Recent research introduces deep learning algorithms such as recurrent neural networks (RNNs) to predict the next activity, one of the most challenging tasks in predictive business process monitoring. However, the RNN-based models use only the last hidden state as a context vector, resulting in the loss of significant historical information, particularly in long sequences. Furthermore, many previous approaches rely primarily on the activities and timestamps of events, disregarding other activities and failing to capture the event log’s multi-view. To address these issues, we propose a novel method for predicting the next activity that combines a transformer network and a multi-view representation of the event log. By adding multi-view information from all attributes recorded in the event log, we hope to increase predictive accuracy. The proposed method captures long-term dependencies between the different views and fuses the multi-view information using the multi-head self-attention mechanism to predict the next activity. Experimental results on six real datasets show the effectiveness of the proposed approach compared to state-of-the-art approaches. Source codes of this paper are available on Github1.