Microservices architecture (MSA) is becoming the preferred choice for web applications due to its scalability and maintainability. However, troubleshooting microservices remains challenging due to the complex dependencies of microservices. Additionally, recent state-of-the-art methods rely heavily on labeled training data. To address these challenges, we propose HERO, a unified framework for both anomaly detection and root cause localization in MSA applications. HERO effectively captures complex dependencies of microservices by utilizing a heterogeneous graph neural network. Unlike existing machine learning approaches, HERO does not require root cause labeled data for training through a novel explainability technique. In our experiments on a widely used MSA benchmark dataset, HERO outperformed existing methods in both anomaly detection and root cause localization tasks.