PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond a Scalar: Distributional Serving Interfaces for Watch-Time Prediction

Xuan Liu, Jingbin Qian, Zhanyu Liu, Hefeng Zhou

arXiv:2609.28383v1cs.IR

TL;DR

Existing watch-time predictors generally serve a single expected or debiased estimate, even though downstream decisions may depend on probabilities over different outcome regions. DSI serves a compact summary of a learned joint distribution to lightweight readouts; across three datasets it achieves the lowest MAE on all three and the best XAUC on two, while supporting reuse for new targets.

  • Problem

    Existing methods generally serve one expected or debiased watch-time estimate, which does not expose probabilities for completion, overplay, or other task-relevant outcomes.

    Source: §1 · p. 1, §3.1 · p. 3

  • Method

    DSI estimates a joint distribution over event type and event time, then uses its frozen provider output for value and ranking readouts.

    Source: §3.2 · p. 4, §3.4 · p. 5

  • Results

    Across three datasets, DSI achieves the lowest MAE among nine baselines on all three and best XAUC on two, reducing MAE versus the strongest baseline by 1.9% to 8.5%.

    Source: §4.2 · p. 6, §5 · p. 9

  • Takeaways & Limitations

    The frozen interface supports new watch-time targets and a separately logged engagement signal using linear heads, while a randomly initialized provider does not reproduce the gain.

    Source: §4.4 · p. 8, §5 · p. 9

    The operational event types may include looping, cumulative playback, or logging artifacts, and the paper examines sensitivity to the exact boundary values.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Watch time is the primary engagement signal in short video feeds, and its prediction directly affects ranking and exposure. Existing methods improve watch time prediction by correcting duration bias or modeling richer distributions, but most expose only an expected or debiased watch time at serving time. Even when video duration is available to later models, the interface gives only one estimate of watch time and no probabilities for completion, overplay, or other regions relevant to downstream tasks. To address this limitation, we propose the Distributional Serving Interface (DSI), which has a distribution provider, a compact, low-dimensional summary, and lightweight readouts tailored to each task. The provider learns a joint distribution over four watch states derived from watch ratio and their event times; rules based on video duration remove incompatible combinations, while a restoration loss preserves accuracy in seconds. The summary reduces this distribution to a small set of event probabilities, time scales relative to duration, and uncertainty statistics. After training the provider, we fix its parameters and train value and ranking readouts that combine the summary with raw context. Across KuaiRec, KuaiRand-1K, and WeChat21, the complete DSI system achieves the lowest MAE on all three datasets, beating the strongest result among nine baselines by 1.9% to 8.5%, and achieves the best XAUC on two. It also leads retrieval metrics that account for video duration when complete systems are compared. With matched readouts held constant, the summary retains information relevant to each task beyond a predicted mean paired with video duration. Using the same lightweight linear heads for each new target, it also performs best on two new watch-time targets and improves a separately logged engagement target, while a randomly initialized provider does not reproduce this gain.

1 Introduction

The paper treats watch-time prediction as a serving-interface problem: a duration-aware scalar still cannot expose which outcomes may occur or when. DSI serves a compact summary of a learned outcome distribution to task-specific readouts, and achieves the lowest MAE among nine baselines on all three evaluated datasets and the best XAUC on two.

1 Introduction

  • Autoplay makes click-through signals less informative about interest, while watch time is logged for nearly every impression and directly shapes ranking and exposure.Source: §1 · p. 1
  • A predicted watch time paired with video duration still exposes only a conditional mean, not probabilities over completion, overplay, or other task-relevant regions.Different conditional distributions can share the same mean and duration; new deployed decisions may otherwise require reusing the point estimate or retraining the estimator.Source: §1 · p. 1
  • DSI freezes a distribution provider and exposes a compact summary to lightweight, task-specific readouts.Its three-stage instantiation derives operational watch outcomes, estimates their event times, and trains the readouts without retraining the provider for a new task.Source: §1 · pp. 1–2
  • Across three public datasets, complete DSI achieves the lowest MAE among nine baselines on all three and the best XAUC on two.Source: §1 · p. 2

2 Related Work

Related work improves watch-time estimation through duration-bias correction or richer distribution modeling, while DSI focuses on the information exposed to downstream decisions. The paper positions that reusable output as complementary to decision-focused, multi-task, and generative recommendation approaches.

