Visualising Web Visitations: A Probabilistic Approach

1
Visualising Web Visitations A Probabilistic
Approach

• Geoff Ellis Alan Dix

2
the problem
visualise activity on web sites. not for
particular analysis eg. web designers but to
give a sense of activity, a human-ness to the
web site
www.hcibook.com
3
Cyber-space is lonely

• browsing web pages is usually solitary
• like to feel that others have either
• trod these paths (pages) before
• or are currently visiting the site
• ambient displays present information in the
  environment or users periphery
• CSCW systems promote awareness for collaborative
  work (eg. Tower, Awareness Maps)

4
Cyber-space is lonely

• browsing web pages is usually solitary
• like to feel that others have either
• trod these paths (pages) before
• or are currently visiting the site
• ambient displays present information in the
  environment or users periphery
• CSCW systems promote awareness for collaborative
  work (eg. Tower, Awareness Maps)

5
web logs have (most of) it

• simple record of all requests to a web site
  (pages, graphics, document downloads etc)
• can count the number of hits to each page
• can obtain paths through a site (based on the
  visitors IP address)
• (well known problems with caching, proxy servers,
  timeouts etc)

6
visualising web sites

• numerous tools for studying structure and usage
• discover navigation problems
• categorise users, predict site usage, provide map
• often based on the premise that sites are
  hierarchical - in general they are not and paths
  through sites are definitely not

MAPA
Narcissus
7
human footsteps

• would like to see

but if web site is visualised as a linked
structure then we are more likely to see
8
alternative representation of site

• structure in terms of links between pages

• ball spring layout
• tension ? popularity
• only show popular links
• - can show paths

9
more self-organising maps

• lots of other pages pulling popular ones
• a page can only be in one place,where do you
  place it?

inspiration

• Kohonen Maps
• 2D arrangement of cells
• unsupervised learning neural network
• normally uses feature vectors (multi-dimensional
  data)
• clusters based on these vectors
• then find cell with closest match for any
  particular page
• similar pages would end up in same cell

?
?
?
?
?
?
10
modified Kohonen

• we use a radically modified version of Kohonen
  Maps which
• uses a similarity matrix (co-occurrence of highly
  used links)
• doesn't have feature vectors but uses the pages
  themselves
• tends to place pages close to, but not in the
  same cell (algorithm has a built in choose a
  neighbour function)

11
bring on Schrodingers equation
inspiration

• from the world of Quantum mechanics, very very
  small particles dont necessarily exist at any
  one particular point in space
• it solves some of the problems if particles are
  treated as wave function, and we can talk about
  the probability that a particle will be at any
  point in space
• we could treat our web pages as wave functions
• through the clustering algorithm, calculate the
  probability that a page will be in any cell on
  the grid
• gt pages not restricted to one place
• can also cope with large sites

12
Quantum Web Fields (1)

• the darker the square, the higher the probability
  that the page wants to be at that point in the
  visualisation space
• all pages are relative to each other

hci/search.html (444 visits)
home (1015 visits)
derived from a web log of 64110 records from HCI
book (Dix et al) website, pre-processed to 11992
visits by 1266 visitors. The 150 most popular
pages, in terms of adjacent pages in session
paths have been used in this example.
13
Quantum Web Fields (2)
hci/ex/chapt7.html (251 visits)
hci/ex/chapt16.html (87 visits)
bimodal - high probability of a visitor coming
from/going to the home page (circled) and also
other exercise pages
14
session paths

• Web fields are not particularly useful on their
  own
• but are the basis for plotting real session paths

• random page placement yet proportionate to
  probabilities in its web field
• subsequent pages are biased by the distance from
  last page in session
• each step occupies a new cell
• trail fades with time (numbered to ease
  interpretation)

15
example session path

• session paths visits cell which tend to be close
  to one another but have the occasional jump
  across the web field

• fairly compact trail
• commonly browsed pages as confirmed by web log
• (contents page, early chapters of the book,
  overviews and resources)

16
more session paths

• 23 page session
• browsing common pages
• shows problem with layout algorithm long
  session paths
• possible solution - fade out?

• 23 page session
• browsing common pages
• shows problem with layout algorithm long
  session paths
• possible solution - fade out?

• hopping rather than strolling?
• actually a sequential pass through most of the
  chapter pages
• may be a web crawler?
• if a common user activity then pages would have
  been clustered

17
multiple page visits

• visitor has alternated between the home page
  (circled) and other pages
• compare with ball spring visualisation (on
  right)
• QWF still gives intelligible path and shows
  home page is distributed across the Web field

18
other paths

• algorithm not constrained by site structure or
  size
• examples of long wandering session paths

19
summary

• designed to give a sense of current human
  activity on a web site rather than for analysis
• can be installed on a web server the initial
  Quantum web fields are calculated from past web
  logs (collaborative filtering/interaction
  history)
• web fields can be updated to reflect changing
  user browsing habits
• visitor paths can be shown but i) there will be
  some latency and ii) probably need to sample
• appears to give recognisable paths for normal web
  browsing activity

20
summary (2)

• can visualise sites where there are many links
  from a page
• can use for large sites without changing the size
  of the grid
• have tried with web logs from other sites
• further work
• different ways of showing the paths (transparency
  etc)
• dealing with multiple paths
• further experiments with tuning the algorithm
• install on a site and run real time
• give to a web site users to display their own
  path
• effective, scientific evaluation with real users
  in a working environment and include some case
  studies!

21
the end