Feature Re-Learning with Data Augmentation for Content-based Video Recommendation

Introduction

作者主要将特征再学习,使得在对应的特征空间相关性强的点能够靠的比较接近比原始的特征空间。

Proposed solution

Augmentation for frame-level features

采用skip sample的方式

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

Augmentation for video-level features

给特征增加噪音,使模型更加具有鲁棒性

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

Model structure

利用一个全连接层将原始的特征映射到新的空间。

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

作者从经验上发现模型的表现对于维度不敏感。

Model training

为了使相关的视频的cosine相似度更大,不相关的视频的cosine相似度更小,因此作者提出triplet ranking loss:

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

优化目标即最小化loss:

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

Evalution

Choice of loss functions:

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

Effectiveness of data augmentation:

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

Effectiveness of feature re-learning:

Feature Re-Learning with Data Augmentation for Content-based Video Recommendation简介

相关文章: