Smolnikov Research / Yambda temporal baseline study / protocol v1
Frozen before evaluation: 2026-11-06. This is not MLPerf execution or neural model training.
Data: yandex/yambda revision dd6f3a19eef5866e346c3270e098baa641a44948, flat/50m/likes.parquet. Use entire file, not first rows. No private data or paid APIs.
Question: How do global temporal cutoffs and recency affect simple next-like recommendation, and how does a deliberately invalid future-informed control differ?
Cutoffs: floor(min timestamp + fraction*(max timestamp-min timestamp)), fractions 0.6,0.7,0.8. These are fractions of observed raw time span, not row quantiles. Retrospective descriptive study, no external preregistration. No tuning after results.
History: timestamp < cutoff. Evaluation: each user with history and a future like is represented by all distinct item IDs at that user's first timestamp >= cutoff. No user joins across cutoffs as independent observations. No temporal gap; this differs from official Yambda protocol.
Candidates: all item IDs appearing in historical likes at that cutoff. Unseen targets remain in evaluation denominator and cannot be hit. Repeated items allowed. Do not sample negative items.
Methods: past_pop = total past like-event count; recent_pop = like count in last 20% of available historical time span; future_control = total count in entire file, restricted to same past catalog (deliberate leakage, never deployable). Rank by count descending, item_id ascending. Same global top-10 for all users. Recent method falls back to zero scores for catalog items not in recent interval. No personalized or neural model.
Metrics: per-user Hit@10 (any target hit), Recall@10 (fraction of target set hit), NDCG@10 (binary relevance, ideal min(10,target size)). Macro average. Wilson 95% interval for Hit only; descriptive uncertainty conditional on observed cohort. Paired Hit differences with normal approximation 95% intervals; multiple comparisons exploratory, not adjusted. Report all cutoffs, no best-only reporting.
Subgroups: all, organic-only target group, algorithmic-only, mixed. The is_organic flag is observational, never a causal treatment. Report sizes and cold-target share, exact duplicates, timestamp ties, counts/unique IDs. Do not infer national representativeness.
Separate reanalysis: pin MLCommons training e309244f7f81ecb00da434cae98355a73fc50ff0, all 60 recommendation/rcp_logs logs. Parse first eval_accuracy>=0.75, elapsed since run_start, samples from epoch_num*train_samples if necessary; compare 3 batch groups, do not treat logs as own training or cross-vendor comparison. Retain all crossings, sample CV, GPU-minute proxy (8/16/32 GPUs per upstream); not measured energy/cost.
Execution: isolated KotikTurbo CPU, one thread, <=1500MiB, shared tests.lock, Moscow 10:00-24:00. Pin dependencies, hashes, commands, environment. Publish code/aggregate metrics/licenses, not raw user histories. Independently recalculate key results via a different code path.