Watch-DMLT: Causal Labeling for Recommenders

• Watch-DMLT is a causal labeling framework that transforms raw watch time into quantile-based, multi-semantics indicators to robustly capture user engagement.
• It employs causal adjustment by grouping on video duration to debias the inherent duration effects and correct for confounding variables.
• Empirical results show improved performance over traditional regression methods, enhancing AUC, GAUC, MAE, MAPE, RMSE, and boosting live user engagement.

The Watch-time Debiased Multiple-semantics-extracting Labeling (DML) framework is a causal labeling approach developed for short-video recommender systems to leverage watch-time feedback in a robust, bias-mitigated, and semantically rich manner. Unlike classical regression-based schemes which operate directly on raw watch time, DML converts watch-time data into multiple discrete indicators that encapsulate different types of user engagement and deploys causal adjustment strategies to explicitly correct for confounders like video duration, allowing for superior alignment with recommendation objectives and improved downstream performance (Zhang et al., 2023).

1. Motivation and Problem Setting

In short-video recommendation, user watch time on item $i$ by user $u$, denoted $T_{u,i}$, is routinely used as the training objective. Traditional regression methods formalize label construction as minimizing squared error between predicted and true watch times: $$\min_{\theta} \sum_{(u,i)} (\hat T_{u,i}(\theta) - T_{u,i})^2.$$ However, this formulation is fundamentally flawed in the short-video context due to two pronounced issues:

• Heavy-tail and outliers: The empirical distribution of watch time across users and videos exhibits an extreme heavy-tail, with a vast majority of views lasting only seconds and outliers stretching to minutes. Squared-error objectives overemphasize these rare long watches, distorting learning.
• Duration bias: Raw watch time $T_{u,i}$ is naturally confounded by the video’s intrinsic duration $V_d$; longer content tends to elicit larger values regardless of user interest, introducing bias and attenuating the recommender’s ability to gauge true user preference (Zhang et al., 2023).

2. Quantile-based Multi-semantics Labeling

Addressing the drawbacks of scalar regression, DML recasts the watch-time learning problem by discretizing the aggregate distribution of $T_{u,i}$ into quantile-based buckets. Letting $F(t)$ denote the CDF of $T$ over the training set, DML defines quantile boundaries: $$Q_{k} = F^{-1}\left(\frac{k}{N}\right), \quad k=0,\ldots,N,$$ with buckets $[Q_{k-1}, Q_{k})$. For each tuple $(u, i)$, DML computes indicator variables: $$y_{u,i}^{(k)} = \begin{cases} 1, & Q_{k-1} \leq T_{u,i} < Q_k \ 0, & \text{otherwise} \end{cases}$$ for $k=1,\ldots,N$. These multi-level binary indicators encode the global rank of $T_{u,i}$, imparting robustness to rare outlier events and alleviating sensitivity to heavy-tail effects.

Beyond fine-grained buckets, two additional binary thresholds extract canonical semantics: $$y_{u,i}^{\mathrm{EV}} = \mathbf{1}\{T_{u,i} \geq t_{50}\}, \quad y_{u,i}^{\mathrm{LV}} = \mathbf{1}\{T_{u,i} \geq t_{75}\},$$ where $t_p = F^{-1}(p)$, yielding “effective views” (above-median engagement) and “long views” (top quartile engagement) (Zhang et al., 2023).

