[论文解读] Debiased Collaborative Filtering with Kernel-Based Causal Balancing
论文介绍了自适应基于核的平衡，用于学习 IPS 的倾向分数并进行 DR 去偏的协同过滤，旨在更好地满足因果平衡约束并降低评分预测中的偏差。
Debiased collaborative filtering aims to learn an unbiased prediction model by removing different biases in observational datasets. To solve this problem, one of the simple and effective methods is based on the propensity score, which adjusts the observational sample distribution to the target one by reweighting observed instances. Ideally, propensity scores should be learned with causal balancing constraints. However, existing methods usually ignore such constraints or implement them with unreasonable approximations, which may affect the accuracy of the learned propensity scores. To bridge this gap, in this paper, we first analyze the gaps between the causal balancing requirements and existing methods such as learning the propensity with cross-entropy loss or manually selecting functions to balance. Inspired by these gaps, we propose to approximate the balancing functions in reproducing kernel Hilbert space and demonstrate that, based on the universal property and representer theorem of kernel functions, the causal balancing constraints can be better satisfied. Meanwhile, we propose an algorithm that adaptively balances the kernel function and theoretically analyze the generalization error bound of our methods. We conduct extensive experiments to demonstrate the effectiveness of our methods, and to promote this research direction, we have released our project at https://github.com/haoxuanli-pku/ICLR24-Kernel-Balancing.
研究动机与目标
- 对观测到的协同过滤数据中的偏差进行动机阐述和解决。
- 通过在 RKHS 中强制因果平衡来弥合学习倾向得分的差距。
- 提出自适应和最坏情况的核平衡以满足平衡约束并降低泛化误差。
- 为所提出的基于核的方法提供理论保证和泛化界限。
- 在真实世界数据集上展示在 IPS 和 DR 估计量上的经验收益。
提出的方法
- 将去偏 CF 与 IPS 和 DR 损失函数结合起来，并将平衡约束与倾向学习联系起来。
- 使用高斯核或指数核在 RKHS 中近似平衡函数，以满足对所有 φ 的因果平衡。
- 开发最坏情况的核平衡（WKB），以最小化 RKHS 函数的潜在偏差。
- 引入自适应核平衡（AKB），通过数据驱动准则选择最具影响力的核函数。
- 提供一个交替训练框架，在该框架中预测、平衡权重和插补模型共同更新（如算法1所示）。
- 推导 RKHS 中 KBIPS 和 KBDR 的泛化误差界，并显示通过 WKB/AKB 控制偏差。
实验结果
研究问题
- RQ1在学习倾向时如何更好地满足因果平衡约束以实现去偏 CF？
- RQ2与有限函数方法相比，基于 RKHS 的核平衡是否能更有效地近似对所有 φ 的平衡？
- RQ3带有自适应或最坏情况平衡的核基倾向估计量（KBIPS/KBDR）是否能减少偏差并提升预测性能？
- RQ4在协同过滤中，基于核的去偏方法的理论泛化保证是什么？
- RQ5与现有的 IPS/DR 去偏基线在标准数据集上的表现相比，核基方法如何？
主要发现
