Pmemo Emotion Recognition Eval

---
name: pmemo-emotion-recognition-eval
description: Binary classification of user-independent emotional states (valence and arousal) from electrodermal activity (EDA) signals. It probes the model's ability to generalize across subjects by using subject-specific thresholds and fusing physiological signals with external music benchmarks. Use when the user wants to benchmark on PMEmo, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1908.03692
  bibtex_key: [REDACTED]
  confidence: high
---

# pmemo-emotion-recognition-eval

> User independent Emotion Recognition with Residual Signal-Image Network — Yin et al. (2019) (arXiv:1908.03692, 2019)

## What this evaluates

Binary classification of user-independent emotional states (valence and arousal) from electrodermal activity (EDA) signals. It probes the model's ability to generalize across subjects by using subject-specific thresholds and fusing physiological signals with external music benchmarks.

## Datasets

- **PMEmo** — total 7962; splits: test (-1)

## Metrics

- `accuracy` **(primary)** — range: percent
  - Percentage of correctly classified instances out of the total number of instances. Computed as (TP+TN) / (TP+TN+FP+FN).
- `F1-score` — range: percent
  - Harmonic mean of precision and recall: 2 * (precision * recall) / (precision + recall).
- `precision` — range: percent
  - Ratio of true positive predictions to the total number of positive predictions: TP / (TP + FP).
- `recall` — range: percent
  - Ratio of true positive predictions to the total number of actual positives: TP / (TP + FN).

## Input / output format

**Input**: Continuous EDA signals (phasic, tonic, or mixed) transformed into image-like representations via cvxEDA decomposition, optionally fused with external music features.

**Output**: Binary classification label (0 or 1) representing low or high emotion for valence or arousal dimensions.

## Scoring recipe

```python
def compute_metrics(y_true, y_pred):
    tp = np.sum((y_true == 1) & (y_pred == 1))
    tn = np.sum((y_true == 0) & (y_pred == 0))
    fp = np.sum((y_true == 0) & (y_pred == 1))
    fn = np.sum((y_true == 1) & (y_pred == 0))
    accuracy = (tp + tn) / (tp + tn + fp + fn)
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
    return accuracy, f1, precision, recall
```

## Common pitfalls

- The evaluation converts continuous V/A regression to binary classification using subject-specific k-means thresholds; applying a fixed global threshold will misalign with the protocol.
- Results are averaged over 10-fold cross-validation; reporting single-split or subject-dependent splits will not match the paper's reported metrics.
- Music features are explicitly noted as subjective and unreliable for user-independent tasks; fusing them without accounting for subject specificity may artificially inflate accuracy.

## Evidence (verbatim from paper)

> As we convert emotion recognition to a binary classification, average classification accuracy, F1-score, precision and recall are adopted as the classifying evaluation criteria. In the pre-experiment of correlation analysing, Root Mean Square Error (RMSE) and Pearson Correlation Coefficient (r) are adopted according to the baseline in [[13]]. We applied 10-fold cross-validation method. After tests of 10 flods were finished, the evaluation indexes were averaged at the end.

## Citation

```bibtex
@misc{yin2019userindependent,
  title={User independent Emotion Recognition with Residual Signal-Image Network},
  author={Yin et al. (2019)},
  year={2019},
  note={arXiv:1908.03692}
}
```

- arXiv: 1908.03692