Will Your Marriage Beat the Odds?

Linear Regression

For the original Turkish dataset, we chose to do Lasso and Ridge linear regressions to reduce model complexity. Ridge puts constraints on the coefficients which in effect, shrinks the coefficient.

Learn More

Linear Regression

×



Residual plot of test and training data sets looked promising.



Close

Heatmap

In addition to linear regression, we also used a heatmap to get an overall look into the distribution of each attribute in our dataset. We used seborn to create the heatmap looking at all 54 survey questions and average values.

Learn More

Heatmap

×



In addition to linear regression, we also used a heatmap to get an overall look into the distribution of each attribute in our dataset. We used seborn to create the heatmap looking at all 54 survey questions and average values.

The heatmap showed significant attributes of the dataset, specifically some of the more negative survey questions. The heatmap was a great way of looking at the dataset as a whole. It shows how much each feature affects our predictions, with higher numbers (darker colors) having the most effect.

Close

Lasso and Ridge

For the original Turkish dataset, we chose to do Lasso and Ridge linear regressions to reduce model complexity. Ridge puts constraints on the coefficients which in effect, shrinks the coefficient.

Learn More

Lasso and Ridge

×



For the original Turkish dataset, we chose to do Lasso and Ridge linear regressions to reduce model complexity. Ridge puts constraints on the coefficients which in effect, shrinks the coefficient. Lasso differs from Ridge in that it helps reduce over-fitting by taking into account magnitudes. Both regression methods find the first point where the elliptical hits the region of constraints. For more information on these analyses click here. When splitting our data into test and training sets, we got an R2 score of 92% indicating that our data is predictable and linear regression is a good fit. Both the Lasso and Ridge analyses yielded R2 scores above 80% and low mean squared error rates below 0.17 indicating our model explains a good amount of the variability around our means. These statistics suggest that we can produce predictions that are reasonably precise. Our lower MSE indicates we are close to finding the line of best fit for our model. It is no surprise that our R2 value decreased once running the analysis on our dataset as opposed to the training set.



Close

Classification

For this next model, a Random Forest was utilized to determine feature importance. This type provides a strong modeling technique that is better than using a single decision tree.

Learn More

Classification

×



For this next model, a Random Forest was utilized to determine feature importance. This type provides a strong modeling technique that is better than using a single decision tree. A con to a decision tree is overfitting which allows the model to learn from the “noise” in the training data which could ultimately have a negative impact on the model’s performance. Since the Random Forest model aggregates a multitude of decision trees, it limits overfitting as well as error due to bias which yielded more useful results. This analysis identified the feature importance (i.e., which of the 54 questions) carried the most weight in predicting outcomes (married or divorced). The six questions below were identified by the Random Forest model as top features in priority order. Our analysis identified a different set of priority features when compared to the top features identified in the Yontem study.



A confusion matrix analysis was conducted to visualize the performance of the Random Forest model and was determined to have an 89% average f1-score from our train test split data.







Close

KNN, LinearSVC, and SGDClassifier

Similar to the Yontem study, our team analyzed the accuracy of three machine learning models on the augmented data set. The models used were K-Nearest Neighbor (KNN), LinearSVC, and SGDClassifier.

Learn More

KNN, LinearSVC, and SGDClassifier

×

Similar to the Yontem study, our team analyzed the accuracy of three machine learning models on the augmented data set. The models used were K-Nearest Neighbor (KNN), LinearSVC, and SGDClassifier. Each model was run with and without feature selection (using the top six feature importance scores identified in the Random Forest model). The same analysis was run on data from the original study only.



Our analysis found the best performing model to be KNN without feature selection on the combined data set, while the Linear SVC was the top performer on the original data set (Turkey only data). Except for KNN, the models achieved better accuracy overall with the original data set. This suggests that cultural differences may be reflected in responses received from friends and family vs. responses from the original data.



The worst performing model was the SGD Classifier with feature selection for both data sets. In all cases, the performance of the models without feature selection was better when compared to the same model run with feature selection. This reflects relatively low feature importance scores from the Random Forest model. Shown below are the classification report and confusion matrix for the best and worst performing models on the combined data set.





Close

Statistical Analysis

Our team used numpy and Plotly to create interactive visuals that reflect the responses of the study at an individual level.

Learn More

Statistical Analysis

×

Our team used numpy and Plotly to create interactive visuals that reflect the responses of the study at an individual level. In the visual below, we have mapped aggregated responses every response at the individual level barring answers marked as neutral. You will find that the divorced population studied in Turkey have varying answers to all of the questions in the survey while the married population seems to be clustered towards answer specifically agreeable on most questions. However, the data does look supringsly different when we isolate the responses between the two groups for the 6 most important questions as identified by the Yotem study.

Link to Plotly interactive

With the isolated 6 questions below, there is a stark difference in the agreeable response rate between the married divorced populations. The researchers called out these six questions as being the strongest indicator of a divorce. The divorced population reflected positively on most questions while the married population barely agreed to any of the 6 questions.

Link to Plotly interactive 2

Close