Academic research repository
The ranking is informed by research in measurement, statistics, recommender systems, social influence, and social-choice theory.
Ratings as measurements
A rating depends on who watched, who voted, the scale, and the aggregation rule. Agreement across sources raises confidence, but correlated popularity and demographic biases can create shared error.
Shrinkage and uncertainty
Bayesian and empirical-Bayes shrinkage reduce small-sample winner’s curse. Future versions should expose intervals and sensitivity, not only point estimates.
Social influence
Muchnik, Aral, and Taylor’s randomised experiment showed that prior positive signals can create persistent upward bias. Ratings can influence later opinions rather than merely record them.
Recommender systems
MovieLens and GroupLens established collaborative filtering benchmarks. Academic evaluation also considers novelty, diversity, calibration, transparency, and user control.
Rank aggregation
Weighted averages are transparent but weight-sensitive. Kemeny methods minimise pairwise ranking disagreement, but are computationally difficult and can hide score magnitude.
Representation
Popularity and source coverage favour widely distributed English-language titles. Separate Swiss and Turkish collections reduce, but do not eliminate, geographic and language bias.
Selected reading
- Muchnik, Aral & Taylor (2013), Social Influence Bias
- GroupLens MovieLens datasets
- Herlocker et al. (2004), Evaluating Collaborative Filtering
- Kemeny ranking aggregation overview
- Shrinkage and empirical Bayes notes
- IMDb non-commercial datasets
- Rotten Tomatoes score overview
- Swiss FDPIC data-protection basis
Research agenda
- Measure rank sensitivity to source weights and missingness.
- Compare critic–audience disagreement by genre, country, and decade.
- Estimate uncertainty and minimum evidence thresholds.
- Test whether extra sources improve satisfaction rather than only reshuffle ranks.
- Audit language, geography, gender, and popularity bias.