W dniu 21 września 2023 r. (czwartek) o godz. 11:00 w IBS PAN (sala 200) odbędzie się spotkanie promocyjne projektu BIPOLAR („Bipolar disorder prediction with sensor-based semi-supervised learning”), realizowanego w Instytucie Badań Systemowych PAN w latach 2022-2023.
Podczas spotkania Pan Kamil Kmita wygłosi referat pt. „Controlling Impact of Partial Supervision in Semi-Supervised Fuzzy C-Means by means of explainability” (streszczenie poniżej).
Podczas referatu przedstawiona zostanie między innymi praca zatytułowana: „Confidence path regularization for handling label uncertainty in semi-supervised learning: use case in bipolar disorder monitoring” (DOI: 10.1109/FUZZ-IEEE55066.2022.9882759), która uzyskała nagrodę „Best Paper Award” na konferencji 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE).
Wyniki projektu publikowane są na stronie: http://bipolar.ibspan.waw.pl/
Streszczenie referatu
Semi-Supervied Fuzzy C-Means (SSFCM) is a classical model enhancing unsupervised Fuzzy C-Means with capability to handle partial supervision in form of categorical labels. Controlling how much this additional information impacts the results of clustering is of uttermost importance. SSFCM introduces a single hypeparameter called the scaling factor to control this impact. In this presentation, we shed more light on how the scaling factor has been interpreted in the literature, and what are the far-reaching consequences of such a mechanism of controlling impact of partial supervision. We use the explainability framework to conduct our analysis. In particular, we focus on the relation between the scaling factor and two variables being optimized by the SSFCM model – memberships matrix U (that denotes how much observations belong to different clusters) and clusters prototypes matrix V (that numerically represents gravity masses of the clusters).
Projekt BIPOLAR otrzymał finansowanie z funduszy norweskich w ramach SMALL GRANT SCHEME Call,
grant NOR/SGS/BIPOLAR/0239/2020-00.
Serdecznie zapraszamy na spotkanie !
