Most test ratings share a training session under 80/20 user time splits and a 30-minute gap rule in MovieLens 100K and 1M

We tested whether time splits separate rating sessions in MovieLens 100K and 1M. For each user, training used the first floor(0.8n) rating rows. Test used the remaining rows. A new session started when the gap between adjacent ratings exceeded 30 minutes. In 100K, 15,878 of 20,381 test ratings shared a session with training: 77.91 percent. In 1M, the count was 142,021 of 202,451: 70.15 percent. The majority result held at three training shares and three gap limits in each release. Removing shared-session test rows kept test data for only 27.25 percent and 30.26 percent of users. Those users had more total ratings than removed users. Moving the split to a session boundary with at least 20 training rows covered 35.74 percent and 40.25 percent of users. A separate source scan matched 19 primary fields per release. A full repeat matched all six analysis files. Earlier work already described rating bursts and time-split problems. This contribution adds exact session overlap counts, coverage costs, and an audit tool for these two releases. Sessions are a gap-based proxy for visits. Rating time is not viewing time. We did not measure prediction error or a causal effect. The package has code, source URLs and hashes, aggregate tables, figures, and repeat instructions. Original source records are excluded under their redistribution terms.

MovieLens · rating sessions · human behavior · time splits · evaluation · reproducibility