publications

2019

1. LISP: Localized Inference through Selective Pre-training for Critical Diagnosis Forecasting Samel, Karan, and Sun, Jimeng GT CS 8903 Special Problems 2019 [PDF]
2. Contextual Model Interpretation: A Model Agnostic Approach to Search For Explainable Data Sub-Sets Samel, Karan, and Chu, Xu GT CS 7001 Research Project 2019 [Poster]
3. A Survey of Spatio-Temporal Graph Neural Networks for Traffic Prediction Samel, Karan, and Dovrolis, Constantine GT CS 7001 Research Project 2019 [Abs] [PDF]

   Deep learning is a growing field due to ongoing collection of data and the increase in compute power. Many of these data collec- tions are represented as graphical structures in comparison to the more common tabular data structures. There has been a recent interest in the deep learning community to leverage these graph structures within models to encode spatial dependencies for better modeling performance. These graph representations often have a temporal component as well that are modeled by recurrent deep learning techniques to create spatio-temporal graph neural net- works (STGNNs). This application is being applied for traffic speed prediction, where sensors in a grid are measured over time. We compare different STGNNs tested on traffic prediction to determine the current advances in the field. We conclude by discussing the growing applications of STGNNs in other pertinent fields.

2018

1. Active Deep Learning to Tune Down the Noise in Labels Samel, Karan, and Miao, Xu In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining 2018 [Abs] [HTML]

   The great success of supervised learning has initiated a paradigm shift from building a deterministic software system to a probabilistic artificial intelligent system throughout the industry. The historical records in enterprise domains can potentially bootstrap the traditional business into the modern data-driven approach almost everywhere. The introduction of the Deep Neural Networks (DNNs) significantly reduces the efforts of feature engineering so that supervised learning becomes even more automated. The last bottleneck is to ensure the data quality, particularly the label quality, because the performance of supervised learning is bounded by the errors present in labels. In this paper, we present a new Active Deep Denoising (ADD) approach that first builds a DNN noise model, and then adopts an active learning algorithm to identify the optimal denoising function. We prove that under the low noise condition, we only need to query the oracle with log n examples where n is the total number in the data. We apply ADD on one enterprise application and show that it can effectively reduce 1/3 of the prediction error with only 0.1% of examples verified by the oracle.

2017

1. Predicting Advertisement Clicks Using Deep Networks: Interpreting Deep Learning Models Samel, Karan The Journal of Purdue Undergraduate Research 2017 [Abs] [HTML]

   Deep learning has achieved state-of-the-art results in a variety of tasks such as classifying images and driverless cars. In this paper, I used deep learning to understand consumer product interests. One of the main goals for advertisement agencies is to develop mathematical models to predict whether consumers will click on their advertisement. Achieving the highest click prediction rate means that these agencies can pay to place their online advertisements effectively to target people most interested in their product. Most existing approaches are based on logistic regression or regression tree models (Trofi mov, Kornetova, & Topinskiy, 2012). The model based on deep learning will be discussed to predict the click rate. The data was from the iPinYou competition, where competitors are tasked to build a model that would achieve a high click through rate (CTR). iPinYou provides advertisement data from nine companies. For each instance in the data, various attributes of the person that the electronic advertisement was sent to were provided as well as if the person clicks on the advertisement. I started with exploratory data analysis by splitting data into different seasons, aggregating different advertisers, and cleaning and generating new attributes. I tested my predictive power using a convolutional neural net and a multiple layer perception model. It was shown that the deep learning models have a competitive predictive power and, at the same time, more interpretable for further analysis.