ViSeDOPS in Short-Video Recommendations

• ViSeDOPS is a framework that transforms raw watch-time into quantile-based, multi-semantic labels to mitigate bias in recommendation systems.
• It employs causal inference and stratified quantiling to adjust for confounders such as video duration and popularity.
• Integrated with multi-task deep learning models, ViSeDOPS improves ranking performance and stabilizes training on implicit feedback data.

ViSeDOPS, in the context of short-video recommendations, refers to frameworks and methodologies that leverage debiased, multi-semantic label extraction from dense watch-time feedback, using causal inference techniques and quantile-based stratification. Such approaches move beyond direct use of raw watch-time, addressing the central biases and variabilities inherent in user engagement metrics and enabling robust optimization of recommendation models while aligning with the core ranking objectives of content surfacing platforms.

1. Limitations of Direct Watch-time Regression

Raw watch-time $T_{u,i}$, where $u$ denotes user and $i$ denotes item, is the primary feedback signal in large-scale short-video recommendation scenarios, modeled as a sample from a joint distribution $P(T|u,i)$. The naive approach minimizes mean-squared error:

$$L_{\mathrm{MSE}} = \mathbb{E}_{u,i} [(T_{u,i} - f(u,i; \theta))^2]$$

for a parametric recommender $f(\cdot; \theta)$. However, this strategy is fundamentally limited by two factors:

• Heavy-tailed Distribution and Outlier Sensitivity: Watch-time exhibits a long-tailed distribution, with most videos watched for seconds but rare items accruing hours via repeat plays. MSE optimization becomes skewed toward outliers, resulting in unstable learning.
• Confounding and Bias Due to Video Attributes: $T_{u,i}$ is not solely reflective of latent user interest $M_{u,i}$; it is confounded by video-side attributes, most notably video duration $V_d$ and popularity. The relevant causal structure:

$$V_d \rightarrow T_{u,i} \leftarrow M_{u,i} \rightarrow (u,i)$$

induces direct and indirect pathways from $V_d$ to $T_{u,i}$, such that models trained directly on $T_{u,i}$ systematically overpreference longer videos regardless of user interest (Zhang et al., 2023).

2. Quantile-based Multi-Semantics Labeling

To mitigate the influence of outliers and align labels with ranking goals, watch-time is transformed into quantile-based discrete semantics:

• Let $F_T$ be the empirical CDF of $T$ in the training data.
• The support of $T$ is partitioned into $K$ intervals by quantiles:

$$Q_k = F_T^{-1}\left(\frac{k}{K}\right),\quad k=0,1,\ldots,K; \quad Q_0=0,\ Q_K=\infty$$

• The label vector $y_{u,i}^{(k)}$ is assigned as:

$$y_{u,i}^{(k)} = \begin{cases} 1, & Q_{k-1} \leq T_{u,i} < Q_k \ 0, & \text{otherwise} \end{cases}$$

for $k=1,\ldots,K$

Common label semantics include:

Label Type	Definition	Purpose
Watch-time Percentile Rank (WPR)	$y_{wpr} = \sum_{k'=1}^k 1/K$ for bin $k$	Soft percentile, $[0,1]$ scale
Effective View (EV)	$y_{ev} = \mathbb{I}(T_{u,i} \geq t_{50})$	Binary: watches above median
Long View (LV)	$y_{lv} = \mathbb{I}(T_{u,i} \geq t_{75})$	Binary: watches above $75^{th}$ percentile

This quantile-based transformation prioritizes relative ordering of watch times, stabilizes learning, and aligns loss functions with ranking performance metrics.

3. Causal Label Debiasing via Stratified Quantiling

Despite quantile transformation, label bias from confounders, especially $V_d$, persists. To neutralize this, causal adjustment is performed using the backdoor adjustment formula:

$$\mathbb{E}[T_{u,i} | do(M_{u,i})] = \sum_{v_d} P(v_d) \;\mathbb{E}[T_{u,i}\mid M_{u,i}, V_d = v_d]$$

Practical debiasing is achieved by stratifying the dataset by $v_d$ (video duration buckets), then performing quantile mapping within each stratum:

• For each $(u,i) \in D^{v_d}$ (duration bucket), labels are defined as:

$$y_{wpr}^d(u,i) = g_{wpr}\big((u,i,T_{u,i}); D^{v_d}\big)$$

where $g_{wpr}$ is quantile assignment in $D^{v_d}$.

