Saeed Ghoorchian

About

My research focuses mainly on building machine learning methods to solve online decision-making problems under uncertainty. To achieve this, I have worked on developing decision-making algorithms capable of learning the optimal action and feature observations in non-stationary environments with causal dependencies. Most existing machine learning methods, particularly in the era of big data, postulate the possibility of information acquisition with no limit and for free. This leads to deploying models that require observing all features' states before predicting a class label. In real-world problems, however, collecting beneficial information is often costly. For example, in online advertising problems, the advertiser can purchase information about target users to display personalized ads. As another example, in medical contexts, obtaining data for treatment recommendations mainly requires additional tests that are time- and money-consuming. How to maximize the model performance while keeping the cost of feature observations as low as possible? What if the reward distribution and features' cost distribution vary over time? Such cases frequently appear in real-world problems; for example, in personalized news recommendations, user preferences over news can change over time and exhibit various seasonality patterns. Hence, we need novel learning methods that, besides individual actions' rewards, learn the observations of the features' states and are robust to distribution shifts in the deployed environment.

Now, what if there are causal dependencies among actions' outcomes? A real-world example of this situation is the Covid-19 development within a country. During the Covid-19 pandemic, containing the virus outbreak has been one of the major concerns of governments. To this end, health authorities attempt to monitor the outbreak and detect the regions likely to become coronavirus hotspots. It is only natural that health authorities seek to find the regions that contribute the most to the total number of daily new cases in the country. Due to mobility between geographical areas, causal relations exist amongst total daily new cases (rewards) of regions (actions). How can the optimal candidate regions be detected for political interventions while dealing with statistical dependencies?

My research during my PhD addresses the questions mentioned above and the like. During my postdoc, I was doing research at the intersection of imitation learning, active learning, and inverse reinforcement learning, to design machine learning methods for explaining sequential decision-making based on demonstrated behavior by a sub-optimal learner.

Prior to my PhD, I completed my master studies in Mathematical Modelling in engineering within an Erasmus Mundus international joint master program at University of Hamburg and obtained a bachelor’s degree in Mathematics from Iran University of Science and Technology.