AmortizedPeriod: Attention-based Amortized Inference for Periodicity Identification

Authors: Hang Yu, Cong Liao, Ruolan Liu, Jianguo Li, Hu Yun, Xinzhe Wang

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Amortized Period surpasses the state-of-the-art methods by a large margin of 28.5% on average in terms of micro F1-score, with at least 55% less inference time.
Researcher Affiliation Industry Hang Yu, Cong Liao, Ruolan Liu, Jianguo Li , Yun Hu, Xinzhe Wang Ant Group, Hangzhou, 310013, China
Pseudocode Yes Algorithm 1 Amortized Period
Open Source Code Yes Code is available at https://github.com/alipay/Amortized Period.
Open Datasets Yes Yahoo Data2: We utilize the publicly available multiple-period data from Yahoo s webscope S5 datasets, specifically the Yahoo-A3 and Yahoo-A4 datasets. and App Flow Data34: We further collect the real traffic flow data for 311 micro-services deployed on 18 logic data centers in the cloud system of Ant Group. ... 3https://github.com/alipay/Amortized Period/data/Ant Data/appflow_data.zip.
Dataset Splits No For each dataset, we simulate 5000 samples and split them into training and testing sets with a ratio of 4 : 1.
Hardware Specification Yes All simulations of training Amortized Period are run using 8 NVIDIA TESLA A100 GPUs with 80 GB of VRAM. All inferences are run on a Mac Book Pro (16-inch, 2019) with a 6-core Intel i7 CPU and 16 GB of RAM.
Software Dependencies No For optimization, we use Adabelief (Zhuang et al., 2020) with β1 = 0.5, β2 = 0.999, and learning rate 1 10 4, since it offers more stable results than Adam. The paper does not specify version numbers for all key software components.
Experiment Setup Yes Unless otherwise specified, in all of our experiments, we set the hidden dimension in the CAB and SAB to 192. For optimization, we use Adabelief (Zhuang et al., 2020) with β1 = 0.5, β2 = 0.999, and learning rate 1 10 4, since it offers more stable results than Adam. We conduct training for 5000 epochs and select the checkpoints with the lowest training loss as the final model.