This paper describes a novel neuro-steered music source separation framework and conducts an extensive evaluation of the proposed system on MAD-EEG, a dataset composed of EEG recordings of subjects attending to a particular in duo and trio music excerpts. We propose an unsupervised non-negative matrix factorisation (NMF) variant, named Contrastive-NMF, that separates a target instrument, guided by the user's selective auditory attention to that instrument, which is tracked in his/her electroencephalographic (EEG) response to music. We analyse the impact of multiple aspects of the musical stimuli, such as the number and type of instruments in the mixture, the spatial rendering and the music genre, obtaining encouraging results, especially in difficult cases where non-informed models struggle. We believe that this unsupervised NMF variant is advantageous for neuro-steered music source separation as it allows us to incorporate additional information in a principled optimisation fashion and does not need training data, which is particularly difficult to acquire for applications involving EEG recording.