To extract an effective feature in prediction of subjective impressions from single-trial neurophysiological recordings, the spatial filter that extracts brain activities related to impressions were constructed using the common spatial pattern (CSP). We focus on subjective preference induced by chords composed of 3 notes with different frequency ratio. Magnetic cortical activities while hearing chords and comparative judgment on pair of them were measured. The predictive model that predicts the scale value of preference was trained using the CSP-based feature for each participant. The result of the evaluation experiment shows that the CSP-based feature improved the mean prediction accuracy in all participants, compared with the other features without spatially filtering. Furthermore, the capability of construction of a spatial filter that extracts cortical activities varying with degree of preference using the comparative judgments was indicated.