Quantification over Time

Abstract

Quantification is the supervised machine learning task that estimates the class distribution in a sample. Therefore, quantification applications typically involve predicting aggregated quantities, such as the prevalence of positive comments about a product, personality or company on a set of social media posts. However, quantification analysis is more informative when performed over time, such as when we are interested in tracking public opinion on social media and relating changes in opinion with relevant events. The vast majority of the literature considers quantification as a standalone task, assuming the output of quantifiers to be independent even when applied to temporal data. This paper proposes a new quantification task, Quantification over Time (QoT), that allies quantification with time series forecasting methods. We propose an approach based on the Kalman filter, which can help improve the performance of standalone quantifications and a general framework that includes both ours and SOTA methods. In an experimental comparison with several textual datasets and numeral datasets, we show that our method outperforms existing methods for QoT in the literature, such as a simple composition of the classify and count method with moving averages and ReadMe2 as a standalone quantifier. We also show that our proposal can outperform several baselines, including recently proposed quantifiers used as standalone approaches. Codes are available at https://github.com/frieli11/quantification-over-time .