Article-Goods Associative Search using Bi-Source Topic Modeling Method
Kim Byounghee, Bado Lee, Seong Jong Ha, Nam-Ik Cho, Byoung‐Tak Zhang · 2011
With the progress of digital convergence, multimodal data is generated in torrents. User-centric retrieval and recommendation services in this environment demand methods for multimodal information retrieval and associative analysis. In this paper, an associativity modeling method is presented for datasets from various sources and results are shown in online article-goods associative search just based on images. The model is named as a Bi-Source Topic Model (BSTM), which is an extension of LDA (latent Dirichlet allocation). An image dataset is constructed with pictures in Korean magazines and an online shopping mall. With BSTM, we can quantify associativities between images from magazine and mall based on the similarity of topic proportions in images. With a testset of goods images for evaluation, it is shown that proposed method results in about 60% success rate based on category information of goods. Given article pictures as queries, various goods are retrieved which contain interesting semantic relations.