Watch-time prediction

  • Watch-time predictors correct duration bias or model richer distributions [13], with several retaining distributions internally during training.Source: §2 · p. 2
  • Discrete competing-risks models estimate joint event-type and event-time distributions, whereas DSI uses deterministic operational partitions and exposes a compact distribution summary downstream.Its duration-dependent support follows coarse-data and interval-censoring theory.Source: §2 · p. 2

From prediction to decision

  • Decision-focused learning trains for downstream loss, while DSI instead asks what a trained watch-time estimator should expose for task-specific readouts.The paper describes this separation as motivated by Bayesian decision theory and complementary to shared-representation or shared-predictor multi-task architectures.Source: §2 · p. 2

Generative recommendation

  • Generative recommendation changes the prediction backbone or candidate generator, while DSI changes the representation passed from a watch-time estimator to downstream scoring.The paper states the two directions can be combined; related systems also generate watch time or use a learned watch-time reward [4].Source: §2 · p. 2

3 Method

DSI replaces a single watch-time estimate at serving with a compact summary of a joint event-type and event-time distribution. It freezes the provider before training lightweight task-specific readouts on the summary and raw context.

3.1 Problem Setup

  • At inference, a mean plus duration leaves probability around completion and overplay thresholds unspecified; DSI instead exposes a joint event–time distribution through G and task-specific H.The readout can produce a seconds-level value and a ranking score.Source: §3.1 · p. 3
  • Operational event targets are derived deterministically from logged watch time and duration: early, mid, completion, or overplay, with completion spanning watch ratios from 0.9 to 1.1.The thresholds are fixed before training; these event types are operational labels, not separately logged actions.Source: §3.1 · p. 3

3.2 Typed Event-Time Estimation

  • The provider masks duration-incompatible type–time pairs, anchors completion at the endpoint, and combines event-time likelihood with log-space restoration of observed seconds.On completion records, the likelihood and restoration terms can compete within the ±10% completion band.Source: §3.2 · p. 4

3.3 Typed Distribution Summary

  • The evaluated summary compresses the 276-value grid to 27 coordinates spanning event and time marginals, per-event time scales and ratios, entropy, and confidence margins.The event marginals and ratio vector preserve typed structure, while time, entropy, and margin statistics preserve scale and uncertainty.Source: §3.3 · p. 4

3.4 Decision Readout

  • With the provider frozen, a shared DCN readout combines the summary and raw context; its value branch predicts seconds while one overplay-trained score serves overplay and long-engagement retrieval.The value decoder is kept deliberately ordinary so the seconds-level comparison focuses on the information its head reads.Source: §3.4 · pp. 4–5

4 Experiments

Across three public short-video datasets, DSI leads the main watch-time comparison and is evaluated through controls that separate exposed representations from task-aligned readouts. With its provider frozen, its summary also transfers to new watch-time targets and a separately logged engagement target.

4.1 Setup

  • The evaluation uses KuaiRec, KuaiRand-1K, and WeChat21 with a fixed impression-level random 8:1:1 split and three seeds.The WeChat21 evaluation is limited to durations below 60 seconds and removes records with watch ratio at least 10.Source: §4.1 · p. 5, §4 · p. 5

4.2 Main Results

  • DSI has the lowest MAE on all three datasets, beats the strongest baseline by 1.9% to 8.5%, and has the best XAUC on two.It ranks second on WeChat21 XAUC, 0.0034 behind CREAD, and first on Long-N@5 and OP-N@5 across all three datasets.Source: §4.2 · p. 6, Table 1 · p. 7

4.3 Readout and Interface Effects

  • Under the shared readout, DSI leads 11 of 12 dataset–metric cells; D2Q is higher by 0.0004 XAUC on KuaiRand-1K.The overplay branch substantially lifts retrieval for every baseline relative to point-estimate ranking, so most of the main-table gap lies in the decision layer.Source: §4.3 · p. 6, Figure 3 · p. 8
  • The compact summary beats the duration-aware mean on MAE, XAUC, and OP-N@5 across all three datasets, though gains clear seed variation only on KuaiRec and WeChat21.Long-N@5 reverses on two datasets; the full 69-bin time marginal matches the compact summary on MAE but has the weakest retrieval.Source: §4.3 · p. 6, Table 2 · p. 7

4.4 Reusability of the Frozen Interface

  • With the provider frozen, DSI has the best mean on all four new-target metrics at both label fractions, and at 1% labels beats the strongest full-label baseline on each.The targets are a watch-ratio threshold and conditional 0.75 quantile on KuaiRec; the 1% condition uses 97,779 impressions, not a few-shot sample.Source: §4.4 · pp. 6–7, Table 3 · p. 8
  • For a separately logged KuaiRand-1K engagement target, adding the frozen summary improves AUC by 0.021–0.024 at every label budget.The random-provider summary performs on par with or below the bypass alone; NDCG@5 also improves, but only the 1% gain clears seed spread.Source: §4.4 · p. 8, Table 4 · p. 8

5 Conclusion

DSI turns a structured watch-outcome distribution into a compact serving interface for lightweight downstream readouts. Across three public datasets, it outperforms nine baselines on key prediction metrics and supports multiple objectives through a frozen provider.

