Nonparametric Estimation of Search Query Patterns

In this poster, we adopted nonparametric regression as a method to identify the unique distribution of query log data collected from the Excite search service in May 2001. In Informetrics, parametric modeling has been widely used in tracing term frequency data, such as Zipf’s law, Lotka’s law, or Bradford’s law. However, these traditional parametric methods have had limited application when detecting distributions for large datasets with a nonlinear pattern and a long tail. This study tested kernel regression as an alternative tool to model nonlinearity of term frequency patterns. The results indicated that the kernel regression produced an improved model fit compared to previous parametric approaches in modeling query patterns.