Data Science with Python: Internship Portfolio

Work completed during the Virtual Data Science with Python Trainee internship at YuvaIntern (Henry Harvin), Aug-Oct 2026. Each week has its own folder with the Python code, generated figures/tables, and the full written report in reports/.

All 6 weeks are in this repo.

Week	Topic	Dataset	Key result
1	Data acquisition, cleaning & preprocessing	Titanic (891 x 12)	891 x 16 clean, 0 missing, no rows dropped; 5-fold CV accuracy 80.25% -> 82.60% vs naive "drop NA"
2	Exploratory data analysis & visualization	Titanic (cleaned)	13 annotated plots + 2 aggregation tables; 1st-class women 97% vs 3rd-class men 14% survival
3	Unsupervised learning & clustering	UCI Wine (178 x 13)	K-Means (k=3) + Ward hierarchical; silhouette 0.285; ARI 0.897 vs true cultivars
4	Supervised learning (classification)	Titanic (cleaned)	6 models compared with 5-fold CV; tuned Gradient Boosting on held-out test: 81.6% acc, F1 0.740, ROC-AUC 0.862
5	Deep learning (CNN, TensorFlow/Keras)	Fashion-MNIST (70k images)	75k-param CNN with augmentation + BatchNorm + Dropout; ~90% test accuracy (90.2% in the committed run), ablation vs MLP and plain CNN
6	Capstone: end-to-end pipeline	Titanic (raw)	Acquire -> clean -> EDA -> supervised + unsupervised; tuned GB 81.6% acc / ROC-AUC 0.862; K-Means k=2 silhouette 0.433 (survival 30% vs 65%), plus a k=4 view that isolates large families (12% survival)

Repository structure

.
├── README.md
├── requirements.txt
├── LICENSE
├── data/
│   ├── README.md
│   ├── raw/                       # titanic.csv, wine.csv, fashion-mnist/
│   └── processed/                 # titanic_cleaned.csv, capstone_cleaned.csv, data_dictionary.csv, sample_cleaned_head10.csv
├── scripts/download_data.py       # re-downloads the raw Titanic CSV
├── week1_data_cleaning/           # 01_explore.py, 02_clean_preprocess.py, 03_additional_analysis.py
├── week2_eda/                     # eda_analysis.py
├── week3_clustering/              # clustering_analysis.py
├── week4_supervised/              # model_pipeline.py
├── week5_deep_learning/           # data_loader.py, dl_pipeline.py
├── week6_capstone/                # capstone_pipeline.py
└── reports/                       # weekly .docx reports

Every weekN_*/ folder writes plots to figures/ and tables to results/.

Setup

git clone <your-repo-url>
cd internship-data-science-python
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Run order

Scripts can be run from any directory. Week 2 needs the cleaned file that Week 1 produces; Week 3 is standalone; Week 4 also needs the Week 1 output; Week 6 is standalone (runs its own condensed cleaning from the raw CSV); Week 5 needs data/raw/fashion-mnist/ (included, or run python scripts/download_data.py).

python week1_data_cleaning/01_explore.py            # profile raw data
python week1_data_cleaning/02_clean_preprocess.py   # writes data/processed/titanic_cleaned.csv
python week1_data_cleaning/03_additional_analysis.py
python week2_eda/eda_analysis.py
python week3_clustering/clustering_analysis.py      # uses scikit-learn's built-in wine data
python week4_supervised/model_pipeline.py           # needs data/processed/titanic_cleaned.csv (Week 1)
python week5_deep_learning/dl_pipeline.py           # TensorFlow; ~30 min on CPU
python week6_capstone/capstone_pipeline.py          # standalone, reads raw Titanic; ~15 s

Random seeds are fixed (random_state=42), so the numbers above reproduce exactly.

Methods at a glance

• Week 1: missingness diagnosed before fixing (chi-square test shows Age is MAR w.r.t. Pclass, so group-wise median imputation); Cabin turned into a HasCabin flag; 15 zero fares re-imputed; Fare outliers compared across Z-score / IQR(1.5) / IQR(3) and winsorised, not dropped; engineered FamilySize, IsAlone, Title. Data dictionary: data/processed/data_dictionary.csv.
• Week 2: univariate, bivariate and multivariate analysis with groupby aggregation tables; findings tied back to Week 1 preprocessing choices (e.g. Age capping removes the 60+ group).
• Week 3: features standardised; k picked with elbow + silhouette (peak at k=3); K-Means cross-checked with Ward hierarchical (169/178 samples agree); PCA used only for 2-D visualisation (55.4% variance); clusters profiled per feature and validated externally against the cultivar labels, which were never used for fitting.
• Week 4: stratified 80/20 split (712/179), scaler fit on train only (no leakage), 6 algorithms scored on 5 metrics via stratified 5-fold CV, GridSearchCV on Random Forest and Gradient Boosting, final test evaluation with confusion matrix, ROC curve, feature importance and learning curve. Top predictors: Title, Pclass, Sex and an Age x Class interaction, which matches the Week 2 EDA.
• Week 5: Fashion-MNIST split 54k train / 6k validation / 10k test; CNN (4 conv layers, BatchNorm, Dropout, GlobalAveragePooling) with random translation/zoom/flip augmentation, Adam + EarlyStopping + ReduceLROnPlateau; per-class precision/recall/F1, confusion matrix, misclassified samples, and an ablation study. Shirt is the weakest class (confused with T-shirt, Pullover, Coat), which is a known property of Fashion-MNIST at 28x28. Note: TensorFlow on CPU with random augmentation is not bit-reproducible even with fixed seeds, so a re-run lands within about 1 point of the committed outputs (committed run: 90.24% test accuracy, MLP baseline 84.1%, unregularised CNN 69.8%). The .docx report quotes an earlier run (89.25%); the conclusions are the same.
• Week 6: full pipeline in one script: raw data -> condensed Week 1 cleaning -> EDA -> tuned Gradient Boosting (stratified split, 5-fold CV, GridSearchCV on F1) -> K-Means segmentation with elbow + silhouette. k=2 is chosen on silhouette; k=4 is shown as a supplementary view because it separates large families (56 passengers, 12% survival) and a high-survival women/children group. Clusters are built without the target, then profiled against survival.

Limitations / future work (Week 3)

Silhouette of ~0.28 means clusters overlap somewhat, which is expected since chemically similar wines exist across cultivars. With only 178 samples from one region, boundaries may not generalise. Next steps would be trying Gaussian Mixture Models (soft assignment), DBSCAN, and a stability check across random seeds.

Limitations / future work (Weeks 4-5)

Week 4: only 891 rows, so test metrics (179 passengers) are noisy; recall (68%) is lower than precision, so a threshold shift or class weights could be tried. Week 5: trained on CPU with a 12-epoch cap; a deeper network or longer schedule might lift the upper-body garment classes slightly, but much of the Shirt confusion is a data limitation.

Limitations / future work (Week 6)

Single small dataset, so the pipeline is a demonstration rather than a generalisable result. Next steps: bootstrap-based cluster stability, probability calibration and threshold tuning for the 68% recall, and external validation on a second dataset.

Reports

reports/ has the full write-ups (methodology, figures, interpretation, limitations).

Author

Aman: B.Tech Software Engineering, NIAT Jaipur · LinkedIn

Released under the MIT License.