| Method | Coat AUC | Coat NDCG@5 | Coat F1@5 | Music AUC | Music NDCG@5 | Music F1@5 | Product AUC | Product NDCG@20 | Product F1@20 |
|---|---|---|---|---|---|---|---|---|---|
| MF | 0.703 ±0.006 | 0.605 ±0.012 | 0.467 ±0.007 | 0.673 ±0.001 | 0.635 ±0.002 | 0.306 ±0.002 | 0.753 ±0.001 | 0.449 ±0.002 | 0.124 ±0.002 |
| + IPS | 0.717 ±0.007 | 0.617 ±0.009 | 0.473 ±0.008 | 0.678 ±0.001 | 0.638 ±0.002 | 0.318 ±0.002 | 0.755 ±0.004 | 0.452 ±0.010 | 0.131 ±0.004 |
| + SNIPS | 0.714 ±0.012 | 0.614 ±0.012 | 0.474 ±0.009 | 0.683 ±0.002 | 0.639 ±0.002 | 0.316 ±0.002 | 0.754 ±0.003 | 0.453 ±0.004 | 0.126 ±0.003 |
| + ASIPS | 0.719 ±0.009 | 0.618 ±0.012 | 0.476 ±0.009 | 0.679 ±0.003 | 0.640 ±0.003 | 0.319 ±0.003 | 0.757 ±0.005 | 0.474 ±0.007 | 0.130 ±0.005 |
| + IPS-V2 | 0.726 ±0.005 | 0.627 ±0.009 | 0.479 ±0.008 | 0.685 ±0.002 | 0.646 ±0.003 | 0.320 ±0.002 | 0.764 ±0.001 | 0.476 ±0.003 | 0.135 ±0.003 |
| + RKBIPS-Exp | 0.714 ±0.003 | 0.618 ±0.010 | 0.474 ±0.007 | 0.676 ±0.002 | 0.642 ±0.003 | 0.318 ±0.002 | 0.763 ±0.001 | 0.463 ±0.007 | 0.134 ±0.002 |
| + RKBIPS-Gau | 0.715 ±0.005 | 0.619 ±0.010 | 0.475 ±0.008 | 0.678 ±0.001 | 0.640 ±0.004 | 0.315 ±0.003 | 0.760 ±0.003 | 0.470 ±0.008 | 0.133 ±0.003 |
| + WKBIPS-Exp | 0.723 ±0.004 | 0.624 ±0.009 | 0.480 ±0.007 | 0.687 ±0.002 | 0.654 ±0.002 | 0.322 ±0.002 | 0.765 ±0.003 | 0.475 ±0.007 | 0.138 ±0.003 |
| + WKBIPS-Gau | 0.722 ±0.004 | 0.625 ±0.008 | 0.479 ±0.007 | 0.686 ±0.002 | 0.650 ±0.002 | 0.321 ±0.002 | 0.763 ±0.003 | 0.476 ±0.007 | 0.137 ±0.003 |
| + AKBIPS-Exp | 0.732* ±0.004 | 0.636* ±0.006 | 0.483 ±0.006 | 0.689* ±0.001 | 0.658* ±0.002 | 0.324* ±0.002 | 0.766* ±0.003 | 0.478 ±0.009 | 0.138* ±0.003 |
| + AKBIPS-Gau | 0.730* ±0.003 | 0.633 ±0.008 | 0.484 ±0.007 | 0.688* ±0.003 | 0.655* ±0.003 | 0.324* ±0.002 | 0.767* ±0.003 | 0.480 ±0.009 | 0.139* ±0.003 |
| + DR | 0.718 ±0.008 | 0.623 ±0.009 | 0.474 ±0.007 | 0.684 ±0.002 | 0.658 ±0.003 | 0.326 ±0.002 | 0.755 ±0.008 | 0.462 ±0.010 | 0.135 ±0.005 |
| + DR-JL | 0.723 ±0.005 | 0.629 ±0.007 | 0.479 ±0.005 | 0.685 ±0.002 | 0.653 ±0.002 | 0.324 ±0.002 | 0.766 ±0.002 | 0.467 ±0.005 | 0.136 ±0.003 |
| + MRDR-JL | 0.727 ±0.005 | 0.627 ±0.008 | 0.480 ±0.008 | 0.684 ±0.002 | 0.652 ±0.003 | 0.325 ±0.002 | 0.768 ±0.005 | 0.473 ±0.007 | 0.139 ±0.004 |
