Aspect-Based Sentiment Drift Detection

Sentiment time series with drift detection

Extracting aspect-level sentiment from 312K Amazon smartphone reviews using PyABSA (0.883 macro F1), aggregated into multivariate time series per product aspect (battery life, camera, display, etc.). Drift detection compares an LSTM-Autoencoder against statistical baselines i.e. PELT and CUSUM across 331 dense products.

Model comparison: LSTM-AE vs PELT vs CUSUM

Implementing Captum Integrated Gradients for token-level explainability: which words drove a sentiment shift? An interactive dashboard allows portfolio monitoring, drift event observation, and model comparison across products and aspects.

Token-level Integrated Gradients attribution

The system is designed for consumer electronics brand teams who need early warning when a firmware update or supply chain change silently degrades a product attribute before it shows up in star ratings.

Early warning alert