Source-linked AI summary
Personalized Task Dependency Graphs for Mitigating Signal Erosion in Multi-Task Recommendation
Fuyuan Liu, Tiandeng Wu, Yaqun Fang, Wei Zhou, Zehao Zhou, Wenping Chen, Qishun Mei, Jiaxin Zhou, Heng Chang, Yi Cao, Jiandong Ding
TL;DR
Industrial multi-task recommendation must optimize heterogeneous conversion objectives, but rigid funnels impose uniform dependencies and can erode signals for sparse, deep tasks. PTDG learns low-rank, instance-specific task graphs with hard causal GCN propagation and sparsity-aware masking. It improves sparse-task performance while preserving dense-task accuracy, with online gains in CVR and eCPM over MMoE.
Problem
Existing MTL methods often enforce uniform dependency strengths in static conversion funnels, overlooking instance-level task variation and signal erosion on sparse deep objectives.
Method
PTDG learns low-rank instance-specific dependency graphs under hard causal masking, uses GCN message passing for adaptive shortcuts, and applies APM to decouple parameters by task sparsity.
Results
PTDG improves sparse, high-value tasks by up to 1.45% while maintaining dense-task accuracy; online testing reports +1.2% CVR and +1.9% eCPM versus MMoE.
Takeaways & Limitations
Personalized topology and structural parameter decoupling provide a favorable trade-off for mitigating signal attenuation in deep conversion objectives without compromising dense interactions.
Abstract
from arXiv · showhide
Optimizing multiple conversion objectives is a core challenge in industrial recommendation, often limited by signal erosion in rigid architectures. Existing Multi-Task Learning (MTL) methods typically enforce uniform dependency strengths across a static conversion funnel, overlooking how task correlations naturally vary based on item characteristics. Hierarchical message passing along these fixed chains leads to cumulative signal attenuation, which degrades performance on sparse, deep-funnel objectives. To address this, we propose the Personalized Task Dependency Graphs (PTDG). While respecting necessary physical causal constraints (e.g., Click -> Pay), PTDG dynamically "rewires" the intensity of dependency pathways for each item via low-rank approximation to ensure structural robustness. We implement a GCN-based propagation with hard causal masking to establish adaptive information shortcuts. Additionally, we introduce an Adaptive Progressive Masking (APM) strategy that decouples shared parameters according to task sparsity, helping to stabilize optimization. Experiments on KuaiRand1K and an industrial dataset show that PTDG significantly improves AUC on sparse conversion tasks by up to 1.45%, while maintaining comparable performance on dense objectives. Online A/B testing shows PTDG improves Conversion Rate (CVR) by 1.2% and effective Cost Per Mille (eCPM) by 1.9% relative to the baseline.
1 Introduction
Industrial recommendation must predict heterogeneous behaviors across a conversion funnel, but rigid task dependencies impose uniform relationships despite instance-specific variation. PTDG addresses this with adaptive topology, causally constrained propagation, and sparsity-aware parameter decoupling.
- Motivation: MTL systems share knowledge across shallow interactions and deep conversion objectives, but commonly assume universal dependency strengths throughout the funnel.This rigid inductive bias limits how task relationships can vary across traffic.
- Motivation: Dependency intensities vary by item: heavy games may follow Click→Download→Pay, whereas lightweight tools may allow Click→Pay.The physical causal order remains necessary, but the relevance of intermediate steps differs across instances.
- Proposed approach: PTDG uses low-rank graph generation to construct instance-specific dependency intensities while respecting causal constraints.This dynamic topology models heterogeneous task correlations rather than enforcing a monolithic funnel.
- Proposed approach: A GCN-based message-passing module with hard causal masking establishes adaptive shortcuts to mitigate signal erosion.The shortcuts are selected within the required causal structure.
- Proposed approach: Adaptive Progressive Masking decouples parameters according to label density, addressing gradient conflicts while reducing computational complexity.The strategy complements topology adaptation by making parameter sharing sensitive to task sparsity.
2 Related Work
Prior recommendation MTL research addresses representation sharing, task weighting, gradient conflicts, and conversion bias, but largely treats task relationships as numerical or globally shared. Dependency-focused methods add sequential structure yet remain limited by fixed hierarchies or costly structural search.
- MTL optimization: MMoE, PLE, and AITM disentangle shared and task-specific representations through gating networks, while PaDiRec generates task-conditioned parameters.These approaches primarily target representation sharing or dynamic parameterization.
- MTL optimization: GradNorm, PMTRec, and MoCoGrad address gradient conflicts through re-weighting or momentum calibration rather than structural task adaptation.The related methods operate mainly on the numerical optimization landscape.
