Multimodal Depression Recognition Using Facial Expression and Speech in Imbalanced Data Scenarios

In depression recognition research, data distribution is often imbalanced, and traditional methods exhibit insufficient classification accuracy for minority samples. To tackle this problem, this study proposes an improved fruit fly optimization algorithm with dynamically adjusted search stepsizes. It integrates this algorithm with AdaBoost to construct a novel classification method, named IFOABoost. Experiments show that on multiple datasets with different imbalance ratios (ranging from 3.7 to 14.2), IFOABoost outperforms classic methods such as CUSBoost, MEBoost, CatBoost, and Imb-XGBoost. Specifically, for male group, the AUC/G-mean/F-score improved by 2.5%/1.9%/2.3%, 5.7%/6.8%/6.1%, 4.4%/3.8%/3.5%, and 2.8%/2.3%/3.7%, respectively. For the female group, the improvements were 2.6%/2.0%/3.2%, 6.8%/6.2%/6.6%, 2.4%/2.3%/3.3%, and 4.4%/4.2%/4.5%, respectively. Ablation experiments further validate the superiority of IFOABoost. Additionally, IFOABoost has a short calculation time, enhancing the efficiency and performance of depression recognition.