5 Conclusion

  • DSI packages a structured distribution of watch outcomes into a compact serving interface for lightweight downstream readouts.The framework has three stages.Source: §5 · p. 8
  • DSI achieves the lowest MAE among nine baselines on all three public datasets and the best XAUC on two.The compact summary also improves MAE and XAUC over a mean paired with video duration when the provider and readout are fixed.Source: §5 · pp. 8–9
  • With a frozen provider, linear heads perform best on two new watch-time targets and improve a separately logged engagement target.A randomly initialized provider does not reproduce this gain.Source: §5 · p. 9
  • Task-specific readouts turn the same estimate into strong rankings for long engagement and overplay.The complete system also leads retrieval metrics that account for video duration.Source: §5 · p. 9

A Implementation Details

The implementation uses a typed event-time provider, a shared task readout, and common training settings for the baselines. Dataset statistics after impression-level preprocessing are reported in Table 5.

A Implementation Details

  • The provider uses 69 upper-edge time bins and shares their vector across the support mask, completion anchor, and restoration.Values above 1800 seconds map to the final bin rather than being discarded.Source: A Implementation Details · p. 10
  • The decision readout uses a matrix-parameterized DCNv2-style backbone with two full-rank cross layers and a two-layer deep branch.Cross and deep outputs are concatenated before task branches; the readout trains for five epochs and retains the checkpoint with lowest validation MAE.Source: A Implementation Details · p. 10
  • The value branch uses 30 training-quantile thresholds, while the ranking branch uses binary cross-entropy on the overplay event-type target.The ordinal branch uses λval = 30 and the ranking branch uses λrank = 0.5.Source: A Implementation Details · p. 10
  • Baselines share a common loader, fixed five-epoch training protocol, and DADF-style field adapter, while CWM, DIFL, and GR retain method-specific backbones.Table 6 records implementation sources and key settings; Table 5 reports dataset statistics after impression-level preprocessing.Source: A Implementation Details · p. 10, Table 6 · p. 10, Table 5 · p. 10

B Metric Definitions

The paper evaluates both continuous watch-time ranking and event-based retrieval, including duration-relative relevance for long engagement and overplay. The retrieval metrics are defined for this study rather than presented as standardized benchmarks.

B Metric Definitions

  • Long-N@5 uses completion-or-overplay relevance, while OP-N@5 uses overplay relevance defined by watch time divided by video duration exceeding β.DSI ranks with its overplay score; standalone baselines use predicted watch time, and matched-readout controls use the common readout score.Source: B Metric Definitions · p. 10
  • XAUC is a global pairwise continuous AUC over watch time, whereas the top-k retrieval metrics adapt NDCG-style ranking to operational event-type relevance.The authors state that the retrieval metrics are not standardized benchmark metrics.Source: B Metric Definitions · pp. 10–11
  • Long-N@5 and OP-N@5 average session-level N@5 equally over sessions with at least two candidates and positive ideal DCG.The ideal ordering is used to compute IDCG@5 for each session.Source: B Metric Definitions · p. 10

C Mechanism Ablations

The ablations isolate contributions from the typed summary, ranking branch, and restoration loss. Together, they show that the summary carries useful readout signal, ranking adds retrieval output, and restoration improves seconds-level prediction.

Input representation

  • Typed input beats raw input on MAE (3.5136 vs. 3.7645) and OP-AUC (0.8888 vs. 0.8440), and remains close to fused input.Adding the raw-context bypass lowers MAE from 3.5136 to 3.5059 but gives no consistent retrieval gain.Source: C Mechanism Ablations · p. 11

Readout branches

  • Adding the ranking branch changes retrieval but barely changes point estimation: MAE shifts by +0.0005 ± 0.0031 seconds and XAUC by less than 10^-4 on average.The trained branch reaches OP-AUC 0.8883 and OP-N@5 0.8808.Source: C Mechanism Ablations · p. 11

Restoration

  • MAE falls from 3.5185 to 3.5059 and XAUC rises from 0.6803 to 0.6827 when restoration is added to the support-masked likelihood.The paired MAE reduction is 0.0125 ± 0.0029 seconds, while OP-AUC changes by less than 0.001.Source: C Mechanism Ablations · p. 11

D Event-Boundary Sensitivity

The event-boundary sensitivity sweep finds that seconds-level MAE and WT-N@5 remain stable, while AUC movements under two sweeps reflect changed relevance labels.