- Research gap: Existing methods predominantly use parallel objectives or globally shared dependency patterns, neglecting structural heterogeneity across industrial applications.This gap motivates learning task relationships that vary with the application or instance.
- Dependency modeling: ESMM and ESCM2 model conversion over the full user space, while AITM transfers information along a fixed conversion chain.These methods exploit sequential behavior structure but retain predetermined dependency organization.
- Dependency modeling: HTLNet uses fixed hierarchical level assignments, limiting instance-level variation, while CSRL incurs O(T^2) or higher search costs.These constraints hinder flexible structural modeling or industrial-latency deployment.
3 Method
PTDG builds item-specific task dependency graphs under causal constraints, then combines causally masked message passing with sparsity-aware parameter masking. Low-rank graph construction and adaptive optimization are designed to preserve task signals while controlling overfitting, gradient conflicts, and serving cost.
- Personalized Topology Learning: PTDG constructs item-specific dependency graphs through low-rank factorization rather than enforcing static dependency strengths.The formulation uses item embeddings to derive latent factors for an adaptive adjacency structure.
- Personalized Topology Learning: User and scenario features provide a personalization bias that modulates item-specific latent factors across tasks.The bias is generated by an MLP and broadcast across the task dimension.
- Causally Optimized Message Passing: A hard causal mask enforces behavior order constraints before GCN propagation, preventing semantic backflow while supporting adaptive task connections.The propagation operates on task features and uses the masked dependency graph for message passing.
- Causally Optimized Message Passing: Adaptive shortcuts bypass redundant intermediate steps in fixed hierarchies, addressing signal erosion while retaining causal constraints.The extracted task node features are used for final conversion-probability prediction.
- Optimization and Efficiency: PTDG combines weighted per-task binary cross-entropy with EMA-normalized task-loss weights to stabilize training.Its low-rank formulation keeps adjacency calculation and GCN message passing linear in the number of tasks, while adding 8 ms end-to-end latency over MMoE.
- Adaptive Progressive Masking: APM learns task-specific masks over shared parameters and adjusts mask rates according to task label positive rates.The strategy is intended to decouple parameter sharing according to task sparsity and stabilize optimization.
4 Experiments
PTDG is evaluated on two datasets against industrial and state-of-the-art multi-task baselines, with results emphasizing sparse, deep-conversion tasks. Ablations show that personalized dependencies, GCN message passing, and adaptive masking each contribute to performance.
- Experimental Setup: PTDG is evaluated on KuaiRand1K with 6 tasks and an industrial dataset with 7 tasks.
- Performance Comparison: PTDG achieves the highest average AUC on both datasets and surpasses MIT by 1.05% on the key deep-conversion Task 6.It also beats MoCoGrad by 1.59% on average.
- Performance on Deep Tasks: 1.45%: PTDG surpasses PMTRec on sparse Industrial Task 3, while exceeding MMoE by 1.77% on the KuaiRand1K follow task.On dense, shallow tasks such as click, performance is comparable but not significantly better.
- Ablation Study: 1.21%: removing the instance-level dependency graph decreases Industrial Dataset Avg AUC.This supports the contribution of personalized task dependencies to heterogeneous multi-task relationship modeling.
- Ablation Study: 0.93%: replacing GCN message passing with simple concatenation lowers AUC, while disabling APM reduces performance by 0.59%.The ablations support GCN-based signal preservation and adaptive masking for mitigating gradient conflicts.
- Sensitivity Analysis: q = 2 gives the best performance in the low-rank dimension study, while APM performs best with r_min = 0.2 and r_max = 0.7.Larger q values degrade performance, and the masking settings correspond to high- and low-positive-sample-rate tasks.
- Online A/B Testing: +1.2% CVR and +1.9% eCPM: PTDG significantly improves both metrics over the deployed MMoE baseline in a two-week online A/B test.The test used 10% base traffic and 10% experimental traffic on a platform serving over 100 million daily active users.
5 Conclusion
The paper presents PTDG as a framework for addressing signal erosion in industrial multi-task recommendation through personalized dependency modeling and structural parameter decoupling. It improves sparse, high-value tasks while maintaining accuracy on dense interactions, with future work targeting larger task sets and cross-domain settings.
- PTDG combines low-rank dynamic graph learning, causal message passing, and structural parameter decoupling to address signal erosion in industrial multi-task recommendation.
- Up to 1.45%: PTDG improves sparse, high-value tasks while maintaining accuracy on dense interactions.
- Future work will explore scaling PTDG to larger task sets and extending it to cross-domain settings.
GenAI Usage Disclosure
Generative AI tools assisted with language editing and code debugging, while all experimental results were produced by author-written scripts and verified against saved artifacts.
- Generative AI tools assisted with language editing and code debugging.
- All experimental results were produced by author-written scripts and verified against saved artifacts.