• Likewise for binary labels:

$$y_{ev}^d(u,i) = \mathbb{I}(T_{u,i} \geq t_{50}(D^{v_d}))$$

$$y_{lv}^d(u,i) = \mathbb{I}(T_{u,i} \geq t_{75}(D^{v_d}))$$

Further stratification by user/item feature hashes can mitigate additional confounding (popularity, user activity). This produces uniform label distributions per bucket, eliminating the direct $V_d$ effect (Zhang et al., 2023).

4. Model Integration and Multi-Task Objective

The debiased, multi-semantic labels are incorporated in a multi-task deep recommender using the Multi-gate Mixture-of-Experts (MMoE) framework:

• Shared embedding and expert layers feed into separate “heads” for each label type:
  • $f_{wpr}(u,i;\theta)$, $f_{ev}^d(u,i;\theta)$, $f_{ev}^v(u,i;\theta)$, $f_{ev}^u(u,i;\theta)$
  • $f_{lv}^d(u,i;\theta)$, $f_{lv}^v(u,i;\theta)$, $f_{lv}^u(u,i;\theta)$

The global loss is a sum of weighted per-label losses: $$L(\theta) = L_{wpr} + \lambda_{ev} (L_{ev}^d + L_{ev}^v + L_{ev}^u) + \lambda_{lv} (L_{lv}^d + L_{lv}^v + L_{lv}^u)$$ Per-task loss formulations:

• $$L_{wpr} = \mathbb{E}[(f_{wpr}(u,i) - y_{wpr}^d(u,i))^2]$$
• $$L_{ev}^* = -\mathbb{E}\big[y_{ev}^*(u,i) \log\sigma(f_{ev}^*(u,i)) + (1-y_{ev}^*(u,i)) \log(1-\sigma(f_{ev}^*(u,i)))\big]$$
• Analogous for $L_{lv}^\ast$

The $\lambda$ weighting parameters are tuned to balance gradient magnitudes across tasks. No architectural or backward-compatibility changes are needed; all debiasing occurs at the label stage.

5. Empirical Comparison and Performance

The debiased, multi-semantics labeling approach was benchmarked against several strong baselines:

• TR: direct watch-time regression
• WLR: YouTube’s weighted-logistic baseline
• OR: ordinal regression on $T$
• D2Q: duration-adjusted quantile regression

Metrics comprised both ranking (AUC, GAUC) and regression error (MAE, MAPE, RMSE). On a 1-hour holdout: | Method | AUC | GAUC | MAE | MAPE | RMSE | |--------|-------|-------|-------|-------|-------| | D2Q | 0.6732| 0.6581|22.67 |2.73 |45.53 | | DML | 0.6763| 0.6617|21.76 |2.60 |42.77 |

Online A/B tests over 5 days on Kuaishou’s major surfacing channels demonstrated uplifts relative to WLR:

• Featured-Video Tab: Watch Time +2.23%, App Usage +0.77%
• Slide Tab: Watch Time +2.06%, App Usage +0.71%

DML’s empirical gains validate the efficacy of extracting multiple semantics, using relative quantile labels, and applying label-construction-time debiasing—all without changes to the deployed recommender backbone (Zhang et al., 2023).

6. Context and Implications

This line of work demonstrates that nuanced, causally debiased label engineering—coupled with multi-task learning—can rectify core deficiencies in dense implicit-feedback recommendation where only watch-time is available. Operating at the label level enables robust, architecture-agnostic bias correction that is compatible with production-scale recommender systems. The stratified quantiling mechanism is extensible to de-confounding against additional side information. A plausible implication is that these methodological principles may generalize to other domains with heavy-tailed, confounded feedback signals, beyond short-video and media recommendation.