Starts 15 May 2024 11:00
Ends 15 May 2024 12:00
Central European Time
CMSP Seminar (Atomistic Simulation Webinar Series): Selecting and weighting informative features automatically with Differentiable Information Imbalance
Luigi Stasi Seminar Room (and via Zoom)
Romina Wild
(SISSA)
Abstract:
Feature selection is a common practice in many applications and tends to come with attached uncertainties such as:
What is the best dimensionality of a reduced feature space in order to retain maximum information?
How can one correct for different units of measure?
What is the optimal relative scaling of importance between features?
To solve these questions, we extend Information Imbalance, an effective statistic to rank information content between feature spaces, to its differentiable form. This Differentiable Information Imbalance is used as a loss function to optimize relative feature weights, simultaneously performing unit alignment and relative importance scaling. The same method can generate sparse solutions.
The optimal size of a reduced space is conveniently found by considering the change in the Differentiable Information Imbalance as a function of the number of non-zero features.
Feature selection is a common practice in many applications and tends to come with attached uncertainties such as:
What is the best dimensionality of a reduced feature space in order to retain maximum information?
How can one correct for different units of measure?
What is the optimal relative scaling of importance between features?
To solve these questions, we extend Information Imbalance, an effective statistic to rank information content between feature spaces, to its differentiable form. This Differentiable Information Imbalance is used as a loss function to optimize relative feature weights, simultaneously performing unit alignment and relative importance scaling. The same method can generate sparse solutions.
The optimal size of a reduced space is conveniently found by considering the change in the Differentiable Information Imbalance as a function of the number of non-zero features.
Zoom registration link to follow in due time.