Spatio-temporal Pseudo Relevance Feedback for Large-Scale and Heterogeneous Scientific Repositories

As larger and larger amounts of data are harvested, finding just the right piece of information out of this noisy and heterogeneous ocean of data remains challenging. Many widely adopted scientific data search engines continue to be mainly based on text semantics. However, it is not uncommon in scientific big data applications to face collected data that do not possess text information. In this scenario, search engines fail to retrieve potentially relevant data. For instance, even though Pangaea, a digital data library and a publisher for earth system science, contains more than 400,000 datasets, more than 98% lack sufficient text information. In this work, we propose a novel pseudo relevance feedback method based on spatio-temporal and text (STT) information for scientific big data: STT-PRF. Although STT-PRF may simultaneously use STT information, we show that the missing values in space, time or/and the text are handled efficiently. STT-PRF is especially robust even without text information. We tested our STT-PRF method using the Pangaea repository on our Cross-DB Search Platform, which is a search engine for scientific big data based on various latent correlations. Experimental evaluations on such standard metrics as nDCG and Precision/Recall show that STT-PRF outperforms the standard baseline methods.