Fourier Learning with Cyclical Data

Authors: Yingxiang Yang, Zhihan Xiong, Tianyi Liu, Taiqing Wang, Chong Wang

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate Fourier learning s better performance with extensive experiments on synthetic and public datasets, as well as on a large-scale recommender system that is updated in real-time, and trained with tens of millions of samples per day. 5. Numerical Simulations In this section, we numerically demonstrate the superiority of Fourier learning over the prior state-of-the-arts on synthetic and public datasets. 6. Fourier Learning in Recommender Systems In this section, we report the performance of Fourier learning implemented on a conversion rate (CVR) prediction model in an industrial recommender system.
Researcher Affiliation Collaboration 1Byte Dance Inc 2Paul G. Allen School of Computer Science & Engineering, University of Washington, WA.
Pseudocode Yes Algorithm 1 Streaming-SGD for F-MLP
Open Source Code Yes The source code and logs used to generate the reported experiments can be found at https://github.com/ Yangyx891121/Fourier-Learning.
Open Datasets Yes We classified the sentiment of tweets using a bag-of-words model over the Sentiment140 Twitter (Go et al., 2009) dataset. We demonstrate Fourier learning s better performance with extensive experiments on synthetic and public datasets, as well as on a large-scale recommender system that is updated in real-time, and trained with tens of millions of samples per day.
Dataset Splits No The paper uses concepts like 'streaming data' and 'online learning' where data arrives sequentially for training, but it does not provide specific percentages or counts for train/validation/test splits.
Hardware Specification No The paper states 'We trained the model on a distributed machine learning platform with 3,600 CPU cores', but it does not specify exact CPU models, GPU models, or other detailed hardware components used for experiments.
Software Dependencies No The paper does not mention any specific software dependencies with version numbers.
Experiment Setup Yes We conducted the experiment over 15 days of data, and plotted the results for lr = 0.1 in Figure 5 (see Appendix I for more results). Under the aforementioned design, we set the base frequency of the Fourier learning model to 2.4192MHz, or (28 days) 1, and set N = 28 4 in (7), providing a frequency band of up to (6 hours) 1. The coefficient ΞΎ is set to 0.5.