With final blocks held out, tied targets allow either top-10 order of two count rules in MovieLens 100K and 1M

Movie rating logs can assign the same final timestamp to several ratings from one user. We tested how this ambiguity changes a fixed-model benchmark in MovieLens 100K and 1M. We excluded each user's full final timestamp block from training. We fit three score rules once and calculated exact top-10 hit-rate bounds over all choices of one final target per user. Equal model scores received their expected credit under uniform score order. Final timestamps were tied for 415 of 943 users in 100K and 2,544 of 6,040 users in 1M. In an exploratory comparison, the paired difference between all-rating count and positive-rating count ranged from -0.954 to +5.355 percentage points in 100K and from -0.902 to +1.358 points in 1M. Each range allows either model order. In 1M, selecting the first rather than last source row among tied targets reversed that order. In the primary model comparison, all-rating count exceeded a shrunk movie mean for every target choice in 1M, with a minimum lead of 0.116 points. The package gives the finite-data method, code, aggregate tables, figures, source hashes, and independent calculation checks. These bounds keep training fixed. They do not cover ordinary leave-one-out retraining, future deployment, or movie viewing order.

MovieLens · timestamp ties · recommender evaluation · partial identification · rating behavior · reproducibility