3. Causal Adjustment and Label Debiasing

Even ranking-based labeling inherits duration bias, as longer videos persistently shift the watch-time percentiles. DML applies a back-door adjustment, rooted in structural causal modeling, to sever spurious associations induced by such confounders. Specifically, it buckets the dataset by discrete video-duration bins and recomputes quantile thresholds within each bin. For group $v_d$, the quantile indicator becomes: $$y_{u,i}^{(k),d} = \begin{cases} 1, & Q_{k-1}^{(v_d)} \leq T_{u,i} < Q_k^{(v_d)} \ 0, & \text{otherwise} \end{cases}$$ Analogous adjustments yield duration-debiased EV/LV semantics: $$y_{u,i}^{\mathrm{EV},d} = \mathbf{1}\{T_{u,i} \geq t_{50}(\mathcal{D}^{v_d})\}, \quad y_{u,i}^{\mathrm{LV},d} = \mathbf{1}\{T_{u,i} \geq t_{75}(\mathcal{D}^{v_d})\}$$ This approach is generalized by also grouping on other features (full video or user attributes), generating a diverse set of debiased labels: $$\{y_{u,i}^{(k),d}\}_{k=1}^N, \quad y_{u,i}^{\mathrm{EV},d},\,y_{u,i}^{\mathrm{EV},v},\,y_{u,i}^{\mathrm{EV},u},\, y_{u,i}^{\mathrm{LV},d},\,y_{u,i}^{\mathrm{LV},v},\,y_{u,i}^{\mathrm{LV},u}$$ These grouped labels block back-door confounding, ensuring model targets reflect user–video matching rather than artifactually high engagement due to content length or popularity.

4. Integration with Recommendation Backbones

The DML label set is directly integrated into a Multi-gate Mixture-of-Experts (MMoE) recommendation architecture. Each labeling task (WPR buckets, EV, LV) is assigned a separate expert gate and output tower. The quantile-bucket tasks are optimized using multi-class or multi-label cross-entropy, while the binary EV/LV tasks employ binary cross-entropy. The composite objective function aggregates losses: $\mathcal{L} = \sum_t \lambda_t \mathcal{L}_t$ where each $\lambda_t$ weights the semantic significance of its corresponding loss. Importantly, no modification to the model architecture or regularization schema is introduced; all debiasing and semantic enrichment is localized to the label construction phase (Zhang et al., 2023).

5. Empirical Evaluation

The DML approach was empirically studied on a large-scale Kuaishou dataset (46 billion daily samples). DML was benchmarked against:

Baseline	Method Description
TR	Direct regression on $T_{u,i}$
WLR	Weighted logistic regression (YouTube-style)
OR	Ordinal regression
D2Q	Deconfounded quantile regression

In head-to-head matrix comparisons against D2Q, DML achieved:

• AUC: 0.6763 vs. 0.6732 (+0.0031)
• GAUC: 0.6617 vs. 0.6581 (+0.0036)
• MAE: 21.7657 vs. 22.6728 (–0.9071 s)
• MAPE: 2.6039 vs. 2.7342 (–4.8%)
• RMSE: 42.7735 vs. 45.5293 (–2.756)

6. Online Impact and Live Deployment

A/B tests spanning two recommendation channels (“Featured-Video” and “Slide-Tab”) over five days revealed that DML surpasses WLR, with increments as follows:

• Using only WPR label: +0.694% watch time, +0.207% app usage (Featured-Video)
• WPR with duration debiasing ($y_{wpr}^d$): +1.048% watch time, +0.332% app usage
• Adding all EV debiased heads ($\{y_{ev}^d, y_{ev}^v, y_{ev}^u\}$): +2.230% watch time, +0.773% app usage
• Adding LV heads instead: +1.549% watch time, +0.566% app usage

These findings validate that quantile-based, multi-semantics labels paired with causal adjustment yield substantial uplift in live short-video recommendation metrics, both in offline ranking accuracy and online user engagement (Zhang et al., 2023).

7. Significance and Implications

DML represents a principled shift away from legacy regression label paradigms by enshrining engagement via percentile semantics and integrating causal debiasing directly into label construction. The approach is agnostic to backbone architecture, modular in terms of which biases are targeted for adjustment, and extensible to additional multimodal or sequential semantics. A plausible implication is that DML-style methodologies, because of their grounding in both distributional and causal theory, may generalize to other domains where feedback is noisy, confounded, or heavy-tailed.