Lost in Noise: When cleaning up clouds the picture. Fuzzy topic modeling and robust low-rank decomposition

This preliminary study assesses the impact of noise-removing techniques, such as Principal Component Pursuit (PCP), on the document-term matrix before topic modeling. Specifically, fuzzy Latent Semantic Analysis (fLSA) is applied to a benchmark dataset of Air France customer reviews to evaluate how different input representations – namely, the standard term-frequency matrix and its low-rank approximation via low-rank decomposition – affect topic coherence and interpretability. Initial results indicate that while fLSA effectively extracts meaningful topics, noise removal via PCP introduces distortions, altering topic structure.