factorization for relation extraction from multi-relation databases. These methods provide compact and expressive lower dimensional embedding for entities in relations that are useful to capture characteristics of the data. In this talk I will first introduce matrix factorization techniques and collective matrix factorization techniques. I will then introduce my work in the field of Statistical Relation Learning for arbitrarily complex relational databases. I will also discuss how sign-rank of a binary matrix as a measure of complexity of relation captures more as compared to the rank of a matrix and how the logistic framework for matrix factorization is able to approximate the sign-rank of a binary matrix. Specific applications of these frameworks to problems in recommendation systems, multi-relation complex networks and knowledge bases will also be discussed. References :
|
|
Seminar 1 : Introduction to RKHS - Part I Speaker : Harish Karnick Date : 05:15 pm, Sunday, October 19th, 2014 Venue : KD102 Summary: In this introductory talk, we started by formulating the learning problem as a risk minimization problem. We then introduced the notion of regularizer to handle overfitting and talked about quantifying complexity of hypothesis. As a next step to define RKHS, several concepts of functional analysis were touched upon. We introduced hilbert spaces of functions discussing briefly advantages of working in hilbert spaces. Last part of the talk was on defining RKHS as hilbert space of functions with a linear bounded evaluation functional. |
| Back to Talks |