EqualityInsight-Project

Research Hypothesis

Countries with higher female graduation rates in science, technology, engineering, and mathematics (STEM) fields will have lower wage inequality.

Final Goal

The increase in the number of female STEM graduates can contribute to higher employment rates, women's economic independence, and ultimately foster economic growth while reducing inequality.
(Aligned with SDG Goal 5 (Gender Equality) and SDG Goal 8 (Decent Work and Economic Growth).)

---

Conclusion

1. Regression Analysis

• Best Model: XGBRegressor was identified as the optimal regression model, achieving the highest R² score of 0.7323 and the lowest RMSE of 6.1634.
• Attempts for Improvement:
  • Applied techniques such as hyperparameter tuning, outlier removal, data augmentation, and polynomial feature addition.
  • However, these approaches provided only limited improvement over the baseline model.
• Key Findings:
  • The results indicate that countries with higher STEM graduation rates tend to exhibit lower wage inequality. The regression model demonstrates a reasonable ability to explain wage gap variability.

2. Classification Analysis

• Initial Model: Logistic regression for binary classification of wage_gap achieved:
  • Accuracy: 72%
  • Recall (Class 1): 88%
  • However, the model suffered from low precision, resulting in many false positives.
• SMOTE for Imbalanced Data:
  • Addressed class imbalance using SMOTE, improving:
    • Accuracy: 75%
    • Precision (Class 1): 85%
  • Drawback: Recall for Class 1 dropped to 61%.
• Cost-Sensitive Logistic Regression:
  • Improved Recall (Class 1) significantly to 94%, addressing the imbalance effectively.

3. Data Source Insights

• Data source significantly influenced wage_gap values, highlighting its importance as a feature.
• Incorporating source-specific effects was critical in improving model performance and interpretability.

4. Conclusion

• Best Models:
  • For regression: The baseline XGBRegressor model is the most suitable.
  • For classification: The cost-sensitive logistic regression model performed best in capturing the nuances of class imbalance.
• Implications:
  • The analysis confirms that countries with higher STEM graduation rates for women are likely to experience less wage inequality.
  • These models provide a reasonable foundation for explaining wage disparities and highlight the role of STEM education in promoting economic equity.

---

1. Machine Learning Tasks

• Regression: Quantitatively analyze the impact of STEM graduation rates on wage gaps.
• Classification: Predict countries with low wage inequality and address data imbalance issues.
  • Apply oversampling techniques such as SMOTE to mitigate data imbalance and improve classification model performance.
  • Compare the performance of models such as XGBRegressor, Random Forest, and Logistic Regression using:
    • Regression Metrics: R², RMSE
    • Classification Metrics: Accuracy, Precision, Recall, F1-score

2. Data Definition

(1) Gender Wage Gap by Occupation (%) - ILOstat (2000-2023)

• Provides gender wage gap by occupation, country, and year.
• Variables:
  • Country Name
  • Gender Wage Gap by Occupation (%)
  • Year

(2) Distribution of Tertiary Graduates by Field of Study - UIS.Stat (2017-2023)

• Contains data on the distribution of tertiary graduates by field of study and gender.
• Includes the proportion of STEM graduates by gender.
• Variables:
  • Country Name
  • Graduation Rate by Field of Study and Gender
  • Year

(3)

• Merge the two datasets by country and year.
• Handle missing values and select key variables:
  • Include STEM graduation rates and wage inequality metrics as critical features for regression models.
• Feature engineering:
  • Add economic and social factors to improve model performance.

3. Data Processing Workflow

(1) Gender Wage Gap by Occupation (%) - ILOstat (2000-2023)

• Handling Rows with Flags:

  • Rows with the "Magnitude nil or negligible" flag: Set values to 0

    This flag indicates negligible or nearly non-existent values, so explicitly setting them to 0 ensures they do not affect the analysis.

  • Rows with the "Category not applicable" flag: Retain as-is (NaN)

    This flag indicates the data is not applicable to certain categories, so keeping them as NaN allows filtering during later analysis.

  • Rows with the "UIS Estimation" flag: Use time-based interpolation for missing values

    Since this flag indicates estimated data, time-based interpolation helps maintain data consistency and reliability over time.

  • Rows without any flags: Apply forward-fill and backward-fill for missing values

    Missing data without flags is likely incomplete, so standard fill methods help minimize missing data impact.

• Filtering Specific Indicators:

  • Retain only indicators relevant to the analysis to prevent unnecessary data processing and interference.

• Removing Unnecessary Columns:

  • Drop columns not directly relevant to the analysis to simplify the dataset.

(2) Distribution of Tertiary Graduates by Field of Study - UIS.Stat (2017-2023)

• Filtering Data:

  • Retain only rows where classif1.label is "Occupation (Skill level): Total".
  • Use data from 2010 onward to focus on recent trends, as older data may have lower relevance.

• Converting obs_value to Numeric:

  • Convert obs_value to numeric for calculations.
  • Non-numeric values are converted to NaN and handled as missing data.

• Calculating Group Statistics:

  • Compute the mean and standard deviation for each source.label group to enable normalization later.

• Merging Statistics:

  • Combine the computed statistics (mean, standard deviation) with the original dataset for group-level normalization.

• Standardizing Values:

  • Normalize values to improve comparability between groups and adjust for distribution differences.
  • This ensures fairness in model training and interpretation.

(3) Data Merging and Preprocessing

• Merging Datasets:

  • Combine data by matching on country and year to create a unified dataset for analysis.

• Column Standardization and Outlier Removal:

  • Ensure uniform column naming conventions for merging.
  • Remove outliers from wage_gap and grad_rate based on the 1st and 99th percentiles to improve model stability.

• Categorical Variable Encoding:

  • Encode source.label as a categorical variable using one-hot encoding for use in machine learning models.

• Defining Features and Target:

  • Independent variables (X): Include grad_rate and one-hot encoded source.label variables.
  • Dependent variable (y): Use wage_gap as the target.

• Splitting Dataset:

  • Split data into training and testing sets to evaluate the model’s generalization performance.