Iris Data Quality Experiment

A controlled experiment to understand the impact of data quantity vs. quality on machine learning models.

Research Question & Hypotheses

RQ: Does increasing the test set size (thus reducing training data) cause a predictable decrease in decision tree accuracy? How does label noise moderate this effect?

H1 (Data Quantity): Accuracy will decline on average as training data shrinks, but the relationship may not be strictly monotonic due to sampling variability.

H2 (Data Quality): Introducing label noise will drastically reduce mean accuracy and inflate variance, especially when training data is scarce.

Key Findings

• On clean Iris data, accuracy declines as training data shrinks, but the trend is not strictly monotonic—random variation in splits matters.
• Introducing only 10% label noise causes a large drop in mean accuracy and dramatically wider error bars, showing that data quality can dominate sample size effects.
• Error bars are essential: without them, small fluctuations could be misinterpreted as meaningful patterns.

Results

The following plots show the mean accuracy and standard deviation over 100 different random train/test splits for each test size.

Accuracy vs Test Size (Clean Data)

Accuracy vs Test Size (10% Label Noise)

The experiment was conducted using a Decision Tree Classifier on the normalized Iris dataset.

How to Reproduce

1. Clone the repository:

   git clone https://github.com/BenyaminMahdavifar/iris-data-quality.git
   cd iris-data-quality