Johan Pouwelse

Master thesis kick-off meeting - Re-using the knowledge of master courses: speciality of math and statistics - Regular weekly meetings at Delft lab. - Make it work! its the operational...

- No checkpoints please, no deeplearning4j, no deep learning at all, no GUI, No APK, no library, no decentralisation, no big data, no attacks (pollution,Sybils,DoS, false promotion, etc.) - [Single...

Recent related work: https://arxiv.org/pdf/2105.04761.pdf "Federated Unbiased Learning to Rank", by Apple Inc.

- ToDo: - post first Kotlin ML code from repo - playlist dataset shortcut - first music recommendation - Graph on your thesis ticket! "_[An error occurred accepting the competition...

Related work, please study and explain to me: https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf

Starting project, simple gradient descent without any library bloat. Evolve further: [Learning to Rank using Gradient Descent](https://icml.cc/Conferences/2015/wp-content/uploads/2015/06/icml_ranking.pdf). Most simple possible machine learning and most simple distributed systems algorithms: 1 packet...

- Focus on Learn-to-Rank. Keep it as simple as possible: gradient decent if possible (performance is not a concern) - Dataset is needed of user, keyword-search, object-clicked, rank-when-clicked - As...

- Learn-to-Rank Grand idea: deploy system which gets better with usage. - Stop with the cheating and shortcuts (e.g. Python, kmeans, jupyter) - Focus: clicklog, gradient decent, and "use a...

- Goal: be the world-first! Nobody has academically pure decentralised federated learning deployed. So without any cheating, no cloud, no coordinator, no rendezvous server, and no central server of any...

- cold start problem: there is no click-log data and no user search queries. - Can we have a few lines bootstrap dataset, not MvP style. Toy example only :warning:...