Two-step Method of Searching Fuzzy Duplicates Using Elements of Graph Theory

This study aims at tackling the problem of searching Fuzzy Duplicates among users of internet-service. In this work I proposed two-step method: first there is a search for candidate-users, and then there is the scoring of users. The first stage was made with such data structure as direct and inverted indexes. The second stage consist of building of the graph from found users. In the graph nodes is users and edges is features by which users were found in the index structures. Further the problem was transformed to binary classification problem for pairs "base user"-"candidate user". Features used in the work were both user-content features and graph-based features. Also graph-based features can be divided on full-graph features and pair-specific features. In the experimental part of the study range of classification algorithms have been tested. The time of work and the quality have been estimated for them. Main advances of the proposed approach are working in on-line, permanent adding of new data and interpretability of the method. Key words: Machine Learning, Classification, Fuzzy Duplicates Detection, Graph Theory.