Preference orderings arise in many artificial intelligence fields when items are ordered by rankers based on their preferences. The presentation of a list of ordered items is often in the format of ordered sequence (from left to right). For example, the following lists show four preference orderings, produced by the four popular search engines when typing "London", where the ordered items are the terms suggested by the search engines:
1) Bing: London tube map, London stock exchange, London weather, London underground, London eye, London zoo
2) Duckduckgo: London has fallen, London broil recipes, London king, London time, London fog
3) Google: London Weather, London tube map, London, London eye
4) Yahoo: London underground, London tube map, London England, London weather, London eye, London Marathon 2016
Rank aggregation for preference orderings is the problem of finding a consensus ordering or consensus rank by combining the original preference orderings, and it is one of the fundamental and most common optimization problems in many AI fields. For example, in meta-search, a query is sent to several search engines, and results from the search engines are aggregated to produce a better result; in group decision making, rank aggregation can help experts adjust their preference orderings, in order to reach a better decision collectively; in crowdsourcing, where users obtain their requested services from Internet, rank aggregation has become one of the biggest challenges to combine large volumes available resources; in recommender systems, aggregating users' preferences over products has been of interest to both industry and academic researchers.
Rank aggregation for preference orderings is also important to the other subjects, ranging from social science to bioinformatics. In social science, rank aggregation has been used to analyze life-course patterns; it is widely studied for determining the winner in competitions and social voting. In bioinformatics, researchers have used rank aggregation to analyze biomedical ordering sequences.
This project aims to develop efficient algorithms on rank aggregation for preference orderings by mapping preference orderings to higher dimensional spaces and by using GPU computation. In order to achieve this, this project will: (1) estimate the median for a set of preference orderings in higher dimensional spaces, because the median is an important factor for evaluating rank aggregation; (2) detect outliers from a set of preference orderings, because the consensus rank can be skewed by the outliers in preference orderings; (3) develop fast learning algorithms to aggregate preference orderings by using GPU computation, because the existing methods are computationally expensive.
The research findings from this project will be published in high-impact conferences and journals to advance the research in the area. The algorithms developed in this project will be released as open source, publicly available on github to benefit academic researchers and industrial developers.
|