Improving discriminative sequential learning with rare--but--important associations