2024
Pistachio Dataset Analysis and Modeling
Built an end-to-end machine learning pipeline over pistachio images and handcrafted features covering classification, outlier detection, and clustering, reducing reliance on manual inspection by combining engineered and deep-learned feature representations with automated model evaluation. Solo ML project.
Headline outcome
Cut experiment turnaround by persisting extracted features as reusable artifacts, and surfaced mislabeled samples the supervised pipeline alone could not see.
Stack
Problem → solution → result
Problem. Neither handcrafted features nor learned embeddings alone separated the classes reliably.
Solution. I combined 28 handcrafted morphological, shape, and color features with deep embeddings from a pretrained CNN and three custom architectures, persisting extracted features to disk.
Result. A richer fused representation, and repeated CNN extraction removed from every experiment.
Problem. Small, heavily imbalanced dataset made models collapse toward the majority class.
Solution. I applied synthetic minority oversampling (SMOTE) inside the pipeline with stratified splits, so resampling and scaling were fit only on training folds.
Result. Improved minority-class recall while preventing the data leakage that inflates scores during hyperparameter search.
Problem. Concatenating tabular features with CNN embeddings produced a high-dimensional, highly correlated space.
Solution. I applied correlation filtering, Recursive Feature Elimination, and PCA/LDA reduction.
Result. Stable classification and clustering on the combined data, with class structure visualizable in 2D and 3D.
Problem. Model choices were made ad hoc, so results weren't comparable.
Solution. I implemented a grid-search selection workflow spanning Naive Bayes, SVM, MLP, and stacking ensembles, persisting fitted models and features.
Result. Repeatable, directly comparable evaluation across all candidates.
Problem. Unlabeled anomalies and latent groupings were invisible to the supervised pipeline.
Solution. I architected an unsupervised layer with outlier detection (Local Outlier Factor, Isolation Forest, One-Class SVM) and a clustering evaluator (KMeans, DBSCAN, Gaussian Mixture) scored on three internal indices.
Result. Surfaced mislabeled and anomalous samples and let cluster counts be chosen on evidence rather than by eye.
Problem. Preprocessing, extraction, and modeling were entangled, so any change forced a costly full re-run.
Solution. I reorganized into separate modules with the cleaned dataset and extracted features saved as reusable artifacts.
Result. Classical ML experiments iterate without re-running CNN feature extraction, reproducible from saved artifacts.