Frequency Domain Predictive Modelling with Aggregated Data

Avradeep Bhowmik, Joydeep Ghosh, Oluwasanmi KoyejoAISTATS 2017 · OpenReview

Abstract

Existing work in spatio-temporal data analysis invariably assumes data available as individual measurements with localised estimates. However, for many applications like econometrics, financial forecasting, and climate science, data is often only accessible in aggregated forms—such as quarterly averages or regional totals. This paper introduces a novel framework for training predictive linear models directly from these aggregates by shifting the problem into the frequency domain, bypassing the need for error-prone data reconstruction.

Introduction

In many empirical sciences, researchers face a data granularity gap. While the goal is often to understand individual-level relationships, the available data is frequently aggregated over time or space. Standard approaches typically involve naive imputation (using group averages as proxies) or complex disaggregation techniques to reconstruct missing samples. Both methods are fraught with risk: naive imputation can lead to the ecological fallacy, where group-level correlations misrepresent individual-level truths, while reconstruction often introduces significant synthetic noise.

This research addresses a fundamental question: Can we provably learn an individual-level linear model using only misaligned, aggregated data? The proposed solution leverages the mathematical properties of the Fourier transform to solve the regression problem directly in the frequency domain.

Methodology: From Time to Frequency

The core innovation of this work is treating aggregation not as a loss of data, but as a signal processing filter.

Aggregation as a Filter

Taking a sliding-window average in the time domain is mathematically equivalent to a convolution. In the frequency domain, this operation simplifies to multiplying the signal's Fourier Transform by a "sinc" function—a specific type of low-pass filter. By viewing the problem this way, the researchers can relate global frequency properties to local time-domain properties.

The Fourier Estimation Algorithm

The proposed algorithm follows three main steps:

  1. Frequency Sampling: The method samples frequencies within a specific low-frequency window $(-\omega_0, \omega_0)$. This window is chosen because low-frequency components are more robust to the distortions caused by aggregation (aliasing).
  2. Sinc Correction: It computes the finite Fourier transforms of the aggregated features and targets, then "undoes" the aggregation effect by dividing by the known sinc function associated with the averaging window.
  3. Direct Optimization: Finally, it solves for the model parameter $\beta$ by minimizing the squared error in the frequency domain. This allows for learning the underlying relationship without ever synthesizing individual-level data points.

Experimental Evaluation

The framework was tested on several real-world and synthetic datasets:

  • Las Rosas Dataset: Agricultural data regressing corn yield against nitrogen levels and topography.
  • Forest Fires Dataset: Predicting burned acreage in Portugal using meteorological variables like humidity and temperature.
  • Comprehensive Climate Dataset (CCDS): Modeling atmospheric vapor levels based on greenhouse gas concentrations and cloud cover.

Key Results

The experiments demonstrated that the Fourier-domain estimator consistently outperforms traditional time-domain baselines. As shown in the figures below, the Fourier estimation error decreases as the window size $\omega_0$ is optimized, eventually approaching the performance of an "Oracle" model trained on the original, non-aggregated data.

Estimation error across Las Rosas, Forest Fires, and CCDS datasets comparing Fourier estimation to time-domain baselines and the non-aggregated oracle.

A significant advantage of this approach is its ability to handle misaligned data. Because the Fourier transform treats the signal globally, features aggregated over different periods (e.g., monthly vs. quarterly) can be corrected individually using their respective sinc functions before being combined in the regression model.

Discussion and Limitations

While the frequency-domain approach offers a robust alternative to time-domain imputation, it relies on certain assumptions. The current framework assumes "weak stationarity," meaning the statistical properties of the data do not change over time. Additionally, the method is currently limited to linear models.

One critical trade-off identified is the aliasing effect. If the frequency window $\omega_0$ is too large, high-frequency noise can leak into the low-frequency estimates, degrading performance. Future work aims to extend this logic to non-linear models, such as neural networks, and to develop automated methods for selecting the optimal frequency window.

Summary updated

Abstract

Existing work in spatio-temporal data analysis invariably assumes data available as individual measurements with localised estimates. However, for many applications like econometrics, financial for...
BibTeX
@inproceedings{bhowmik2017frequency,
  author = {Avradeep Bhowmik and Joydeep Ghosh and Oluwasanmi Koyejo},
  title = {{Frequency Domain Predictive Modelling with Aggregated Data}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2017},
  url = {https://openreview.net/forum?id=lXBn_Zwykzf},
}
Figure