Using Hadoop MapReduce, Apache Spark, or another distributed computing technology, analyze the Netflix Prize Dataset. (Click the "Download" link in the upper right corner, not the "uci.edu" URL near the bottom of the page.) The README in that file describes the format of the data.
We're looking for movies that are well-loved by users who dislike movies most users like.
Find the M movies which have been rated the highest across all users (of movies which have been rated by at least R users). (If there's a tie for the Mth spot, prefer most recent publication then alphabetical order of title.) These are the "top movies."
Of users who have rated all top M movies, find the U users which have given the lowest average rating of the M movies. (If there's a tie for the Uth spot, prefer users with the lower ID.) These are the "contrarian users."