Aiming at main challenges of Web mining and personalized service currently, basic K-Means algorithm of clustering techniques was researched, including algorithm flow and limitations. To solve shortcomings of pre-determining cluster number, heavily dependent on initial center selection and particularly sensitive to noise as well as edge data in basic K-Means algorithm, improved density-based adaptive K-Means algorithm was presented. It conducts steps of initial classification and K means iterative to reduce impact of above problems and improve clustering quality. Experiments on Web log clustering also verified its effectiveness.