Recruitview Eval

---
name: recruitview-eval
description: This benchmark evaluates multimodal models on predicting continuous personality traits and interview performance scores from video, audio, and text inputs. It probes the model's ability to perform fine-grained behavioral analysis and regression across psychometric targets. Use when the user wants to benchmark on RecruitView, or asks about evaluating this task. Reports Spearman's ρ.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2512.00450
  bibtex_key: [REDACTED]
  confidence: high
---

# recruitview-eval

> RecruitView: A Multimodal Dataset for Predicting Personality and Interview Performance for Human Resources Applications — Amit Kumar Gupta et al. (2025) (arXiv:2512.00450, 2025)

## What this evaluates

This benchmark evaluates multimodal models on predicting continuous personality traits and interview performance scores from video, audio, and text inputs. It probes the model's ability to perform fine-grained behavioral analysis and regression across psychometric targets.

## Datasets

- **RecruitView** — total 2011; splits: train (-1), val (-1), test (-1)

## Metrics

- `Spearman's ρ` **(primary)** — range: [-1, 1]
  - Measures the rank correlation between predicted and ground-truth continuous scores. Values range from -1 to 1, where 1 indicates perfect agreement in ranking.
- `Kendall's τ-b` — range: [-1, 1]
  - A rank correlation coefficient that measures the correspondence between two rankings, adjusting for ties.
- `Concordance Index (C-Index)` — range: [0, 1]
  - The probability that for a randomly selected pair of instances, the predicted scores preserve the correct ordering relative to the ground truth.
- `Pearson's r` — range: [-1, 1]
  - Measures the linear correlation between predicted and actual continuous values.
- `MSE` — range: [0, ∞)
  - Mean Squared Error, calculated as the average of the squared differences between predicted and actual values.

## Input / output format

**Input**: Multimodal input consisting of video clips (16 FPS, 16 frames, 224×224 resolution), audio (resampled to 16 kHz), and text transcripts (tokenized to max length 512).

**Output**: Continuous regression scores for 12 targets (Big Five personality traits, overall personality score, and interview performance metrics).

## Scoring recipe

```python
def compute_metrics(preds, golds):
    spearman = spearmanr(golds, preds).correlation
    kendall = kendalltau(golds, preds).correlation
    c_index = concordance_index(golds, preds)
    pearson = pearsonr(golds, preds).correlation
    mse = mean_squared_error(golds, preds)
    return {'spearman': spearman, 'kendall': kendall, 'c_index': c_index, 'pearson': pearson, 'mse': mse}
```

## Common pitfalls

- Metrics are computed per-target and then macro-averaged; reporting only the overall macro-average without per-target breakdowns obscures task-specific performance.
- Early stopping is based on validation Spearman correlation, not MSE or accuracy, which may lead to suboptimal convergence if models optimize for rank rather than absolute error.
- The dataset uses pairwise-derived continuous labels, so models must handle regression on potentially noisy or correlated psychometric targets.

## Evidence (verbatim from paper)

> We evaluate using Spearman's ρ, Kendall's τ-b, Concordance Index (C-Index), Pearson's r, and MSE. Metrics are computed per-target and macro-averaged for overall performance.

## Citation

```bibtex
@misc{gupta2025recruitview,
  title={RecruitView: A Multimodal Dataset for Predicting Personality and Interview Performance for Human Resources Applications},
  author={Amit Kumar Gupta et al. (2025)},
  year={2025},
  note={arXiv:2512.00450}
}
```

- arXiv: 2512.00450