Efficient Similarity Join Method Using Unsupervised Learning

This paper proposes an efficient similarity join method using unsupervised learning, when no labeled data is available. In our previous work, we showed that the performance of similarity join could improve when long string attributes, such as paper abstracts, movie summaries, product descriptions, and user feedback, are used under supervised learning, where a training set exists. In this work, we adopt using long string attributes during the similarity join under unsupervised learning. Along with its importance when no labeled data exists, unsupervised learning is used when no labeled data is available, it acts also as a quick preprocessing method for huge datasets. Here, we show that using long attributes during the unsupervised learning can further enhance the performance. Moreover, we provide an efficient dynamically expandable algorithm for databases with frequent transactions.