Research projects
My ongoing work is within the fields of recommender systems and music information retrieval:
I did some other work in recommender systems and within the broader field of artificial intelligence:
Ongoing
• Explainability of music playlists
Dave
Musical tracks, or songs, are independent entities. Even so, songs are connected by hidden links behind the bare audio wave. Dave is a Research prototype by Giovanni Gabbolini and Derek Bridge, and is designed to unveil such links, or segues.
Play with Dave Click the Spin button, and check out segues generated by Dave from pairs of randomly-chosen songs.
For more datails:
Giovanni Gabbolini and Derek Bridge. 2021. Generating Interesting Song-to-Song Segues With Dave. In UMAP ‘21.
The paper on Dave was awarded as:
- James Chen best student paper award at UMAP ‘21;
- Best paper award ‘21 within the school of computer science in UCC.
And, Derek had the chance to talk about our work on the BBC Radio 6 Radcliffe and Maconie Show! Listen back.
IPSim
We use segues from Dave to build an interpretable and accurate similarity measure, called IPSim.
For more details:
Giovanni Gabbolini and Derek Bridge. 2021. An Interpretable Music Similarity Measure Based on Path Interestingness. In ISMIR ’21.
Sam
We use segues from Dave to decorate playlists. As such, we create tours, that is a sequence where songs alternate with segues.
For more details:
Giovanni Gabbolini and Derek Bridge. 2021. Play It Again, Sam! Recommending Familiar Music in Fresh Ways. In RecSys ‘21;
Past
• Recommender Systems and Instability
In this project, jointly with Edoardo D’Amico (UCD) and the RecSys@PoliMi group, we show how the output of matrix factorization change when the initialization of embeddings change. We refer to this phenomenon as instability.
We introduce a new family of algorithms that generalize matrix factorizations. The new algorithms are more stable, and have better accuracy on the long-tail.
This project was the subject of two papers and of my M.Sc thesis:
Giovanni Gabbolini, Edoardo D’Amico, Cesare Bernardis, and Paolo Cremonesi. 2021. Analyzing and Improving Stability of Matrix Factorization for Recommender Systems. In JIIS.
Edoardo D’Amico, Giovanni Gabbolini, Cesare Bernardis, and Paolo Cremonesi. 2021. On the instability of embeddings for recommender systems: the case of matrix factorization. In SAC ‘21.
Giovanni Gabbolini, Edoardo D’Amico: Exploiting the Long Tail Recommending Less of More, Nearest Neighbors Matrix Factorization. In PoliTesi (link)
• ACM RecSys Challenge 2019
I participated to the RecSys Challenge 2019 as a member of the PoliCloud8 team, inside the RecSys@PoliMi group.
The challenge, sponsored by Trivago, focused on travel metasearch. We had to develop a session-based and context-aware recommender system using various input data to provide a list of accommodations matching user needs. We were given a dataset of $20$M interactions, $1$M users and $1$M accommodations.
In the final leaderboard, our team ranked in the $10^{th}$ position over $575$ teams. Our solution was based on a stacked ensemble of three algorithms powered by over $200$ handcrafted features. We explain in details our solution to the challenge in a paper, published at RecSys:
Edoardo D’Amico, Giovanni Gabbolini, Daniele Montesi, Matteo Moreschini, Federico Parroni, Federico Piccinini, Alberto Rossettini, Alessio Russo Introito, Cesare Bernardis, and Maurizio Ferrari Dacrema. 2019. Leveraging laziness, Browsing-Pattern Aware Stacked Models for Sequential Accommodation Learning to Rank, Procs. of the ACM Recommender Systems Challenge 2019, 2019.
• Face2Face
In the context of the Deep Learning class @ Politecnico di Milano, I contributed to design and implement a novel face authentication algorithm. The algorithm is based on Convolutional Neural Networks (CNNs).
For more details, see a report and the source code.