Abstract:
Problem: Software effort estimation is a crucial task that makes a project successful by
completing it in estimated time and budget. Traditional method often struggles because of the
agile project nature and lack accuracy making it difficult for the project managers to estimate
the accurate time needed to complete the project as a result project can get delayed and exceed
budget. This project aims to tackle these challenges by building a more accurate and
interpretable model that combines advanced machine learning techniques with explainability
driven feature selection method.
Methodology: The proposed solution involves a novel approach using china dataset. Initially,
data preprocessing is performed to handle skewness, outliers and scale features. We then used
SHAP (Explainable AI) values, derived from initial models, to select the most impactful
features. Our core model is a heterogeneous ensemble, combining Linear Regression, Decision
Tree, Random Forest, XGBoost, and CatBoost.Also experimentally developed and integrated
a custom neural network SHAP-Guided attention mechanism. Optuna is used to fine-tune the
hyperparameters for the neural network component.
Results: Testing on the China dataset resulted in excellent performance for the final ensemble
model (combining the base learners trained on SHAP-selected features), achieving an R² of
approximately 0.99 and a Mean Magnitude of Relative Error (MMRE) around 1.14%. This
indicates high accuracy and reliability on the test data. However, the evaluation also revealed
that the experimental SHAP-guided attention neural network component performed poorly
(negative R²), likely due to data limitations and complexity challenges, and it was therefore
excluded from the final successful ensemble prediction.
