Training via Confidence Ranking

Jian Zhu

Keywords: loss function

Abstract: Model evolution and constant available data are two common phenomenon in large-scale real-world machine learning application, e.g. ads and recommendation system. To adapt, real-world system typically operates both retraining with all available data and \textit{online-learning} with recent available data to update models periodically with the goal of better serving performance for future data. However, if model and data evolution results in a vastly different training manner, it may induce negative impact on online A/B platform. In this paper, we propose a novel framework, named Confidence Ranking, to design optimization objective as a ranking function with two different models. Our confidence ranking loss allows directly optimizing the logits output for different convex surrogate function of metrics, e.g. AUC and Accuracy depending on the target tasks and datasets. Armed with our proposed methods, our experiments show that the confidence ranking loss can outperform for test-set performance on CTR prediction and model compression with various setting against the knowledge distillation baselines.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)

TL;DR: We devise a series of loss function for training a new better model than deployed one in real-world machine learning system.