Dataset Overview

Dataset Name

• Name: sports_activities_dataset
• Source: UCI ML repository
• Link: Daily and Sports Activities

Dataset Files

• Main file: csir_cdri_test.ipynb
• Auxiliary files: prediction-pca.ipynb, pytorch-model.ipynb

Data Preprocessing

• Input: Patient activity data with 5‑second windows, containing 125 observations, each with 45 features.
• Processing steps:
  • Step 1: 225 features, including min, max, mean, skewness, and kurtosis of five units for each of the 9 sensor axes.
  • Step 2: 225 features representing the top 5 DFT peak magnitudes of five units for each of the 9 sensor axes.
  • Step 3: 225 features indicating the frequencies corresponding to the peaks from Step 2.
  • Step 4: 495 features representing 11 selected autocorrelation values of the time series.
• Output: 1,170 features; each feature file is normalized to the range [0,1] and includes patient ID and activity ID.

Dataset Applications

• Model testing: Includes two parts, one using a 9,120 × 1,170 matrix, the other using a 9,120 × 30 matrix.
• PCA application: PCA performed on the original matrix, but did not achieve the expected results.

Model Performance

Actual dataset (9,120 × 1,172)

• Activity prediction:
  • Best model: Gradient Boosting Classifier, Accuracy: 0.9368
  • Other models: Bagging Classifier, Random Forest Classifier, ExtraTrees Classifier, Decision Tree
• Patient and activity prediction:
  • Best model: Bagging Classifier, Accuracy: 0.8245
  • Other models: Gradient Boosting Classifier, Random Forest Classifier, Decision Tree, kNN (k=3)

PCA dataset (9,120 × 32)

• Activity prediction:
  • Best model: ExtraTrees Classifier, Accuracy: 0.8767
  • Other models: Gradient Boosting Classifier, Random Forest Classifier, Bagging Classifier, Neural Networks (DNN)
• Neural Network (PyTorch):
  • Best model: With Adam Optimizer, Accuracy: 0.8951
  • Other models: With Adam Optimizer + Karpathy constant, With RMSProp optimizer, With SGD Optimizer
• Neural Network (scikit‑learn):
  • Best model: MLP with Adam + ReLU, Accuracy: 0.8065
  • Other models: MLP with Adam + Sigmoid

Additional task (PyTorch on PCA dataset)

• Model: With RMSProp optimizer, Accuracy: 0.5043