Schematic representation of CANSTAR (bottom), a convolutional attention network designed to predict the stellar activity contribution to line shifts, κ(t), from the time series of distortion coefficients, g_i(t). The input coefficients are treated as independent channels and processed through a series of one-dimensional convolutional layers that extract local temporal features. The convolutional outputs are then passed to a transformer encoder module, which consists of a multi-head self-attention part (top) with different ‘heads’, where each one of them focuses on different temporal dependencies within the data, and their outputs are combined, normalised, and propagated through additional attention layers. Finally, the resulting features are flattened and projected through a linear layer to produce the predicted κ(t) time series.