We analyse team dynamics in the popular e-sports game DotA2 using an approach that combines convolutional and LSTM networks with a feature-temporal attention mechanism. Our goal is to identify strategic behaviours that lead to successful goals, such as scoring kills, during World Championship matches. Each team’s formation is represented by a polygon, which feeds an RNN that learns kill events from this polygon under its area, diameter, and moments around the centroid. By exploiting the attention mechanism, our network highlights the most relevant features at each time step, providing insights into strategic team movements and formations. Our results demonstrate the effectiveness of our approach in capturing critical dynamics that influence the outcome of engagements in DotA2.