Deconfounding Duration Bias in Watch-time Prediction for Video Recommendation

1 INTRODUCTION

• Watch time is an engagement signal that platforms seek to predict accurately when recommending candidate videos to users.The introduction identifies watch time as a scarce resource of user attention and links accurate prediction to engagement and daily active users.Source: §1 · p. 1
• As recommendations pursue greater watch time, long-duration videos can become overexposed, undervaluing users’ real interests and risking bias amplification that harms diversity and personalization.Duration bias is less explored than popularity- or position-related biases in recommendation studies [1] [2] [31] [39] [40].Source: §1 · p. 2
• The paper models duration as a confounder that affects both watch-time prediction and video exposure, distinguishing the natural effect to preserve from the exposure effect to eliminate.Duration’s influence on video exposure represents unfair preference for long videos; its effect on watch time reflects users spending longer on intrinsically longer videos.Source: §1 · p. 2
• D2Q uses duration-based data groups and adjusted watch-time quantile labels to remove exposure bias while sharing model parameters across groups.The framework preserves duration’s natural effect on watch time while removing its effect on video exposure.Source: §1 · p. 2
• Offline evaluation and live experiments report advantages over existing baselines, including more accurate watch-time predictions and improved real-time video consumption on Kuaishou.The authors also report that performance first improves and then declines as the number of duration groups increases.Source: §1 · p. 3

2 RELATED WORK

• Weighted Logistic Regression (WLR) estimates expected watch time from watch-time-weighted samples, but long videos can receive larger weights that amplify duration bias.The authors adapt WLR for both streaming services without unclicked samples and settings with click and watch behaviors.Source: §2 · p. 3
• Recommender-system bias methods include causal embeddings [3] [40], inverse propensity weighting [26], and causal intervention [31] [37] [39].The paper classifies its approach as causal intervention, removing duration’s effect on video while preserving its effect on watch time.Source: §2 · p. 3
• Backdoor adjustment uses do-calculus to intervene on relevant variables and remove an undesired causal edge [21].The passage notes applications in healthcare [14], bioinformatics, socioeconomics, and recommender-system popularity-bias adjustment for CTR prediction [31] [39].Source: §2 · p. 3

3 CAUSAL MODEL OF WATCH-TIME PREDICTION

• The causal graph treats duration as affecting both watch-time prediction and video exposure, with the latter reflecting recommender preference for long videos.The graph also represents interest effects from users and videos to watch time, and duration may affect training through differing sample sizes and weights such as WLR.Source: §3 · pp. 3–4, Figure 4 · p. 4
• The model’s variables are user representation U, video representation V, video duration D, and watch time W.U includes demographics, context, and interaction history; V includes video topics.Source: §3 · pp. 3–4
• Duration affects watch time along D→W and D→V→W paths; the direct path reflects the natural effect to capture, while the exposure path can bias video distribution toward longer videos.If the exposure path is not removed, the passage says predictions risk feedback-loop bias amplification.Source: §3 · p. 4

4.1 Deconfounding Duration

• Backdoor adjustment averages group-conditional watch-time predictions over the marginal duration distribution after removing duration’s effect on video exposure.The paper proposes discretizing the duration distribution and estimating watch time within each group.Source: §4.1 · pp. 4–5

4.2 Data-Splitting based on Duration Quantiles

• Duration-quantile splitting forms groups to reduce interference from long-watch-time samples, but separate models per group would grow too large for production use.With shared parameters and original watch-time labels, training is equivalent to not splitting the data.Source: §4.2 · p. 5

4 BACKDOOR ADJUSTMENT FOR DURATION BIAS

• D2Q removes duration’s biasing effect on video exposure while preserving its intrinsic effect on watch time, using duration-dependent quantile labels to support parameter sharing.The framework is designed to achieve duration debiasing and parameter sharing together.Source: §4 · p. 4, §4.3 · p. 5

4.3 Estimating Watch-time per Duration Group

• D2Q predicts watch-time quantiles across duration groups with one shared model, then maps each prediction through its group’s empirical watch-time distribution.The method finds the new video’s duration group before mapping the predicted quantile back to a watch-time value.Source: §4.3 · p. 5
• Res-D2Q adds a duration-adjustment tower and further improves prediction accuracy; increasing group counts can eventually hurt performance as groupwise distribution estimates become less reliable.The paper also notes that duration should remain a model input to distinguish samples across groups.Source: §4.3 · p. 5

5.1.1 Data.

• The offline Kuaishou dataset contains 1,211,885,691 training samples and 134,653,965 test samples, with all algorithms sharing input features.Source: §5.1.1 · p. 6

