Same audit methodology as the e-commerce case, five orders of magnitude apart: 35 rows there, 25 million here.
An audit of 25,000,095 movie ratings, 1995–2019, by Alex Kwon. Verdict in one line: the data is immaculate; the ruler moved. Every structural promise the dataset makes holds exactly, and its 25-year timeline is three measuring instruments stitched together at seams its own documentation never mentions.
Ratings per year. The shaded bands are the three instruments; the marks are events the timestamps confess: the 1997 seam (a 48-day hole between the imported EachMovie data and native MovieLens), the Nov 1999 press wave (Gladwell, Nightline, Ebert: a 10× jump in one month), the 2003 half-star switch, and the 2014 relaunch.
Marks: ✂ the 48-day seam · ▲ the media wave · ½ the scale change · v4 the relaunch. Hover any bar for its number.
822 movies have 200+ ratings in both whole-star eras. Band them by how the native era scores them, and compare the same films across eras: a clean staircase. Movies the native era calls terrible looked 0.8 stars better under the imported instrument; the beloved barely moved. Consistent with the six-to-five scale conversion squeezing the bottom, and with 1996 audiences meeting these films as new releases. Either way: averages that cross 1997 inherit it.
Small n at the extremes (3 and 29 movies); the finding rests on the monotonic gradient across all seven bands, not on any single band.
Heavier raters score lower, and they cast most of the votes. The dot is each group's average rating; the right column is their share of all 25M ratings.
Users under 20 ratings are excluded by the sample's own design; the casual web visitor does not exist in this dataset at all.
| Minimum ratings | Winner | Average | Ratings |
|---|---|---|---|
| none | an 820-way tie of perfect scores | 5.00 | 1 each |
| 1,000 | Planet Earth II (television) | 4.48 | 1,124 |
| 10,000 | The Shawshank Redemption | 4.41 | 81,482 |
855 movies hold a perfect 5.0; 820 of them earned it from exactly one person. Honest rankings publish the floor next to the winner. Most-rated is a different question: Forrest Gump, 81,491 — nine ratings ahead of Shawshank.
Also on the books: three timestamp scars (three ratings on one 1995 second, the 48-day 1997 seam, and 506 accounts all terminating on the morning of May 22, 1998, the signature of an undocumented batch migration).
Analysis: Alex Kwon. Definitions, methods, and every query: movielens-case-study (FINDINGS.md, REPORT.md, sql/01–04, setup.py rebuilds the database from the canonical GroupLens download). Receipts: Konstan et al. 1998 (AAAI WS-98-08) for the EachMovie seed; Harper & Konstan 2015 (ACM TiiS) for the version history. Companion case: ecommerce-case-study. Every figure independently recomputed before publication.
Cite: Kwon, A. (2026). Three Rulers, One Dataset: an audit of MovieLens 25M. github.com/collapseindex/movielens-case-study (machine-readable: CITATION.cff in the repository). License: report CC BY 4.0; code Apache-2.0. Disclosure: AI-assisted implementation; methodology and validation framework developed in prior research; every figure recomputed from raw data by the runnable code in the repository.
Alex Kwon · ask@collapseindex.org · more case studies · github.com/collapseindex