(A) study on video content recommendation scheme based on topic using latent Dirichlet allocation

Abstract

In this thesis, a noble method for video content recommendation is proposed based on Latent Dirichlet Allocation (LDA), which is adequate for large and unstructured data. The description texts tagged in video contents are mainly utilized for extracting content topics. In spite of many advantages of Dirichlet prior in inference process, the Dirichlet distribution assumes that topic proportions are independent so the similarities among the topics are hard to be induced. Therefore using the per document (item) topic proportions, we suggest a way of representing the items as profile vectors. In this study, the text descriptions tagged on video contents are treated as items or documents. By doing so, we can compute the similarities of items based on topics so that the documents (items) can be semantically compared. From the experiments, we can conclude that the similarity between topics tends to decrease when the number of topics decreases. However, as a result, it may affect recommendation results as the number of topics changes. It has shown that our proposed method effectively works for item recommendation compared to those of YouTube in terms of the document (item) similarities.