Principal Component Analysis of User Association Patterns in Wireless LAN Traces

Wireless networks have gained its popularity quickly in recent years. As the usage increases, there is also an increasing need to understand the characteristics of wireless users. Among all the properties to describe user behaviors in 802.11-based wireless LANs, their association patterns to access points (APs) play a very important and fundamental part. In this poster, we apply principal component analysis (PCA) to unearth the common pattern of user association in wireless networks. The questions we seek to answer by applying the PCA technique are: (1) Are users similar to one another in their association pattern in long run? (2) Does individual user show consistent daily association pattern across multiple days? (3) If the answer to question 2 is yes, then how do we find some summarized presentation of the daily association pattern of a user? (4) Can users be grouped using the summarized presentation obtained in question 3, leading to groups that show similar association pattern? Throughout the analyses we find that for university campuses, the whole user population is diverse enough that the major common trends of association, even if it may exist, is fairly insignificant. This observation is consistent from the traces we studied about generic users. However, if we focus on a user group in which the individual users have some common attributes, the common trends in association patterns become much stronger for the group. We further investigate the individual user association pattern across days and find most users show a clear consistent trend in its daily association patterns. The principal components (PCs) of the daily individual association data set can be used to characterize individual users and summarize their association behaviors. Based on the principal components of individual user association matrices, we further propose a way to group users. We define the similarity index between two users by performing a weighted sum of inner products of the PCs from each of the user pairs. Using this definition, we provide a method to distinguish users that display similar association patterns from the others, and identify such users as a sub-group from the whole population. We show that by grouping users based on the similarity index, the members of sub-groups have much more significant common trends in the association patterns than randomly generated group, or the whole user population. In this paper we use three WLAN traces collected from university campuses, including University of Southern California (USC) [2], Dartmouth College [4], and University of California at San Diego (UCSD) [3]. The USC and Dartmouth trace are collected from all types of wireless devices on campus. UCSD trace is from a specific project targeting at PDA users. We select to analyze the traces for one whole semester/quarter from the studied universities, including fall quarter 2002 for UCSD, spring quarter 2004 for Dartmouth, and summer semester 2005 for USC. II. MATRIX REPRESENTATION OF USER ASSOCIATION PATTERNS