Beyond a Scalar: Distributional Serving Interfaces for Watch-Time Prediction
Xuan Liu, Jingbin Qian, Zhanyu Liu, Hefeng Zhou
cs.IR
2026-09-24
DSI freezes a typed watch-time distribution into a 27-d summary for light readouts. MAE beats nine baselines by 1.9–8.5% on KuaiRec, KuaiRand-1K, and WeChat21.
Short-video feeds barely use clicks. Watch time is the main engagement signal, and it feeds ranking and exposure. Existing methods already correct duration bias or fit richer distributions in training, but at serving time most of them still emit one expected or debiased number of seconds. Even if a later model sees video duration, the pair is still one conditional mean plus \(d\).
Ten seconds is near completion on a 12-second video and an early exit from a 60-second one. The same \((\hat y, d)\) can come from distributions that put very different mass on completion and overplay (watch ratio above duration). A new decision then either reuses that point estimate or retrains the estimator.
DSI turns one logged watch record into a reusable serving interface in three stages.
No new label is collected. Watch ratio \(\gamma=y/d\) is cut into four operational events: early (\(\gamma<0.3\)), mid (0.3–0.9), completion (0.9–1.1, pinned at duration), and overplay (\(\gamma>1.1\)). Time is discretized into 69 bins, one-second resolution through 60 seconds, then wider up to 1800 seconds. Inside the completion band, 95, 100, and 108 seconds on a 100-second video share the completion coordinate.
The provider outputs a joint distribution over the 4×69 grid and masks illegal type–time pairs given duration. Early cannot sit past the duration; completion can occupy only the duration bin. The loss is a support-masked likelihood plus a log-domain SmoothL1 restoration between the distributional mean and logged seconds.
The provider is then frozen. The 276-cell grid is reduced to a 27-dimensional summary: event probabilities and logits, long-watch mass, \(\log(1+\mu)\) and \(\mu/d\), per-type time scales, entropies, and margins. A light DCNv2 readout on the summary plus raw context predicts seconds on one branch and an overplay ranking score on the other. That score is also evaluated for broader completion-or-overplay retrieval. A later task fits a new head on the same frozen summary.
On KuaiRec, KuaiRand-1K, and WeChat21 (\(d<60\), drop \(y/d\ge 10\)), full DSI has the lowest MAE among nine baselines (VR, WLR, D2Q, CWM, DIFL, TPM, CREAD, EGMN, GR), 1.9% to 8.5% below the strongest baseline on each set. XAUC is best on the first two; on WeChat21 it is 0.0034 behind CREAD. Long-N@5 and OP-N@5 lead on all three.
| Dataset | DSI MAE | Best baseline MAE | DSI XAUC | Best baseline XAUC |
| KuaiRec | 3.5059 | 3.8327 EGMN | 0.6827 | 0.6605 D2Q |
| KuaiRand-1K | 12.3244 | 12.5668 D2Q | 0.7367 | 0.7345 D2Q |
| WeChat21 | 12.9278 | 13.2964 D2Q | 0.7020 | 0.7054 CREAD |
Controls separate the interface from the head. Under one shared readout, D2Q / TPM / EGMN / DSI frozen representations show that most of the retrieval jump comes from the task-aligned ranking branch. DSI still wins 11 of 12 dataset–metric cells, with much smaller margins than the main table. Inside one provider, the 27-d summary beats \([\log(1+\mu),\mu/d]\) on MAE, XAUC, and OP-N@5. The OP-N@5 edge matches the design: overplay is explicit in the summary. The 69-bin time marginal matches MAE and is weakest on retrieval, because the event-type axis has been integrated out.
With the provider frozen, linear heads on KuaiRec for watch-ratio \(>0.5\) and the 0.75 quantile beat the strongest full-label baseline already at 1% labels (Ratio-AUC 0.8852 vs CREAD 0.8746). On KuaiRand-1K, like / follow / comment / forward, a label never used to train the provider, gains 0.021–0.024 AUC over raw context; a randomly initialized provider does not.
Short-video ranking rarely optimizes one number of seconds. Completion, overplay, duration-normalized scores, and later engagement heads want different slices of the same distribution. DSI's claim is narrow: the estimator should serve a small summary, and each task head should read it. That is the same staged-interface idea as industrial scoring, calibration, and reranking, with a distributional object instead of a scalar.
A large share of the main-table retrieval lead is the aligned head, not a magic representation. After matching the readout, the summary still carries residual information beyond a duration-aware mean, most clearly on overplay retrieval. The interface is worth the extra machinery when new targets keep arriving.
The four events are deterministic slices of watch ratio, not independently observed actions. The logs cannot tell deliberate replay from auto-loop or logging artifacts. The main split is a warm-start 8:1:1 random split at impression level; users and items repeat across splits, with no chronological shift. WeChat21 is further filtered on duration and extreme ratios. Matched readouts erase much of the main-table retrieval gap, so complete-system numbers are not representation quality. A time-only provider and an unstructured 27-d summary perform similarly in the appendix, so the typed grid is not the only working parameterization. There is no online A/B.