Dan Pomerantz
My name is Dan Pomerantz. Since Fall 2009, I have been working for Microsoft as part of the Bing Shopping team. In Fall 2010, I moved to Montreal and continued to work remotely for my team as well as teaching some courses at McGill.
This page is mostly about my work as a Master's student at McGill University. My area of study focuses on robotics and machine learning, specifically on a movie recommender project.
In addition to project reports, I have posted my CV
Teaching
I am currently teaching several introductory courses. Here are links to the course webpages which all use the same layout.
Research Project
Context Dependent Movie Recommendations Using a Hierarchical Bayesian Model: In this paper, I explored providing context-dependent recommendations to users. This allows users to get recommendations for specific settings, such as "on a date" or "on a Wednesday night." This paper was published in the 2009 Canadian Conference of Artificial Intelligence (AI 2009) Abstract Pdf .
I also published my thesis in expanded work with the same name. pdf
Robot localization Using Local Matching: During the term of fall 2007, Yogesh Girdhar and I worked on a computer vision project on vision based object recognition and localization. The idea is to apply an interest operator both in the training and query phases. This allows for local image matching and helps address issues related to occlusion. Additionally, we use polar coordinates and Fourier transforms in order to achieve rotation invariance. We test this algorithm to a robot trying to figure out what part of a video sequence it is looking at.
Sentiment Classification in Movie Reviews: During this same term, I also independently worked on a sentiment classification project for machine learning. I studied applying machine learning algorithms to classify movie reviews as "good" or "bad." The methodology used was first creating a subjectivity classifier that could filter out many of the "factual" sentences that were not useful for sentiment. For example, the sentence "The character is very evil and a horrible villain in the movie" contains many seemingly negative words, but does not mean that the movie is bad, so our approach removes these sentences. (It is possible that we could gain information related to what sentences we filter (for example the fact that the writer thought the character was that "evil" may be a compliment to the movie), but we assumed that once we labeled something as "objective" it wasn't useful.) Additionally, I examined reducing the dimension of the space using a feature selection approach based on information gain. The paper (pdf) and slides are available for download.
Here are some miscellaneous perl/C examples: that may be useful for comp206 students.