Details
This project was done as part of the course CS771: Machine Learning Techniques, in the Fall ‘17 term at IIT Kanpur under Prof. Purushottam Kar, Department of Computer Science and Engineering, IIT Kanpur.
Abstract
The process of reviewer assignment to the papers of a conference is a very challenging and sensitive task. The choice of reviewers for a particular paper plays a crucial role in determining whether the paper is accepted into the conference. Automating the task of choosing reviewers is one of the recent challenges being actively researched by the machine learning community. The state of the art paper-matching system is the Toronto Paper Matching System which uses an LDA based topic modelling approach to identify topics in the entire paper and a simple dot product to assess similarity between reviewer topics and paper topics. Our approach builds on this line of work and develops an alternating optimization approach for completing the matrix of relevance scores between paper vectors and author vectors.
We did an extensive survey of the feild and understood the currently most prevalent techniques for automated paper-reviewer assignment like the Toronto Paper Matching System, and the Robust Paper-Reviewer Assignment Model. We implemented multiple modifications and techniques for improving the TPMS system by improvising on the Latent Dirichlet Allocation technique used in the generative model, and adding some intuitive biases. We also implemented an alternating optimization approach for completing the matrix of relevance scores between paper vectors and author vectors.
Amrit Singhal
MS in Machine Learning
My research interests include machine learning, reinforcement learning and artifical intelligence.