Learning with Partial Multi-Outlooks

How to deal with data that abounds in heterogeneous applications, known as multi-outlook learning, is of imperative importance to the creation of general-purpose and flexible intelligent systems. Prior works envision that each data instance appears in all outlooks, however, in practice, it is often the case that every outlook suffers from information incompleteness due to the data availability issue (e.g., privacy concerns). In such a case, existing learning models tend to be fooled by the ambiguous semantics conveyed by the missing outlooks. To fill the gap, we in this paper propose a new learning paradigm, named Generative Outlook Reproducing via Repository (GORR), which draws insight from the human analogy of capturing the commonalities among outlooks to reconstruct the missing outlooks. Specifically, GORR leverages the feature relatedness across outlooks to construct an outlook repository. The instances, once being projected onto the outlook repository, would have complete feature representations, where the missing outlooks are generated from the observed ones. Learner trained on the outlook repository then enjoys a complete feature information and thus is capable to perform accurate predictions. Extensive experiments are carried out on both synthetic and real data sets, demonstrating the effectiveness of GORR.