5.1 Offline Evaluation

• D2Q splits data by duration and predicts group-wise watch-time quantiles; Res-D2Q adds duration information to its network architecture.The offline comparison varies the number of duration groups across 1, 10, 20, 30, 50, and 100.Source: §5.1.1 · pp. 6–7, Figure 5 · p. 6

5.1.4 Result-I: Overall performance.

• D2Q and Res-D2Q with 30 duration groups peak across XAUC, XGAUC, and MAE; duration-split D2Q outperforms LR at every tested group count.With no splitting, D2Q has performance comparable to LR.Source: Table 2 · p. 7, §5.1.4 · p. 7

5.1.3 Metrics.

• XAUC and XGAUC measure ranking agreement, which the paper says is often valued more in practice than MAE because recommendations rank predicted values.XGAUC computes per-user XAUC and averages it with weights proportional to user sample size.Source: §5.1.3 · p. 7

5.1.5 Result-II: Effect of number of duration groups.

• XGAUC improves as duration groups increase, then declines when groups become too numerous; Res-D2Q is superior to D2Q after splitting.The decline is attributed to reduced groupwise sample size and accumulated estimation error.Source: Figure 6 · p. 7, §5.1.5 · pp. 7–8

5.2.3 Metrics.

• Res-D2Q improved WatchTime by +0.746% over VR in a statistically significant 24-hour Kuaishou App experiment.The reported average watch-time improvement from production algorithms is around 0.1%; differences in interaction constraint metrics between Res-D2Q and VR were not statistically significant.Source: Table 3 · p. 8, §5.2.4 · p. 8

5.2.2 Experimental setup.

• For live evaluation, users were randomly bucketed and 5% experienced each tested recommender; like, follow, share, and comment counts complemented WatchTime.The live metric was total time spent watching videos by users in the bucket.Source: §5.2.2 · p. 8, §5.2.3 · p. 8

• Duration can spuriously favor longer videos despite weak user-interest match, and recommendation feedback loops can amplify this bias and harm personalization.Source: §6 · p. 8
• D2Q uses a causal-graph formulation to remove duration's bias on video exposure while preserving its natural effect on watch-time prediction.Source: §6 · p. 9
• Offline evaluation and live experiments on Kuaishou show more accurate watch-time estimates and improved real-time video consumption; performance first rises, then declines as groupwise sample size falls.Source: §6 · p. 9

• YouTube's method defines positive videos as clicked, impressed videos and uses Odds based on positive-sample count, total-sample count, and positive-video watch times.Source: §A.1 · p. 10
• Odds is connected to expected watch time E(T) using the recommender system's click-through rate p_ctr.Source: §A.1 · p. 10
• The section approximates 1/(1−p_ctr) with a Taylor expansion because p_ctr is around 0.01, then combines the equations to calculate expected watch time.Source: §A.1 · p. 10

A.2 WLR: Adapting YouTube Method to Our Scenario

• The top-down, full-screen setting has no click action, so click-through rate is not well-defined for this adaptation.Instead, samples are labeled by whether watch time surpasses the empirical q60 quantile.Source: §A.2 · p. 10
• Positive samples are weighted by watch time, while negative samples receive unit weight.Source: §A.2 · p. 10
• The average probability of exceeding q60 is 0.4, so it cannot be approximated as click-through rate; expected watch time is obtained using Equation 11.The probability is learned with classical log-loss for binary classification.Source: §A.2 · p. 10

B NETWORK ARCHITECTURE IN EXPERIMENTS

• All experimental methods share the same network structure except Res-D2Q, which adds an MLP for duration adjustment.Source: B NETWORK ARCHITECTURE IN EXPERIMENTS · p. 10
• The network architecture is divided into input, encoding, and output layers.Source: B NETWORK ARCHITECTURE IN EXPERIMENTS · p. 10
• The duration-adjustment MLP is specific to Res-D2Q rather than shared across all experimental methods.Source: B NETWORK ARCHITECTURE IN EXPERIMENTS · p. 10

B.1 Input Layer

• The input layer embeds dense, ID, and raw video-duration features, producing 32-, 512-, and 32-dimensional embeddings, respectively.The mini-batch size is 512.Source: §B.1 · p. 10

B.2 Encoding Layer and Output Layer

• The concatenated embeddings are projected to E, then a three-layer MLP with sizes 256, 128, and 64 produces hidden matrix H.Source: §B.2 · p. 10
• The output layer uses H directly for VR and a sigmoid activation for other methods; Res-D2Q also concatenates the duration-tower output with H.Source: §B.2 · p. 10