ADS Capstone Chronicles Revised

29

For individual food recommendations, classification models displayed a greater range of performance due to the increased variability in the dataset. The XGBoost Classifier once again emerged as the top-performing model, achieving a precision score of 0.989. This demonstrates its ability to adapt to the broader range of nutritional values and patient scores present in individual food data. The Random Forest model followed closely with a precision of 0.980, further highlighting the strength of ensemble methods in handling complex datasets. Neural Networks also performed well, achieving a precision of 0.982, while the Support Vector Classifier, which was included specifically for the individual food dataset due to its smaller size, achieved a respectable precision of 0.945. Logistic Regression, although still a strong baseline, showed its limitations in capturing the variability of individual food data with a precision of 0.922. The combined results, presented in Table 6, reveal that ensemble methods like XGBoost and Random Forest consistently outperformed other models across both datasets. This performance can be attributed to their ability to model complex relationships and interactions within the data. Neural Networks also demonstrated strong performance, particularly on menu data, where the limited variability allowed the model to achieve near-perfect scores. The inclusion of Support Vector Classifier for the individual food dataset added value, as its performance highlights its effectiveness for smaller, less complex datasets.

5.2 Evaluation of Results of Classification Models - Classifying Recommendation

Classification models were used to predict whether menu items or individual food items are “recommended” or “not recommended” for diabetic patients. These predictions were based on the binary target variable Recommendation, derived from the patient score. Evaluation metrics included accuracy, precision, recall, and F1 score, with precision emphasized as the most critical metric due to the need to minimize false positives. Incorrectly recommending a food item could result in serious health risks for diabetic patients, making this a priority in the evaluation process. 5.2.1 Combined Results: Menu and Individual Food Data For menu recommendations, the XGBoost Classifier consistently achieved the best performance, with a precision of 0.994 and nearly perfect metrics across all evaluation categories. Ensemble methods like Random Forest and Neural Networks also performed exceptionally well, achieving precision scores of 0.987 and 0.992, respectively. The Logistic Regression model, serving as a baseline, recorded a lower precision score of 0.894, indicating its limited capacity to handle the nuanced relationships within the dataset. The similar performance of ensemble methods and neural networks on menu data reflects the lack of variability in menu items, as many items from the same restaurant or similar establishments often share common nutritional characteristics.

231