Visualizing and Analyzing Web Browser History Data
using Eureka and SpotFire

CMSC 838B: Information Visualization Application Project

Matthias Mayer
Research Visitor at HCIL
mayer@informatik.uni-hamburg.de
March 6th, 2001

Motivation

About 80% of all accessed web pages have been visited befor by the same user (Cockburn & McKenzie 2001). Revisiting these pages, however, is no easy task. Today Back Buttons and Favorites are used to revisit pages in current or previous sessions. These mechanisms could be enhanced using better and more adequate interfaces. The patterns with which users access web pages should help to find how these new interfaces should look like.

This small project uses two commercial data visualization tools in order to find patterns in my own web browsing behaviour. Therefore I use a Netscape history file which documents about 7 months of my browsing activities. Thus, Visualization here is used to reveal characteristics after browsing has already taken place - not to support the user during his sessions.

In my PhD project I develop a tool to visually support web users during their sessions by showing their history. More about these 'Browsing Icons' can be found on the web (Mayer 2001).

Questions

What Questions must be answered in order to develop better tools for revisiting web pages?
Some questions are:

• How often are pages revisited and in what intervals?
• Which pages are visited often?
• Can we charachterize sessions? E.g. how often is just the last page important to the user?

Data

• I used a Netscape Communicator 4.7 history file, produced on my laptop under Windows98.
• Data cover 7 months: Oct. 16th 1999 - May 15th 2000.
• 5096 web pages were accessed (rows in the file).

I wrote a simple Java application to convert the Netscape file in input to Eureka and SpotFire. This step includes the calculation of derived data such as duration between first and last visit. Each row in the file contains information about one URL:

index Simple linecount URL the URL of the page Title the title of the page Kind I assigned each etry one of the following kinds: file (on my laptop) server homepage query comp. sci. department site our lab's website secure accesses unknown Host the host part of the URL FirstVisit date fo the first visit FirstDaytime time in the day of the first visit LastVisit date fo the last visit LastDaytime time in the day of the last visit DiffFirstLast 1 difference between first and last visit in milliseconds DiffFirstLast 2 difference between first and last visit in days, hours, minutes numOfVisits number of visits to this page

Visualization Tools

I used InXight Eureka 1.1 (www.inxight.com, formerly Table Lens) and SpotFire 5.1 (www.spotfire.com).

Restrictions

There are several major restrictions due to the kind of data:

• The date lack certain important information because Netscape doesn't store them:
  each single visits date and duration,
  widget by which page was loaded,
  window in which it was displayed, etc.
• just one person's data, thus highly biased.
• I used another machine in my office as well, so data are missing.
• I sometimes changed the time on my computer, so there are errors in the time fields.
• These restrictions show the necessity for a well planned study with modified browsers and a larger number of users. This small project just tries to find out to what extent the tools would support the analysis of a better data set.

Selected Visualizations

Fig. 1a Eureka

Fig. 1b SpotFire

Getting rid of errors: An accidentally inserted outlier was obvious in both tools at first sight (one data point was dated early 2001).

Fig. 2

Fig. 2 shows the data first sorted by kind, then by number of visits. Findings:

• More than half of all web pages were just visited once (see right column).
• Few pages show extremely high numbers of visits. Maximum: 456, then 86, 80, 74, ..., decreasing logarithmically).
• All 'kinds' show this pattern.
• The negative difference between the visit dates point at an error in the data: I sometimes changed the date of my machine in order to circumvent expiration dates...

Fig. 3

Fig 3. shows the data sorted by 'number of visits' and then by 'kind'. Findings:

• All kinds show both less and more frequented pages (compare two rightmost columns).
• Our institute's site (third 'kind' from bottom) shows the highest percentage of large number of visits.
• Just a few of the pages that were visited often were also visited over a long period of time (compare orange line and right column).

Fig. 4

Fig. 4 shows a SpotFire Scatterplot. Mapping: x - number of visits, y - difference between visits (in 'artificial' dates beginning at 1.1.1970), size - number of visits, color - kind, markers are jittered. The homepage would be further on the right. Findings:

• As in Eureka it was obvious that there were just a few pages which were visited very often or over a long period of time.
• Counting single objects or looking at content was much easier than in Eureka.
• It was harder to estimat proportions due to the large number of hidden items.

Fig. 5

Fig. 5 shows a SpotFire Histogram View. Mapping: x - host, y - sum of entries with this host. Findings:

• Queries are counted as different pages, thus the high green and black columns. They should be eliminated to explore the rest.
• A lot of different pages were visited on java.sun.com and our lab's site.
• Good: clustering of items (Eureka just provides grouping, but still keeps single items).

Fig. 6

Fig. 6 shows a SpotFire profile chart (a parallel coordinates visualization developed by Inselberg). Mapping: Six parallel axes show values of six attributes, each in it's own scale. Entities are connected by lines.
I filtered the view to show just items that were visited more than 20 times and over a period of at least five days. Findings:

• Again, the broker site shows the most hits. (Well, fall 1999...)
• cybergeography.com was visited over the longest period of time.

Fig.7

Fig. 7 shows a SpotFire 3D ScatterPlot. Mapping: right to left - date of first visit, bottom to top - number of visits, depth - hosts. Number of visits was redundantly coded to size, kind to color. Findings:

• I liked the spatial feeling in this visualization.
• My march vacation was visible as large valley (in fact, i was in 'Valle Gran Rey').
• Top bubbles represent again the broker pages.
• The cluster of black pages on the left in the valley are visits to an online dictionary. When I came home from vacation I wrote an English paper.

Conclusion

• Both tools provide good overviews.
• They are good to verify or reject expected patterns and hypotheses.
• Intimate knowledge of the domain and dataset were important for interpretation.
• More insights could be found if we had richer data.
• None of the tools made it easy to get the following numbers: I visited 5098 pages and made 12179 visits. This is an average of 24 pages and 58 visits per day (Cockburn & McKenzie 2001 found the average visits to be 42 per day).

Critique of Tools

Eureka

• + easy to learn.
• + One view shows overview, proportions, distributions, correllations and even details on demand.
• - Focus is still hard to control.
• - Statistical numbers can't be retrieved.
• - lacks a lot of functionality on the detail level (e.g. sums, scrolling whole pages)

SpotFire

• + Variety of specific visualizations.
• - Therefore, more difficult to learn.
• + Better access to single data entries.
• - occlusion impedes seeing proportions.

Both tools could be enhanced by

• improving interactivity (selection, adjusting the focus etc.).
• a better concept for grouping, clustering.
• better undo semantics.
• allowing to extract statistical information easily.
• enabling modifications of data (e.g. to get rid of outliers).
  
  

References

We currently build a bibliography for history visualizations on the HCIL wiki web:
http://jazz.cs.umd.edu:8080/hcil/historyviz.wiki?cmd=get&anchor=HistoryViz

(Cockburn & McKenzie 2001)
Cockburn, Andy and Bruce McKenzie (2001). What Do Web Users Do? An Empirical Analysis of Web Use. International Journal of Human-Computer Studies, In Press. 2001.
http://www.cosc.canterbury.ac.nz/~andy/papers/ijhcsAnalysis.pdf

(Mayer 2001)
Mayer, Matthias (2001). Browsing Icons. Website which accompanies the PhD project.
http://www.hfbk.uni-hamburg.de/lem/mayer/phd/

[go to top]