TY - JOUR
T1 - A Covariate-Guided Graph Attention Network for Depression Recognition via Pure Facial Movements
AU - Zhao, Bohua
AU - Liu, Zhenyu
AU - Zhang, Haibo
AU - Yuan, Jiaqian
AU - Ding, Yi
AU - Hu, Bin
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026/6/1
Y1 - 2026/6/1
N2 - Depression recognition based on facial movements has garnered widespread attention in recent years. One key challenge faced by existing studies is that individual heterogeneity in facial movements severely affects model performance. This paper, for the first time, proposes an innovative solution targeting two critical aspects: data acquisition and model design. In data acquisition, we replace the traditional interview task with a video-watching paradigm to eliminate mouth movements induced by verbal expression to obtain pure facial movement flows. In model design, we treat facial action units (AUs) as covariates to align discriminative depression cues to mitigate the impact of individual heterogeneity. Specifically, our proposed FDSNet statistically selects AUs showing significant differences between depressed and healthy groups and combines the AUs-based scoring module and graph attention network to guide the model in dynamically focusing on discriminative key video segments. Experimental results show that FDSNet achieves superior accuracy and generalization performance compared to SOTA methods. Our work suggests that controlling data acquisition to ensure pure facial movement data, along with employing a personalized modeling strategy, are both critical to mitigating individual heterogeneity and enhancing model performance.
AB - Depression recognition based on facial movements has garnered widespread attention in recent years. One key challenge faced by existing studies is that individual heterogeneity in facial movements severely affects model performance. This paper, for the first time, proposes an innovative solution targeting two critical aspects: data acquisition and model design. In data acquisition, we replace the traditional interview task with a video-watching paradigm to eliminate mouth movements induced by verbal expression to obtain pure facial movement flows. In model design, we treat facial action units (AUs) as covariates to align discriminative depression cues to mitigate the impact of individual heterogeneity. Specifically, our proposed FDSNet statistically selects AUs showing significant differences between depressed and healthy groups and combines the AUs-based scoring module and graph attention network to guide the model in dynamically focusing on discriminative key video segments. Experimental results show that FDSNet achieves superior accuracy and generalization performance compared to SOTA methods. Our work suggests that controlling data acquisition to ensure pure facial movement data, along with employing a personalized modeling strategy, are both critical to mitigating individual heterogeneity and enhancing model performance.
KW - Depression recognition
KW - affective computing
KW - facial movements
KW - graph attention network
KW - individual heterogeneity
UR - https://www.scopus.com/pages/publications/105034082753
U2 - 10.1109/TCSVT.2026.3661206
DO - 10.1109/TCSVT.2026.3661206
M3 - Article
AN - SCOPUS:105034082753
SN - 1051-8215
VL - 36
SP - 8181
EP - 8192
JO - IEEE Transactions on Circuits and Systems for Video Technology
JF - IEEE Transactions on Circuits and Systems for Video Technology
IS - 6
ER -