| marp | author | size | theme | title | paginate | header | footer | style |
|---|---|---|---|---|---|---|---|---|
true |
Simone Meneghello, Alessio Piraccini, Gianluca Tori |
58140 |
default |
PreFer data challenge |
true |
ODISSEI 2024, PreFer data challenge |
Simone Meneghello, Alessio Piraccini, Gianluca Ivo Tori |
section {
font-size: 25px;
text-align: left;
}
img {
max-width: 95%;
max-height: 95%;
object-fit: contain;
}
h2 {
position: absolute;
top: 70px;
left: 50px;
}
|
Simone Meneghello, Alessio Piraccini, Gianluca Ivo Tori
-
Recent data priority: Key insights for predicting fertility (2021-2023) likely stem from surveys around 2021. Focus on recent surveys, supplemented by relevant past data.
-
Feature selection: Despite numerous survey questions, only a few features may be crucial. Identifying and selecting these key features is essential to streamline the dataset.
-
Handling missing values: Missing data might not be random and could hold valuable information. Some missing responses result from survey branching logic. Longitudinal data may allow filling gaps with past information.
-
Key predictors: Likely predictors include age, relationship status, economic situation, and existing children. However, the dataset's breadth may reveal unexpected patterns.
-
Task Overview: classify if a person will have a child within 2021-2023 using 2020 data.
-
Dataset Details:
- Initial dataset: 6,418 rows, 31,634 columns.
- Cleaned dataset: 987 rows, 25,868 columns (removed rows with missing outcomes and columns with all missing values).
-
Background Dataset:
- Longitudinal format with repeated entries per subject across years.
- Filtered to include subjects with available outcomes.
- Contains key predictive features (e.g., age, income, civil status).
- Used a univariate decision tree with stratified 5-fold cross-validation for feature evaluation.
- Performance decreases over time, supporting the relevance of recent surveys.
- Most surveys show limited key variables.
- Key variables are desire for children, income, age, civil status, gynecologist visits
Key considerations:
- missing data can often be attributed to survey logic or non-participation
- missing data in key variables, like future child plans, often correlates with not having a child
We used a longitudinal approach for imputation, filling gaps with the most recent past responses.
| quest2015 | quest2016 | quest2017 | quest2018 | quest2019 | quest2020 | quest2020_imp |
|---|---|---|---|---|---|---|
| 4 | -- | 4 | -- | 5 | 5 | 5 |
| -- | -- | 5 | 3 | -- | -- | 3 |
| 2 | -- | -- | -- | -- | -- | 2 |
- Propensity to have a child is the strongest predictor.
- Income, relationship status, and Big Five traits are significant.
- Immediate plans for children and desire for children correlate with having children.
- Recent cohabitation and individuals in their thirties are more likely to have children.
- Mostly married or never married have children.
- Income shows no clear difference between groups.
- Used stratified shuffle split (k=10, 30% validation) for validation to improve test performance alignment.
- Focused on tree-based models: Random Forest, Extra Trees, and Gradient Boosting (CatBoost, LightGBM, XGBoost).
- Employed Bayesian hyperparameter tuning.
- Combined CatBoost, LightGBM, Extra Trees using weighted voting (CatBoost weight: 0.5).
- Final model: weighted voting of CatBoost, LightGBM, Extra Trees; CatBoost is the best standalone model.
- Intention of having a child and related features.
- Relationship status: start year of cohabitation, civil status, and marriage flag are significant.
- Age is the third most important feature.
- Additional predictors include "Go to the gynecologist" and Big Five personality traits (agreeableness, conscientiousness).
- Higher SHAP values for individuals desiring children, especially older subjects.
- Increased predicted probability for people in their thirties planning another child soon.
- Married individuals, particularly those marrying young, show higher predicted probabilities.
- Recent cohabitation (within ten years) and age 28-38 correlate with higher SHAP values.
- Visiting a gynecologist indicates a higher expected probability, supporting its role in identifying pregnant women.
- Achieved ~90% accuracy with the final model; sophisticated techniques offer slight performance improvements but aren't essential.
- Key predictors of fertility are intention to have a child and relationship status, as expected.
- Personality traits, analyzed via factor analysis, can enhance prediction accuracy.
- Focused on investigating how Big Five personality traits affect childbearing propensity.
- According to this model the personality traits are Extraversion, Agreeableness, Conscientiousness, Emotional Stability, and Openness; viewed as spectrums. It's based on factor analysis of psychometric data, validated across cultures.
- Survey questions are rated on a 5-point scale, missing values are imputed using previous year data due to trait stability.
- Used Factor Analysis on survey questions to obtain five traits to be used as features for the final model.
| Neuroticism | Extraversion | Agreeableness | Conscientiousness | Openness | |
|---|---|---|---|---|---|
| (+) Get upset easily. | (-) Keep in the background. | (+) Sympathize with others’ feelings. | (+) Like order. | (+) Have excellent ideas. | |
| (+) Have frequent mood swings. | (-) Am quiet around strangers. | (-) Am not really interested in others. | (-) Often forget to put things back in their proper place. | (+) Am full of ideas. | |
| (+) Get stressed out easily. | (-) Don’t talk a lot. | (+) Am interested in people. | (-) Leave my belongings around. | (+) Have a rich vocabulary. |