Wydarzenia w dniu 06.05.2019:
SEMINARIUM ODWOŁANE:
SEMINARIUM
Oddziału Fizyki Dielektryków i Spektroskopii Molekularnej
Dnia: 06.05.2019 roku (poniedziałek)
o godzinie 10:00 w auli
referat pt.:
Lepkościowe konsekwencje samoorganizacji molekuł dioli
wygłosi
SEMINARIUM
Oddziału Fizyki Dielektryków i Spektroskopii Molekularnej
Dnia: 06.05.2019 roku (poniedziałek)
o godzinie 10:00 w auli
referat pt.:
Who is the winner? Artificial Intelligence and Machine Learning in the Science of Rating and Ranking
wygłosi
University of Bergen, Norway
We are living in a data-driven world. Advances in artificial intelligence (AI) could impact nearly all aspects of society: the labor market, transportation, healthcare, education, and national security. The booming IT industry collects large amounts of user preference and behavioral data to make various decisions. In many cases, these data are in diverse rank format instead of numerical values over alternatives – for example, how voters rank candidates, consumers choose one product over another, search engines rank webpages. Ratings and rankings are everywhere, but how exactly do they work? Who is the winner? This seminar presentation aims to share and discuss recent progress in machine learn-ing and decision-making from rank-data as well as provide an overview of the funda-mental ideas behind mathematical rating systems, i.e. reference to classical properties and algorithms for handling rank data.
Preference aggregation is the process of combining multiple individually ranked lists (of alternatives) towards choosing a winner from the list of options. The question we ex-plore here is can winners or representative rankings be predicted using machine learn-ing methods. The primary research objective conducted in experimental study is learna-bility of linear ranking of the alternatives in pairwise comparison. With its different con-figurations, the set of agents (voters) have preferences (votes) over a set of alternatives (candidates). Taking as input the preferences of all agents (so called profile), the mech-anism outputs an aggregated preference ranking of all alternatives or calculate the out-come as a single winner. Specifically, we focus on winner prediction based on the fol-lowing voting rules:
1. Kemeny voting (chooses the ranking which is closest to the individual rankings based on the total number of pairwise switches.)
2. Borda voting (which assigns n-1 points to a top ranked choice, n-2 points to sec-ond ranked choice, down to 0 points for a bottom ranked alternative. Finally, ranks alternatives of n candidates according to total number of points.)
3. Dodgson voting (which elects Condorcet winner by making swaps of adjacent candidates. The winner is the candidate that needs the minimum number of swaps).
Clearly, aggregated rank learning problem has a strong impact on identifying the winner, as determining Kemeny winner is NP-hard (over 4 candidates). The experimental study performs a comparison of several machine learning methods. In particular we explore XGBoost, Linear Support Vector Machines, Multilayer Perceptron and regularized linear classifiers with stochastic gradient descent for Borda, Kemeny and Dodgson methods with the goal of establishing the best trade-offs between search time and performance. We analyze the performance of our approaches on two datasets: one obtained from Spotify and a high-dimension synthetic dataset.
Key words: rank aggregation, preference aggregation, supervised machine learning, multi-agent systems, computational social choice