Juho Rousu (Aalto)
Monday 2016-12-05 11.00 – 12.00
Lecture hall AS3, TUAS building
Predicting structured data
In modern data analysis, data does not always come in neat tables of numbers, each column corresponding to a variable of interest. In particular, many data have internal structure, with statistical dependencies arising between close parts of the structure. In my talk I will introduce structured (output) prediction, a branch of machine learning that aims to take advantage of the data structures (e.g. sequences, trees, networks) and the embedded statistical dependencies, in order to make predictive models more accurate and inference more efficient. I will illustrate the methods by two examples: Network response prediction (Su et al, ICML 2014), that asks to predict a subnetwork that responds to a stimulus arising in a node of a larger network, and Metabolite identification (Brouard et al. 2016, ISMB 2016), where the outputs correspond to graphs representing small molecules.