D Event-Boundary Sensitivity

  • Seconds-level MAE and WT-N@5 remain stable across one-at-a-time event-boundary perturbations on all three datasets.The sweep uses defaults α=0.3, ρ=0.9, and β=1.1 for the other boundaries; each row retrains on the full training split with one seed.Source: D Event-Boundary Sensitivity · p. 11, Table 8 · p. 11
  • Long-AUC under ρ and OP-AUC under β measure target-definition sensitivity because those sweeps redefine long-engagement and overplay relevance.The α sweep leaves both retrieval labels fixed.Source: D Event-Boundary Sensitivity · p. 11

E Interface Robustness Controls

The controls find similar performance across provider-supervision and summary-form alternatives, with only small, task-specific differences. Their results are consistent with an interface-level reading, while ablations isolate the roles of the typed summary, ranking branch, and restoration.

E Interface Robustness Controls

  • Ablations show the typed summary carries most fused-readout signal, the ranking branch supplies OP-AUC without changing MAE, and restoration improves MAE with OP-AUC nearly unchanged.OP-AUC is undefined for the value-only branch.Source: Figure 4 · p. 12
  • The typed and time-only providers perform similarly on the main tasks, with small differences aligned to seconds-level supervision and overplay support.The typed provider has the only uniformly signed edge on OP-N@5; the time-only provider is slightly better where seconds-level supervision matters.Source: E Interface Robustness Controls · p. 12, Table 9 · p. 12
  • Typed and time-only providers also perform similarly on both transfer tasks, with differences within or near seed-level variation.This comparison uses the frozen-transfer protocol.Source: E Interface Robustness Controls · p. 12
  • The typed summary has a small, uniform OP-N@5 edge, but no consistent advantage over matched-dimensional duration-relative statistics.The plain summary is slightly better in MAE and XAUC on KuaiRec and slightly worse in MAE on WeChat21.Source: E Interface Robustness Controls · pp. 12–13, Table 10 · p. 13
  • Replacing provider supervision or summary form leaves performance essentially unchanged, consistent with an interface-level reading of the gains.The controls change one component while keeping the exposed interface and rest of the pipeline fixed.Source: E Interface Robustness Controls · pp. 11, 13

F External Engagement Association

The analysis compares event-type-derived relevance with an independently logged engagement mark, while keeping that mark out of model inputs and supervision.

F External Engagement Association

  • Long relevance and overplay relevance are derived from operational event types, not independently recorded actions.The analysis uses the KuaiRand-1K test split's same-impression engagement mark, defined as the union of separately logged like, follow, comment, and forward actions.Source: F External Engagement Association · p. 13
  • The separately logged engagement mark is excluded from model inputs and does not supervise the provider or ranking branch.Source: F External Engagement Association · p. 13
  • The association analysis reports aggregate and duration-stratified overplay associations.Source: F External Engagement Association · p. 13

G WeChat21 Protocol and Feature Parity

The WeChat21 evaluation uses a duration-below-cap subset and applies the same preprocessing rules to every method. Methods share processed features and evaluation code, with past-only user-history features.

G WeChat21 Protocol and Feature Parity

  • The WeChat21 evaluation removes impressions with d≥60 seconds because the metadata’s 60-second cap is not an exact video endpoint, and removes records with y/d≥10.The resulting dataset contains 6,037,792 impressions; both rules are fixed before training and applied identically to every method.Source: G WeChat21 Protocol and Feature Parity · p. 13
  • Within each dataset, all methods use the same processed split, raw-feature block, and evaluation code; DSI supplies its frozen estimator output to its readout.The provider objective and value and ranking branches follow the same design across all three datasets.Source: G WeChat21 Protocol and Feature Parity · p. 13
  • Past-only user-history statistics are shifted by one impression, preventing current and later outcomes from entering features.WeChat21 lacks intra-day exposure timestamps, so records retain dataset order within each user-date pair.Source: G WeChat21 Protocol and Feature Parity · p. 13

Related papers

Loading 2609.28383v1…