Mouse Behavior Classification: 6-Model Comparison 3-CLASS BEHAVIOR

Social Behavior Classification • 5 Independent Runs with Different Random Seeds • 200 Epochs Training

697,003
Total Samples
76.62%
Best Accuracy (XGBoost)
76.22%
Best F1 (CNN)
0.14%
Best Stability (LightGBM)

Experiment Configuration Task: 3-class behavior classification | Data: 697,003 samples (Train: 487,902 / Val: 104,550 / Test: 104,551) | Features: 34 DLC features | Split: Stratified 70/15/15 with seeds 42-46 | Metrics: Mean ± SD across 5 runs

📊 Key Observations

XGBoost achieves the best accuracy (76.62%), followed by CNN (74.19%)
CNN outperforms LSTM significantly (74.19% vs 69.95%), suggesting spatial features are more important than temporal
LightGBM shows the best stability (CV = 0.19%), extremely consistent across runs
Deep learning models (CNN) now competitive with ensemble methods due to larger dataset
GMM performs very poorly (7.26%), unsupervised clustering fails on behavior classification

Figure 1. Overall Model Performance Comparison (3-Class Behavior)

Accuracy, Weighted F1, and Macro F1 scores across 6 classification models with error bars (±1 SD)

Figure 2. Model Ranking by Accuracy

Sorted by mean accuracy (5 independent runs)

Figure 3. Model Stability Analysis

Coefficient of Variation (CV%) - Lower indicates more stable performance

Figure 4. Per-Class F1 Score Heatmap

Estimated F1 scores for each behavior class across all models (3 classes)

Figure 5. Accuracy vs Weighted F1 Score Trade-off

Each point represents a model; bubble size indicates stability (smaller = more stable)

Table 1. Detailed Performance Metrics (3-Class Behavior Classification)

Mean ± standard deviation across 5 independent runs with different random splits