Cluster-Based Symbolic Compression of Time Series for Scalable Forecasting and Analysis

The increasing volume and velocity of time series data make it very challenging to perform effective and scalable predictive modeling. This paper presents a novel time series compression and transformation pipeline to address this challenge. We have combined three key components in a novel approach: sliding window segmentation for identifying local patterns, centroid-based clustering for converting continuous values into discrete symbols, and run-length encoding for compact representation of similar symbolic sequences. This combination allows us to convert raw time series data into easily understandable numerical symbolic representations, capturing essential time patterns while significantly reducing the volume of data. We utilized real-world metrics from Azure Function system-wide traces, including the number of applications, functions, invocations, and average execution time, which were collected every 5 minutes over a 14-day period, to evaluate the proposed method. Experimental results demonstrate impressive compression ratios while preserving pattern interpretations with lightweight computation, highlighting the method's effectiveness in reducing the learning cost for regression and forecasting tasks. The resulting model-agnostic representation can be easily applied to a variety of machine learning architectures, including recurrent and transformer-based models, providing an effective solution for scalable time series analytics in various fields, including both cloud and edge computing environments.