ΑΙ@AUEB talk: "Spectral Algorithms for Ranking Regression" by Stratis Ioannidis, Northeastern University, USA, Tuesday, 28 June 17.15
ΑΙ@AUEB talk (hybrid presentation)
Tuesday, 28 June 17.15 (Greek time)
Room: Ground floor, Troias building
and virtually via MS Teams:
Speaker: Stratis Ioannidis, Northeastern University, USA
Title: Spectral Algorithms for Ranking Regression
Abstract: We consider learning from rankings, i.e., learning from a dataset containing subsets of samples ranked w.r.t. their relative order. For example, a medical expert presented with patient records can order them w.r.t. the relative severity of a disease. Rankings are often less noisy than class labels: human experts disagreeing when generating class judgments often exhibit reduced variability when asked to compare samples instead. Rankings are also more informative, as they capture both inter and intra-class relationships; the latter are not revealed via class labels alone. Nevertheless, the combinatorial nature of rankings increases the computational cost of training significantly. We propose spectral algorithms to accelerate training in this ranking regression setting; our main technical contribution is to show that the Plackett-Luce negative log-likelihood augmented with a proximal penalty has stationary points that satisfy the balance equations of a Markov Chain. This observation yields fast spectral algorithms for ranking regression for both shallow and deep neural network regression models.
Compared to state-of-the-art siamese networks, our resulting algorithms are up to 175 times faster and attain better predictions by up to 26%
Top-1 Accuracy and 6% Kendall-Tau correlation over five real-life ranking datasets.