[论文解读] Feedback Control for Multi-Objective Graph Self-Supervision
ControlG 将多目标图 SSL 重构为闭环时间分配，利用感知、Pareto 感知规划和 deficit 基于 PID 控制来调度单目标块并提高迁移性能。
Can multi-task self-supervised learning on graphs be coordinated without the usual tug-of-war between objectives? Graph self-supervised learning (SSL) offers a growing toolbox of pretext objectives: mutual information, reconstruction, contrastive learning; yet combining them reliably remains a challenge due to objective interference and training instability. Most multi-pretext pipelines use per-update mixing, forcing every parameter update to be a compromise, leading to three failure modes: Disagreement (conflict-induced negative transfer), Drift (nonstationary objective utility), and Drought (hidden starvation of underserved objectives). We argue that coordination is fundamentally a temporal allocation problem: deciding when each objective receives optimization budget, not merely how to weigh them. We introduce ControlG, a control-theoretic framework that recasts multi-objective graph SSL as feedback-controlled temporal allocation by estimating per-objective difficulty and pairwise antagonism, planning target budgets via a Pareto-aware log-hypervolume planner, and scheduling with a Proportional-Integral-Derivative (PID) controller. Across 9 datasets, ControlG consistently outperforms state-of-the-art baselines, while producing an auditable schedule that reveals which objectives drove learning.
研究动机与目标
- 由于每次更新混合导致的不稳定性和干扰在多目标图 SSL 中的存在及其原因
- 提出一种时间分配方法，对单一目标进行调度而非逐更新混合
- 开发一个感知-规划-控制环，包括状态估计、Pareto 感知规划和 PID 执行
- 提供可审计的调度，揭示哪些目标在何时驱动学习
提出的方法
- 将多任务图 SSL 解释为一个包含感知、规划、控制三环的闭环调度问题
- Sense：通过表示梯度变化来估计每个目标的谱需求，并通过基于 MGDA 的梯度几何来估计干扰
- Plan：使用对数超体积敏感性并结合综合难度状态的 tempered 值，计算 Pareto 感知分配
- Control：实现一个 deficit 跟踪的 PID 控制器，以实现离散的单任务块和可审计的调度
- 调度决策映射到基于 softmax 的随机选择器，以进行块级任务更新
实验结果
研究问题
- RQ1是否可以通过时序分配在不牺牲性能的前提下避免在图 SSL 中按更新混合目标？
- RQ2如何量化谱需求和梯度干扰以驱动自适应调度？
- RQ3将 Pareto 感知的对数超体积规划器与 PID 控制结合，是否在多样化图基准上带来鲁棒改进？
- RQ4生成的训练调度是否可审计且可解释，哪些目标驱动学习？
主要发现
| Method | Cora | CiteSeer | Chameleon | Squirrel | Actor | PubMed | Wiki-CS | Co-CS | Arxiv | Avg Rank |
|---|---|---|---|---|---|---|---|---|---|---|
| BGRL | 77.14 ± 2.2 | 35.28 ± 2.1 | 64.87 ± 1.3 | 46.11 ± 1.2 | 31.22 ± 0.5 | 79.94 ± 1.2 | 80.36 ± 0.3 | 93.05 ± 0.4 | 71.24 ± 0.8 | 9.2 |
| DGI | 75.64 ± 1.6 | 63.84 ± 1.6 | 66.27 ± 0.9 | 48.17 ± 1.0 | 30.46 ± 0.7 | 81.28 ± 1.1 | 77.95 ± 0.9 | 93.17 ± 0.4 | 70.86 ± 0.6 | 9.6 |
| GRACE | 68.88 ± 2.1 | 58.96 ± 2.3 | 66.23 ± 0.4 | 54.06 ± 1.2 | 28.80 ± 0.5 | 80.54 ± 0.8 | 79.78 ± 0.3 | 93.62 ± 0.2 | – | 9.1 |
