Joint Prediction of Individual Activity Location and Start Time Based on Mobile Phone Signalling Data via Multi‐Head Self‐Attention

Activity location and start time are crucial and challenging to predict in activity‐based models (ABMs). For most individuals, activity location and start time are intrinsically correlated. However, current ABMs usually predict them using cascaded or independent architectures. Cascaded approaches are susceptible to error propagation, while independent structures often fail to capture task correlations due to information isolation. Consequently, these methods cannot effectively reflect the spatiotemporal constraints and collaborative nature of decision‐making. To address these limitations, we propose SALT‐JP, a joint prediction model based on a multi‐head self‐attention mechanism. It uses a hard parameter sharing multi‐task learning framework to collaboratively predict the next activity location and start time. The framework features a Transformer‐based shared encoder, which effectively integrates diverse input features and captures long‐range sequential dependencies. Furthermore, SALT‐JP can predict the previously unvisited locations by introducing a dynamic candidate set generation mechanism. Experimental results showed that SALT‐JP significantly outperforms baseline models such as LSTM and DeepMove. It achieves a Top‐1 location accuracy of 72.3% and reduces the start time prediction MAE to 28.56 min. These results validate the efficacy of jointly modelling location and time via shared self‐attention representations. The proposed method offers a promising solution for individual mobility modelling.