Aman Atman supervised by Dr. Santosh Nannuru received his Master of Science – Dual Degree  in  (LCD). Here’s a summary of his research work on Space-Time Aware Neural Networks for Efficient Multivariate Time Series Imputation:

Time series are ubiquitous and are of interest because of their impact. For example, air quality and traffic affect our health and time. Stock market fluctuations can change our finances. Missing values in time series data is a frequently occurring problem in multi-sensor data analysis. Imputation of these missing values is a vital pre-processing step for forecasting and other downstream tasks like anomaly detection. In this work we address the problem of accurately imputing missing values in multivariate time series. We propose two novel architectures– P-GUTS and SAINT. These neural networks leverage reinforced space-time inductive biases and sparse attention mechanisms to reduce error on various benchmark imputation datasets. We use attention to avoid error compounding problems faced by auto-regressive methods. We operate directly on space-time product graphs to avoid information loss. P GUTSleverages multiple representations of data using different temporal pooling factors and effectively aggregate them using U-networks. SAINT is an efficient network which uses channel independent networks and sparse space-time graph transformers to prune connections from the product graph which can otherwise become huge. P-GUTS has more parameters, performs more computations and can perform better than SAINT in multiple datasets. However, both these architectures show consistent improvement in performance in comparison to the state-of-the-art networks on various imputation benchmarks and across increasing missing rates. We also test these networks on spatiotemporal forecasting with encouraging results. Sparsity in SAINT further makes it practical also for long-term forecasting.

May 2025