A neural-embedded discrete choice model: Learning taste representation with strengthened interpretability