| + DR-BIAS | 0.726 ±0.004 | 0.629 ±0.009 | 0.482 ±0.007 | 0.685 ±0.002 | 0.653 ±0.002 | 0.325 ±0.003 | 0.768 ±0.003 | 0.477 ±0.006 | 0.137 ±0.004 |
| + DR-MSE | 0.727 ±0.007 | 0.631 ±0.008 | 0.484 ±0.007 | 0.687 ±0.002 | 0.657 ±0.003 | 0.327 ±0.003 | 0.770 ±0.003 | 0.480 ±0.006 | 0.140 ±0.003 |
| + MR | 0.724 ±0.004 | 0.636 ±0.006 | 0.481 ±0.006 | 0.691 ±0.002 | 0.647 ±0.002 | 0.316 ±0.003 | 0.776 ±0.005 | 0.483 ±0.006 | 0.142 ±0.003 |
| + TDR | 0.714 ±0.006 | 0.634 ±0.011 | 0.483 ±0.008 | 0.688 ±0.003 | 0.662 ±0.002 | 0.329 ±0.002 | 0.772 ±0.003 | 0.486 ±0.005 | 0.140 ±0.003 |
| + TDR-JL | 0.731 ±0.005 | 0.639 ±0.007 | 0.484 ±0.007 | 0.689 ±0.002 | 0.656 ±0.004 | 0.327 ±0.003 | 0.772 ±0.003 | 0.489 ±0.005 | 0.142 ±0.003 |
| + SDR | 0.735 ±0.005 | 0.640 ±0.007 | 0.484 ±0.006 | 0.688 ±0.002 | 0.661 ±0.003 | 0.329 ±0.002 | 0.773 ±0.001 | 0.491 ±0.003 | 0.143 ±0.003 |
| + DR-V2 | 0.734 ±0.007 | 0.639 ±0.009 | 0.487 ±0.006 | 0.690 ±0.002 | 0.660 ±0.005 | 0.328 ±0.002 | 0.773 ±0.003 | 0.488 ±0.006 | 0.142 ±0.004 |
| + RKBDR-Exp | 0.730 ±0.003 | 0.631 ±0.005 | 0.482 ±0.006 | 0.682 ±0.002 | 0.648 ±0.003 | 0.323 ±0.002 | 0.765 ±0.004 | 0.460 ±0.006 | 0.138 ±0.003 |
| + RKBDR-Gau | 0.726 ±0.005 | 0.630 ±0.008 | 0.480 ±0.008 | 0.683 ±0.002 | 0.652 ±0.003 | 0.325 ±0.002 | 0.766 ±0.003 | 0.469 ±0.007 | 0.134 ±0.004 |
| + WKBDR-Exp | 0.735 ±0.005 | 0.637 ±0.009 | 0.483 ±0.006 | 0.685 ±0.003 | 0.654 ±0.003 | 0.325 ±0.002 | 0.773 ±0.003 | 0.489 ±0.008 | 0.142 ±0.003 |
| + WKBDR-Gau | 0.732 ±0.003 | 0.638 ±0.007 | 0.483 ±0.005 | 0.687 ±0.001 | 0.655 ±0.002 | 0.327 ±0.002 | 0.773 ±0.002 | 0.490 ±0.005 | 0.142 ±0.004 |
| + AKBDR-Exp | 0.745* ±0.004 | 0.645 ±0.008 | 0.493* ±0.007 | 0.692 ±0.002 | 0.661 ±0.002 | 0.328 ±0.002 | 0.782* ±0.003 | 0.498* ±0.008 | 0.147* ±0.003 |
| + AKBDR-Gau | 0.746* ±0.004 | 0.646* ±0.008 | 0.492 ±0.007 | 0.694* ±0.002 | 0.664* ±0.002 | 0.332* ±0.002 | 0.782* ±0.005 | 0.503* ±0.006 | 0.148* ±0.004 |
- 自适应核平衡（AKB）在不同数据集和核函数下始终优于基线平衡方法的去偏性能。
- 使用高斯或指数核的 AKB 在 Coat、Music、Product 数据集上实现了最佳的综合 AUC、NDCG@5 和 F1。
- 核平衡方法（KBIPS/KBDR）优于随机核平衡（RKB）和矩量平衡（MB）基线，AKB 提供最强的增益。
- 理论泛化界定表明在 RKHS 假设下使用 KBIPS/KBDR 的误差减少。
- 经验结果显示 AKBDR 结合核平衡在许多 IPS/DR 基线上实现统计显著提升（如 AKBDR-Exp 和 AKBDR-Gau）。
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。