[论文解读] Learning Disentangled Representations for Recommendation

Jianxin Ma, Chang Zhou+3 more|arXiv (Cornell University)|Oct 31, 2019

Generative Adversarial Networks and Image Synthesis被引用 100

一句话总结

引入 MacridVAE，从用户行为中学习宏观和微观解耦表示，提升推荐性能并获得可解释、可控的表示。

ABSTRACT

User behavior data in recommender systems are driven by the complex interactions of many latent factors behind the users' decision making processes. The factors are highly entangled, and may range from high-level ones that govern user intentions, to low-level ones that characterize a user's preference when executing an intention. Learning representations that uncover and disentangle these latent factors can bring enhanced robustness, interpretability, and controllability. However, learning such disentangled representations from user behavior is challenging, and remains largely neglected by the existing literature. In this paper, we present the MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE) for learning disentangled representations from user behavior. Our approach achieves macro disentanglement by inferring the high-level concepts associated with user intentions (e.g., to buy a shirt or a cellphone), while capturing the preference of a user regarding the different concepts separately. A micro-disentanglement regularizer, stemming from an information-theoretic interpretation of VAEs, then forces each dimension of the representations to independently reflect an isolated low-level factor (e.g., the size or the color of a shirt). Empirical results show that our approach can achieve substantial improvement over the state-of-the-art baselines. We further demonstrate that the learned representations are interpretable and controllable, which can potentially lead to a new paradigm for recommendation where users are given fine-grained control over targeted aspects of the recommendation lists.

研究动机与目标

• 动机：说明在推荐系统中需要解耦表示以提高鲁棒性和可解释性。
• 提出 MACRo-mIcro Disentangled Variational Auto-Encoder (MacridVAE)，将高层意图与低层项目属性分离。
• 通过概念原型和类别分配发展宏观解耦，通过 beta-正则化的 KL 目标实现微观解耦。
• 展示学习到的表示具有可解释性，并为用户提供对推荐的可控性。

提出的方法

• 提出一个生成模型，其中用户交互 x_u 是由潜在变量 z_u 和项目概念指示 C 生成的。
• 通过将 z_u 分裂为 K 个概念特定部分 z_u^(k) 并通过原型 m_k 将项目与 one-hot 概念向量 c_i 关联起来，实现宏观解耦。
• 使用变分编码器 q_theta(z_u|x_u,C) 和 基于余弦的原型分配以防止模式塌陷。
• 应用 beta-VAE 风格的目标，对 KL 项进行强化以促进微观解耦和维度间独立性。
• 使用采样 softmax 对大规模项目集进行可扩展解码，g_theta^(i)(z_u^(k)) 指导在概念 k 中项目 i 的似然性。
• 将先验 p_theta(z_u) 采纳为高斯分布，并使用基于余弦的解码器以对齐微观与宏观因素。
• 提供基于束搜索的方法用于用户可控推荐，以应对潜在维度的渐进变化。

实验结果

研究问题

• RQ1 宏观层面的解耦是否能够按概念分离用户意图，同时保持跨概念的偏好？
• RQ2 微观解耦是否能够在每个维度产生独立的潜在因子并提升推荐鲁棒性？
• RQ3 在真实数据集上，结合宏观和微观解耦是否能提升预测性能？
• RQ4 学习到的表示是否具备可解释性并可供用户控制？
• RQ5 基于余弦的原型分配是否相比内积方案更能缓解模式塌陷？

主要发现

数据集	方法	NDCG@100	Recall@20	Recall@50
AliShop-7C	MultDAE	0.23923 (±0.00380)	0.15242 (±0.00305)	0.24892 (±0.00391)
AliShop-7C	β-MultVAE	0.23875 (±0.00379)	0.15040 (±0.00302)	0.24589 (±0.00387)
AliShop-7C	Ours	0.29148 (±0.00380)	0.18616 (±0.00317)	0.30256 (±0.00397)
ML-100k	MultDAE	0.24487 (±0.02738)	0.23794 (±0.03605)	0.32279 (±0.04070)
ML-100k	β-MultVAE	0.27484 (±0.02883)	0.24838 (±0.03294)	0.35270 (±0.03927)
ML-100k	Ours	0.28895 (±0.02739)	0.30951 (±0.03808)	0.41309 (±0.04503)
ML-1M	MultDAE	0.40453 (±0.00799)	0.34382 (±0.00961)	0.46781 (±0.01032)
ML-1M	β-MultVAE	0.40555 (±0.00809)	0.33960 (±0.00919)	0.45825 (±0.01039)
ML-1M	Ours	0.42740 (±0.00789)	0.36046 (±0.00947)	0.49039 (±0.01029)
ML-20M	MultDAE	0.41900 (±0.00209)	0.39169 (±0.00271)	0.53054 (±0.00285)
ML-20M	β-MultVAE	0.41113 (±0.00212)	0.38263 (±0.00273)	0.51975 (±0.00289)
ML-20M	Ours	0.42496 (±0.00212)	0.39649 (±0.00271)	0.52901 (±0.00284)
Netflix	MultDAE	0.37450 (±0.00095)	0.33982 (±0.00123)	0.43247 (±0.00126)
Netflix	β-MultVAE	0.36291 (±0.00094)	0.32792 (±0.00122)	0.41960 (±0.00125)
Netflix	Ours	0.37987 (±0.00096)	0.34587 (±0.00124)	0.43478 (±0.00125)

• MacridVAE 在五个真实世界数据集上明显优于最先进基线，尤其是在稀疏、较小的数据集中。
• 宏观解耦使多样化的用户兴趣建模成为可能，并通过在概念内共享信息来缓解数据稀缺性。
• 微观解耦正则化产生更独立的潜在维度，提升鲁棒性与可解释性。
• 使用余弦相似度进行原型分配可防止模式塌陷，并比内积带来更有意义的概念聚类。
• 学习到的表示是可解释的，维度与人类可理解的概念对齐，并为潜在的用户可控推荐提供可能。