Fast Fourier transform (FFT) is a foundational primitive in data mining, supporting a broad range of pipelines for feature extraction, filtering, compression, and large-scale learning.
However, modern workloads, including high-dimensional signals, la...
Fast Fourier transform (FFT) is a foundational primitive in data mining, supporting a broad range of pipelines for feature extraction, filtering, compression, and large-scale learning.
However, modern workloads, including high-dimensional signals, large-scale tensors, and continuously arriving observations, often exhibit strong frequency-domain structure in which only a small, structured portion of the spectrum is informative.
In such settings, conventional practice still relies on full transforms followed by discarding most coefficients, and in streaming scenarios it frequently resorts to full or frequent retraining, incurring unnecessary computational cost and limiting responsiveness as data grow.
This thesis develops FFT-centric theory and algorithms that explicitly exploit frequency-domain structure to scale data mining across both offline and online regimes.
The proposed partial-spectrum computation directly evaluates task-relevant Fourier components with provable, user-controlled approximation error, eliminating the overhead of computing unused coefficients.
We further enable rapid, automatic accuracy-speed reconfiguration in multidimensional settings through an efficient optimization formulation, making partial Fourier computation practical under changing data shapes and requirements.
Beyond spectral access, this thesis develops frequency-informed learning mechanisms for data mining: lightweight, invertible frequency-domain transformations are learned to align multi-way data and reduce the effective rank for downstream tensor factorization, and an online coupled factorization framework is presented to support real-time updates via frequency regularization and adaptive forgetting rather than full retraining.
Experiments on diverse real and synthetic datasets demonstrate consistent and substantial acceleration, achieving up to 19x speedup while preserving accuracy and improving reconstruction quality, compression efficiency, and anomaly detection performance under practical budgets.
Collectively, the results show that FFT can serve not only as a fast transform, but as a scalable design principle for approximation, representation, and continuous learning in data mining systems.