| MVGRL | 79.32 ± 0.9 | 59.68 ± 3.9 | 55.00 ± 1.1 | – | 31.17 ± 0.7 | 78.60 ± 0.9 | – | 90.51 ± 4.5 | – | 12.4 |
| p_link | 76.72 ± 1.0 | 52.76 ± 1.9 | 57.41 ± 3.6 | 36.56 ± 1.3 | 29.00 ± 0.6 | 81.28 ± 0.5 | 79.22 ± 0.3 | 91.92 ± 0.2 | 70.42 ± 0.5 | 13.1 |
| p_recon | 78.76 ± 0.5 | 64.86 ± 1.5 | 69.39 ± 1.3 | 52.33 ± 0.8 | 27.75 ± 0.4 | 76.28 ± 0.9 | 79.88 ± 0.4 | 94.85 ± 0.3 | 71.08 ± 0.4 | 6.8 |
| p_minsg | 76.44 ± 1.1 | 61.74 ± 0.8 | 64.78 ± 1.6 | 47.61 ± 0.9 | 29.16 ± 0.7 | 80.76 ± 1.0 | 79.95 ± 0.2 | 93.44 ± 0.3 | 70.94 ± 0.6 | 10.2 |
| p_decor | 43.86 ± 9.0 | 34.22 ± 5.7 | 52.59 ± 0.8 | 40.88 ± 1.1 | 25.92 ± 0.9 | 75.66 ± 1.3 | 72.60 ± 4.7 | 93.39 ± 0.3 | 68.52 ± 1.2 | 15.6 |
| p_par | 67.22 ± 1.3 | 52.74 ± 1.7 | 57.06 ± 3.4 | 40.96 ± 0.6 | 28.26 ± 1.3 | 74.20 ± 0.8 | 73.56 ± 0.2 | 92.31 ± 0.2 | 69.86 ± 0.7 | 14.7 |
| AutoSSL | 80.30 ± 1.5 | 64.72 ± 1.0 | 70.09 ± 2.0 | 49.99 ± 3.0 | 29.76 ± 0.7 | 80.80 ± 0.7 | 79.82 ± 0.3 | 93.61 ± 0.2 | – | 6.4 |
| WAS | 74.12 ± 3.3 | 57.66 ± 3.5 | 57.59 ± 6.3 | 43.52 ± 2.2 | 29.68 ± 0.5 | 78.62 ± 2.1 | 77.40 ± 1.6 | 92.62 ± 1.3 | 70.68 ± 0.9 | 12.9 |
| Uniform | 77.14 ± 0.9 | 58.68 ± 1.2 | 66.23 ± 1.4 | 49.93 ± 1.0 | 29.26 ± 0.5 | 82.38 ± 0.5 | 80.48 ± 0.2 | 93.97 ± 0.2 | 71.32 ± 0.4 | 7.3 |
| ParetoGNN | 78.26 ± 0.9 | 58.32 ± 1.1 | 65.35 ± 2.2 | 45.00 ± 1.9 | 28.95 ± 1.0 | 67.72 ± 3.3 | 79.94 ± 0.1 | 93.96 ± 0.2 | 70.56 ± 0.8 | 10.8 |
| PCGrad | 79.44 ± 0.8 | 59.66 ± 1.4 | 66.84 ± 1.4 | 49.11 ± 1.3 | 29.63 ± 0.5 | 83.12 ± 0.5 | 80.40 ± 0.2 | 94.07 ± 0.2 | 71.48 ± 0.3 | 5.8 |
| CAGrad | 80.40 ± 0.4 | 62.66 ± 0.3 | 66.75 ± 0.6 | 50.34 ± 1.4 | 29.67 ± 0.5 | 82.84 ± 0.9 | 79.94 ± 0.3 | 94.31 ± 0.1 | 71.62 ± 0.4 | 5.2 |
| Random | 79.06 ± 0.3 | 58.10 ± 1.3 | 63.07 ± 1.8 | 46.22 ± 0.7 | 29.57 ± 1.0 | 80.56 ± 0.8 | 80.10 ± 0.2 | 93.39 ± 0.2 | 71.18 ± 0.5 | 9.4 |
| Round-Robin | 78.52 ± 2.0 | 59.90 ± 1.1 | 64.61 ± 1.1 | 48.17 ± 1.1 | 30.07 ± 0.5 | 80.78 ± 0.5 | 80.22 ± 0.1 | 94.11 ± 0.2 | 71.28 ± 0.4 | 7.4 |
| ControlG | 81.92 ± 0.9 | 66.48 ± 1.1 | 69.54 ± 1.0 | 53.18 ± 0.9 | 31.20 ± 0.6 | 84.24 ± 0.7 | 80.45 ± 0.3 | 96.14 ± 0.2 | 72.86 ± 0.3 | 1.4 |
- ControlG 在九个数据集上的节点分类、链接预测和节点聚类任务上具有最佳平均排名。
- ControlG 在同向同质图（如 Cora、PubMed）上显示出强劲增益，在异质图上仍具有竞争力，优于基于加权的基线。
- ControlG 在各数据集的节点分类平均排名为 1.4，链接预测为 1.9，节点聚类为 1.8。
- 在 ogbn-arxiv 上，ControlG 达到 72.86% 的准确率，比 CAGrad 提升 1.2%。
- 消融实验显示谱需求和规划器组件发挥了贡献作用；移除它们会降低性能。
- ControlG 产生可审计的调度，揭示了哪些目标被优先考虑以及何时被考虑。
更好的研究，从现在开始
从论文设计到论文写作，大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成，并经人工编辑审核。