[论文解读] Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering

Ruining He, Julian McAuley|arXiv (Cornell University)|Feb 4, 2016

Generative Adversarial Networks and Image Synthesis参考文献 36被引用 203

一句话总结

本文通过在 Bayesian Personalized Ranking framework 内结合深度 CNN 视觉特征、基于 epoch 的时尚演化以及轻量级非视觉动态，开发了时序演化、具视觉感知的一类单类协同过滤模型用于时尚推荐。

ABSTRACT

Building a successful recommender system depends on understanding both the dimensions of people's preferences as well as their dynamics. In certain domains, such as fashion, modeling such preferences can be incredibly difficult, due to the need to simultaneously model the visual appearance of products as well as their evolution over time. The subtle semantics and non-linear dynamics of fashion evolution raise unique challenges especially considering the sparsity and large scale of the underlying datasets. In this paper we build novel models for the One-Class Collaborative Filtering setting, where our goal is to estimate users' fashion-aware personalized ranking functions based on their past feedback. To uncover the complex and evolving visual factors that people consider when evaluating products, our method combines high-level visual features extracted from a deep convolutional neural network, users' past feedback, as well as evolving trends within the community. Experimentally we evaluate our method on two large real-world datasets from Amazon.com, where we show it to outperform state-of-the-art personalized ranking measures, and also use it to visualize the high-level fashion trends across the 11-year span of our dataset.

研究动机与目标

• 对隐式反馈（购买）建模用户对时尚的个性化排序。
• 捕捉随时间变化影响时尚选择的视觉因素。
• 在更好地解释和推荐的前提下，将视觉演变与非视觉的时序动态进行解耦。
• 实现在大规模时尚目录上对稀疏、长尾商品的可扩展训练。

提出的方法

• 在矩阵分解中扩展视觉交互：x_hat = alpha + beta_u + beta_i + <gamma_u, gamma_i> + <theta_u, theta_i>.
• 将高维 CNN 图像特征 f_i 通过 E 映射到 K'-维视觉空间：theta_i = E f_i.
• 通过随时间变化的分量建模时序演化：E(t) = E + Delta_E(t); theta_i(t) = E(t) f_i 且按需计算 theta_u(t)。
• 引入时间加权和基于 epoch 的分段以捕捉非线性时尚变化：theta_i(t) = E f_i ⊙ w(t) + Delta_E(t) f_i，以及 epoch 特定参数 Theta_ep。
• 结合非视觉的时序动态 beta_i(t) 和 beta_{C_i}(t) 以将视觉效应与非视觉效应分离。
• 使用 Bayesian Personalized Ranking (BPR) 进行优化，并采用在参数拟合与时尚 epoch 分段 Lambda 之间交替的坐标上升法。

实验结果

研究问题

• RQ1 如何在一类协同过滤框架中纳入视觉外观及其时序演化？
• RQ2 相较于固定时间模型，基于 epoch 的分段是否能更好地捕捉随时间的非线性时尚动力学？
• RQ3 引入轻量级非视觉时序动态对预测性能与可解释性有何影响？
• RQ4 在真实时尚数据集上，视觉信息模型在 warm-start 和 cold-start 情况下的表现如何？

主要发现

• 所提出的 TVBPR 家族在两个大型 Amazon Women’s and Men’s Clothing 数据集上优于最前沿的个性化排序方法。
• 基于 epoch 的分段比固定参数时间模型更好地捕捉非线性时尚演变。
• 通过嵌入矩阵从深度 CNN 提取的视觉特征的引入，在稀疏、长尾商品设置下提升了推荐质量。
• 在总体层面的时序动态（时尚趋势）显著提升了个性化排序的效果，但在各用户层面的漂移在他们的数据集中影响较小。
• 该模型支持对 11-year-span 的学习时尚趋势的可视化，展示了复杂的非线性视觉演化。