Active Congruency-Based Reranking

genre, abstract expressionism. 8 Ben Shalom et al. Active Congruency-Based Reranking Frontiers in Digital Humanities | www.frontiersin.org August 2016 | Volume 3 | Article 7 8. DiscUssiOn Our method combines two different approaches – similarity betterment by graphical models and Spectral Clustering Co-occurrence Stability based on spectral analysis. The experiments demonstrate that the contributions of these two steps, which both tap into group congruency, albeit using very different approaches, partly overlap. It is worth noting that the graphical model suggested in Pellegrini et al. (2010) and Wolf et al. (2011b), which partially resembles our similarity betterment method, was not designed to be query specific, and, therefore, it does not consistently improve retrieval results. The main differences between the suggested model and the previous ones are summarized in Table 2. The SCCS procedure, while highly effective on the Genizah dataset, was much less effective on the Art dataset. We hypothesize that this stems from the lack of transitivity violation triplets in the Art data. In addition, the features used in the art dataset capture only a fraction of the visual information in a painting. The Graph-pageRank and Graph-density methods of Zhang et al. (2012), which were designed primarily to combine multiple similarities together, are not competitive in the context of our experiments. However, they did extremely well (in the original paper) when combining local and holistic features. Note that the nature of the experiment is different, since previous work focused on the fusion of ranking from various sources, while we deal with a single, very noisy, similarity matrix. Relevance feedback can provide a significant amount of actionable information. While in casual web search, it may not be realistic to expect the user to mark relevancy, collecting results is an inherent part of scholarly research of the type considered here. In this work, we employ such information to alter the graph structure of our problem. It might be beneficial to employ such information even at an earlier stage and alter the underlying pairwise similarities. One limitation of our technique is the usage of multiple parameters. Due to computational reasons, it is infeasible to optimize these parameters in an exhaustive way. It is likely that results could improve further by discovering better parameter values. Note, however, that the parameters are intuitive and interpretable: the transitivity potentials χ(lqi, lqj, lij) reinforces faint connections when appropriate, by assigning a high probability to the clique (1,1,1); we offer (Section 3.5) a unique way to calibrate the trade-off parameters cij and cα; the meaning of the trade-off parameter β is given in Section 3.3; the parameters of SCCS and the relevance feedback are mostly discrete and have clear meanings. Our method is partly motivated by join finding based on handwriting similarity in the Cairo Genizah. Digital Paleography holds the promise of scalability: it allows processing of sizable collections of images, and even more practically at the moment, the comparison of all small subsets of such collections, thereby finding links that were previously unknown. As the Genizah visual search engine is making its online debut as one of the