Pairs at equal times carry most rating match excess in MovieLens 100K and 1M under uniform tie weights and a user shuffle reference

We tested adjacent star-rating matches in the complete MovieLens 100K and 1M releases. We gave equal weight to all orders within each user and timestamp group. A uniform shuffle of each user's full rating history supplied the reference. Total match excess was 5.88 and 6.18 percentage points. Pairs at equal recorded times carried 69.35% and 60.86% of the excess counts. The 95% user resampling ranges for these shares were 66.73% to 71.88% and 59.77% to 61.90%. Between distinct times, equal weight match rates were 33.53% and 35.43%, above references of 29.85% and 30.26%. An exact dynamic program kept the first and last labels of each time group coupled. It gave attainable boundary match limits of 17.54% to 54.92% and 18.84% to 58.28%. Both intervals contain the reference. A selected subset with one-record groups on both sides retained excess matches without order ambiguity. Thus, repetition is not limited to tied times, but the full boundary result depends on how unknown order is treated. The package contains the method, proof, original code, source URLs and hashes, aggregate tables, all bootstrap values, and figures. Separate parsing, exhaustive order checks, input-row shuffling, and a complete repeat passed. The results describe rating entry in selected users of one service. They do not identify viewing order, a causal memory process, or recommendation performance.

MovieLens · human behavior · rating repetition · timestamp ties · partial order · dynamic programming · reproducibility