Description of the Application
Recommender systems play an indisputable role in our life as they
take part in various day-to-day decision making processes, ranging from
deciding which book to buy, which job to apply for, which movie to
watch, or which music to listen to. State-of-the-art algorithms powering
recommender systems nowadays operate on massive amounts of
user-generated data, typically given as interactions between system
users and available items; and sometimes enriched by content
descriptions of items (e.g., product features, music or movie genres) or
additional user information (e.g., demographics). These algorithms most
commonly involve computationally intensive matrix factorization
techniques or neural networks.
We research novel personalized recommendation methods, in particular
in the domains of music, videos, and microblogs. This involves computing
high-dimensional feature vectors from the raw data multimedia items are
composed of and creating detailed user profiles that describe aspects
such as users' preferences for item diversity or mainstream items in the
recommendation list. Also, training recommendation models, in particular
hybrid ones that integrate a wide range of user, item, and interaction
data, is a highly demanding and computationally complex task. So is
evaluating a variety of models trained on different datasets, using
different parameter settings, and evaluation metrics.
The massive computing capability and memory capacity of the MACH-2
allows us to conduct large-scale recommendation experiments,
investigating hundreds of algorithmic combinations in parallel, and
leveraging real-world data of billions of data points.