Web-log Session Clusters

Web-portals and other e-commerce sites often segment their visitors to provide better personalized services. When a web-page is requested, the server log records the user's IP address, the URL retrieved, access time, etc. These logs can be analyzed to segment visitors based on their `cow-path' or trajectory through the website, as described by the sequence of pages visited, page contents, time spent on each page, etc. In a recent work, the use of a weighted Longest Common Subsequence (LCS) [BG01] similarity metric to describe how similar two trajectories are was suggested. This metric determines the LCS of the two trajectories, and then scales it by what fraction of the total visit time is spent in the longest common subsequence. Alternatively, one can use a vector-space model, wherein entries in the data matrix $\mathbf{X}$ indicate time spent in a particular session (column) on a particular page (row). In this subsection, we present results of OPOSSUM and CLUSION for the data presented in [BG01]. We randomly selected 3000 sessions (out of 23310) from a community portal, http://www.sulekha.com/. The index / root page of the web-portal was removed since it was visited by almost everyone for a considerable amount of time and, hence, provided no discriminatory information. Figure 3.7 compares results for a vector-space-based approach using cosine similarity with LCS. The cosine measure shows some large dark diagonal regions indicating compact clusters of sessions, but it turns out that these clusters are sessions where the majority of the time was spent on a category index page (level 2 on the portal's site map). The LCS is able to capture a larger percentage of the total similarity (amount of `grayness') in the diagonal regions, showing a better and more balanced grouping. Such visualization can be used to select the appropriate similarity measure for a given clustering objective, and to evaluate the overall clustering quality. For example, CLUSION shows that clustering visitors into 20 groups was successful despite the extreme sparsity ($\sim$1%) in figure 3.7(b).

Figure 3.7: Web-log session clustering using a vector space model and cosine similarity (a) versus using weighted longest common sub-sequence similarity (b). The cosine similarity is far less sparse and is dominated by major category index pages, while the LCS shows better isolation among the clusters.

(a)	(b)

We also used value-balanced OPOSSUM to cluster web-log sessions which yields clusters with comparable total web-surfer exposure time. These clusters might be particularly useful for new formats in target advertising campaigns. It simplifies advertising campaign management by enabling the portal to offer fixed prizes for ad delivery exposure to each cluster since they represent comparable attention times.