Missing Data Imputation Using C

This project represents my first university project, developed to explore classic missing data imputation techniques using C. The goal was to implement and evaluate five methods on three real-world datasets, with each method coded from scratch and adapted to the specific structure and missingness patterns of each dataset.

Missing data imputation is a critical preprocessing step in any data pipeline. This project systematically compares several popular techniques to highlight their strengths, limitations, and contextual suitability.

---

Techniques Implemented

Each imputation strategy was implemented from scratch in C, including logic tailored to each dataset:

Method	Description
Linear Regression Imputation	Builds a regression model to predict and impute missing values using correlated variables.
k-Nearest Neighbors (k-NN)	Uses similarity-based voting among nearest neighbors to estimate missing entries.
Mean/Median Imputation	Fills missing values using column-level statistical estimates (mean or median).
Listwise Deletion	Removes entire rows with missing entries — easy but causes potential information loss.
Hot Deck Imputation	Imputes missing values using similar “donor” observations from the same dataset.

---

Datasets Used

Dataset	Description	Size	Key Features with Missing Data
Pima Indian Diabetes	Medical dataset of diagnostic measures and outcomes	768 rows × 9 columns	Plasma Glucose, BMI, Blood Pressure
World Happiness Report	Global socio-economic indicators of happiness	1,949 rows × 11 columns	Perceptions of Corruption
Melbourne Housing	Real estate listings from 2016–17	13,580 rows × 21 columns	Land Size, Building Area

---

Sample Evaluation Summary

Method	Pima Indians	World Happiness	Melbourne Housing
Linear Regression	✅ Excellent	✅ Excellent	✅ Excellent
KNN	✅ Good	✅ Excellent	⚠️ Poor
Mean/Median	✅ Accurate but repetitive	✅ Consistent	⚠️ Limited uniqueness
Listwise Deletion	⚠️ Significant data loss	✅ Effective	✅ High performance
Hot Deck	⚠️ 2.4% missing left	✅ Excellent	✅ Good

See /analysis for full breakdowns.

---

Repository Structure

├── data/
│   ├── raw/            # Original datasets
│   └── imputed/        # Post-imputation outputs by method
├── src/
│   ├── linear_regression/
│   ├── knn/
│   ├── mean_median/
│   ├── listwise_deletion/
│   └── hot_deck/
├── analysis/           # Result interpretations per method
├── LICENSE
└── README.md

---

Key Learnings

• Linear Regression and KNN provide robust, context-aware imputation when feature correlation is strong.
• Mean/Median is effective in reducing missingness but may oversimplify distributions.
• Listwise Deletion, while easy to implement, often discards valuable data and should be used cautiously.
• Hot Deck's performance varies widely depending on dataset density and feature spread.

---

License

This project is licensed under the MIT License. See the LICENSE file for more information.