|BERNARDIS CESARE||Cycle: XXXIV |
Section: Computer Science and Engineering
Tutor: TANCA LETIZIA
Advisor: CREMONESI PAOLO Major Research topic
:Reproducibility and Replicability issues in Hybrid Recommender SystemsAbstract:
Recommender Systems (RS) are software tools used to provide suggestions for items or other entities to the users by exploiting various strategies. Depending on the type of information used to generate recommendations, RS can be divided in different categories. Hybrid Recommender Systems (HRS) try to combine multiple categories together, in order to obtain all their benefits and limit their drawbacks in a single model. Moreover, they can face scenarios that approaches belonging to single categories can not face independently (cold-start, ramp up, ...).
Recent studies have raised the attention on reproducibility and replicability issues that affect many works belonging to the literature. The source code used to perform the experiments is rarely disclosed, various evaluation procedures are adopted without a precise justification, weak baselines are propagated from work to work and a careful hyper-parameter optimization is often not performed.
As a consequence, it is hard to replicate the results obtained by the authors of many works belonging to the state of the art, and often the respective findings can be questioned. Of course this strongly impacts the quality of the research.
There is a big interest in the community in shedding light over these problems, finding solutions and guidelines in order to improve replicability and reproducibility in future works.
While some research in this direction has been performed on deep learning based collaborative approaches, there is a lot of effort to be put in hybrid approaches.