An Improved LSTM based Method for User Behavior Classification in VR Scenes

Time series classification is a common machine learning problem. With the rapid growth of the virtual reality (VR) industry, the number of users participating in VR is increasing rapidly, and analyzing user behavior in VR scenes is becoming increasingly important. Compared to general time series classification tasks, the input sources in VR are more complex and diverse, and traditional time series classification models struggle to handle this complexity. In this study, we propose a multi-source time series classification network that utilizes an attention mechanism to combine multiple input sources such as left and right controllers, head-mounted displays, and buttons. By focusing on key features using an attention mechanism, we can provide more accurate information and improve the overall performance. Our approach achieved high accuracy in user action classification tasks in VR scenes.