Traditional depression detection has relied on content-based approaches that analyze linguistic features, but fail to capture the complex relationships of users on social media and their temporal patterns. Thus, we propose a semantic ego-network graph neural network that combines multiple similarity components and constructs the network using a combined similarity metric. Our methodology combines three similarity metrics: linguistic similarity using LIWC features and BERT sentence embeddings (α = 0. 4), temporal similarity using posting patterns and circadian rhythms (β = 0. 3), and psychological similarity modelling expressions of mental health disorder symptoms (γ = 0. 3). The semantic ego-networks are used within a graph attention network that learns from user relationships. We evaluate our model against a baseline RoBERTa model using a total of 10000 posts from depression and control groups, sampled from the Reddit Mental Health Dataset (RMHD). Experimental results suggest that our proposed semantic ego-network model provides superior performance compared to baseline content-based models, achieving higher recall and F1 scores. The ablation studies show that a combined similarity approach provides better performance compared to a single component approach. Our model successfully captures indicators of depression using semantic user relationships that traditional content-based models miss. Our findings suggest that semantic ego-networks using combined similarity of multiple similarity components offers significant advancement in early detection of depression. Our similarity approach offers a foundation for development of mental health early detection.