Whole dataset context-aware prediction on the null values in time series data for faster inferencing with low complexity

Abstract

Null value handling in time series datasets demands heavy and explicit data imputation models, which add latency during inferences for various tasks. This concern raises the following question: is it possible to avoid these explicit data imputation models to perform task predictions directly during inference without imputation? As a pioneer, this paper proposes ZeroTIP, a knowledge distillation (KD)-based Zero Time Imputation for Prediction strategy for workload predictions without having additional data-imputation models. During the training period, a student network is forced to reason the missing or null values implicitly and mimic the inference (workload prediction task) while taking synthetically corrupted data as input and being supervised by the pretrained teacher network that contains representations of the original dataset. Only the student network is used during inference. ZeroTIP reduced the inference time by almost 99.9% by avoiding explicit data imputation. A version of ZeroTIP, called ZeroTIP-DI, was deployed for the data imputation task to evaluate the significance of ZeroTIP in reasoning data context and pattern. For a prediction length of 48 and 96, ZeroTIP-DI achieved an average improvement of 38.37 (97.08%) and 21.67 (95.08%) times the baseline, highlighting its superiority.