Workout Calorie Prediction

Can calories burned in a workout be predicted from heart rate, session and body measurements, and which kind of model does it best? For our machine learning course at Thompson Rivers University, Gursahib Singh and I built a seven-stage Python pipeline that covers the whole syllabus: data quality, exploratory analysis, PCA, decision trees, ensembles, regularisation, SVMs, neural networks, classification, clustering and cross-validation.

The Kaggle dataset has 20,000 sessions and 54 columns with no missing values. HIIT sessions averaged 1,653 calories against 897 for yoga, and an ANOVA confirmed the workout-type effect (p < 0.001) while diet type had none (p = 0.36). Session duration had the strongest correlation with calories burned (0.81).

PCA over 41 numeric features needed 21 components for 95% of the variance: the first captured body composition and the second heart-rate intensity. The scatter of duration against calories falls on perfectly straight lines, a sign the dataset is synthetic, which the report notes as a limitation.

We compared six ensembles (Random Forest, ExtraTrees, AdaBoost, Gradient Boosting, stacking and voting), Lasso, Ridge and ElasticNet, SVR with three kernels and four neural network sizes, then checked the strongest with 5-fold cross-validation and a learning curve. K-Means grouped sessions into three loose personas (silhouette about 0.19).

As submitted, Gradient Boosting won with test R² 0.997 and RMSE 27.7 calories, and 0.989 in cross-validation.

Plain Ridge and Lasso scored a perfect 1.000. A linear model fitting exactly means some input already contains the answer. The culprit was cal_balance, a column that ships with the dataset and equals calorie intake minus calories burned, checked on all 20,000 rows. With intake also in the inputs, the target was a simple subtraction away.

We had excluded three other leaky columns, but missed this one. Rerunning the same Gradient Boosting setup without cal_balance drops test R² to 0.65. Adding back workout type, which the pipeline had silently dropped as a text column, lifts it to 0.9999 on this synthetic data: the honest finding is that duration, workout type and experience level almost fully determine the target here, which says little about real wearable data.

The broad pipeline was the assignment; the leak was the lesson.