Project for the course Online Learning Application. The course proposes the adoption of machine learning techniques, mainly in the field of online learning for real-world applications. In particular, the course focuses on four applications of great interest to the market: pricing in e-commerce scenarios, online matching applications, digital advertising, and social influence. Additionally, the course provides non-trivial learning techniques, such as non-stationary bandit algorithms, combinatorial bandits, Gaussian process-based bandits, Gaussian process regression, and learning from data generated by bandit processes. Students are required to produce prototypes of algorithms capable of working on data closely related to real-world applications.
The project required developing a scenario simulating customers buying products on an online e-commerce website with a certain demand curve. The general objective was to maximize the cumulative expected margin and minimize the regret for the seller deciding the best price to associate to each product using bandits algorithm (TS and UCB) and the information learned on the customer demand. Going deeper into the project, less information about the user were given and more difficult was to maximize profits and reduce losses. This means that more advanced bandit algorithm techniques were used.