Multiple instance learning (MIL) has shown great potential in addressing weakly supervised problems in which class labels are provided for sets (bags) of instances. The main challenge in MIL comes from the lack of knowledge on the pertinence of each individual instance in class discrimination. In this paper we propose TensMIL2, a generic unsupervised feature extraction procedure based on non-negative PARAFAC (CP) decomposition, combined with instance selection and MIL classification, that is efficient also for partially observed datasets. Evaluation of our algorithm in standard MIL benchmark datasets showed that TensMIL2 is performing better than state-of-the-art algorithms in most of the cases. Moreover, the comparison of the proposed feature representation via CP decomposition to the previously used features, showed an increase in performance in most of the cases, in both full and partially observed (90% missing values) datasets.