Web Log, Text, and Other Data Mining Wayne Kao
What is Data Mining? “Automated extraction of hidden predictive information from large databases” -Kurt Thearling “Quickly and thoroughly explore mountains of data, isolating the valuable, usable information -- the business intelligence” -SPSS site
Possible Questions (Chi) Usage –How has info been accessed? How frequently? What’s popular? –How do people enter the site? Where do people spend time? How long do they spend there? –How do people travel within a site? What are the [un]popular paths? –Who are the people accessing the site? From what geographical location? From what domains?
Possible Questions (cont) Structural –What information has been added? Modified? Remained the same but moved? Usage + Structural –How is new info accessed? When does it become popular? –How does introducing new information change navigation patterns? Can people still navigate there to the desired info? –Do people look for deleted information?
Usability Testing Common usability testing techniques: Interviews Ethnographic and/or lab-style observations Surveys Focus groups Good qualitative data Problems with these techniques: Time and effort are costly Small sample sizes – quantitative results? (Spool) How can we get usability testing more involved in the design cycles, so we can find problems and potential problems earlier? Design Evaluate Prototype
Remote Usability (Waterson) Analyze clickstreams in the context of the task and user intentions Human observers not present Want methods that are –Easy to deploy on any website –Compatible with range of OS and browsers Mobile computing adds further usability challenges –Small screen sizes –Limited and/or new interaction techniques –Devices are used in environments beyond the desktop
Apache Web Log [29/Mar/2002:03:58: ] "GET /~sophal/whole5.gif HTTP/1.0" " "Mozilla/4.0 (compatible; MSIE 5.0; AOL 6.0; Windows 98; DigExt)" [29/Mar/2002:03:59: ] "GET /~alexlam/resume.html HTTP/1.0" "-" "Mozilla/5.0 (Slurp/cat; [29/Mar/2002:03:00: ] "GET /~tahir/indextop.html HTTP/1.1" " "Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1)“ [29/Mar/2002:03:00: ] "GET /~tahir/animate.js HTTP/1.1" " "Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1)“
Analog - One traditional tool Reports number of requests, info about client machines, entry/exit points, charts (Chi et al.) Generated on a daily basis Typical stats Prettier stats
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
Evolution of Web Ecologies Rather than hits, focus intermediate representation on (C)ontent, (U)sage, and (T)opology, sorted by URL. –URL1: {day1: …} {day2: …} –URL2: {day1: …} Visualize an entire web site in a small amount of space Show temporal changes
Disk Tree Visualization Breadth first traversal Each ring represents a tree level All leaf nodes guaranteed some angular space (360 / # leaves) Tree linksline mark in X and Y Page access frequency line size/brightness Lifecycle stagecolor: new, continued, deleted
Disk Tree Visualization (cont) Pros –No occlusion problems since it’s 2D plane –Can use the 3rd dimension for other info (e.g. time) –Aesthetically pleasing to the eye (?) Cons –Difficult to see any page-level detail –Confusing color choices
Time Tube Visualization Put Disk Trees along spatial axis Rotated so that each slice gets equal screen area Focus+context Animation: Can fly through tube, mapping time onto time
Interaction Model Can rotate slices with a button click Can focus a slice by clicking on it Flicking gestures move slices around Right-clicking zooms to an area Mouseovers display more information about a node in a side window Can bring up pages in the browser Animation of slices
Real-world Analyzes Deadwood: Shows pages becoming [un]popular Shows effects of a redesign
Real-world Analyzes (cont) Added items are being used Deleted items aren’t negatively impacting the rest of the site
Comments Gives only a broad view of the data with no real way to get at the specifics Interaction seems very advanced Not sure how intuitive the whole idea of a circular tree is – seems kind of gratuitous
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
Association Rule? Quantitative rule that describes associations between sets of items –Not qualitative because no domain knowledge necessary for text mining Implication X Y where –X: set of antecedent items –Y: consequent item Example: 80% of people who buy diapers and baby powder also buy baby oil.
Association Rule? (cont) Support/predictability/conditional probability –Percentage of items in the total set that satisfies the union of items in the antecedent and in the consequent item Confidence/prevalence/joint probability –Percentage of articles that satisfy both the antecendent and the consequent item
Association Rule Visualization Must visualize –Antecedent items & consequent items –Associations between antecedent and consequent –Rules' support –Confidence Traditional ways of visualizing it –2D matrix –Directed graph
2D Matrix (figure 1) Antecedent and consequent items on axes Metadata icons in the cells that connect the antecedent to consequent contain support and confidence values Association rule: B C
2D Matrix (cont) Pros: one-to-one binary relationships Cons: –Hard to see association rules in many-to-one relationships (A+B C or A C and B C) –Grouping antecedents adds complexity –Object occulusion
Directed graph nodes = items edges = associations Cons: –Dozen or more items tangled display –Selecting edges to display multiple rules requires significant human interaction
Confusing?
“Novel” Technique Matrix: rule-to-item –rows = topics –columns = item associations –blue/red = antecedent and consequent Bar graph = confidence/support Can use queries to filter Mouse zooming to support context/focus
“Novel” Technique Advantages Handles hundreds of multiple antecedent association rules View topics and associations simultaneously Individual items clearly shown No antecedent groups Few occulusions because metadata is plotted at the far end and bar graph is scaled No screen swapping, animation, or serious interaction required
“Novel” Technique Demo Demo shows scalability ~9 MB news article corpus of 100,000+ documents Use word and concept-based text engines Words evaluated on whether they’re interesting depending on their position in documents Suffices removed and common prepositions, pronouns, adj’s, gerunds ignored Build a table of antecedents, consequents, confidences, and supports -> feed into viz
Conclusions Rule-to-item association Very clear visualization if limited to a few dozen rules Most web log visualizations jump to using a graph; this paper forces you to think twice.
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
VISVIP Captures individual movement between pages rather than aggregates Shows paths - sequence of URLs
Topology Directed graph Force-directed algorithm –Spring-like force –Nodes repel each other with force inversely proportional to the distance between them (i.e. closer nodes means closer pages) –Final force pulls nodes toward center
Content URLs abbreviated – bd.gif ge/abd Color-coded by content type Mouseover reveals all the abbreviated information
Simplification Common problems –Noise nodes not significant to paths - image and mailto nodes –Over-connectivity - link back to home page or company logo Solutions –Delete all edges connected to a node –Make one node the graph root –Focus on a subset of the graph
Path Sequence Showing subject paths as straight lines didn't work –Hard to follow single jagged path –Multiple paths overlapped Spline representation –Each path is a smooth curve overlaid on the graph –Colors for groups of subjects (e.g. novices)
Path Sequence (cont) User path-oriented layouts –Simpler structure than when path is laid over a graph of the entire site
Path Timing Vertical bar with base on node, its height proportional to time spent on page Animation runs through pages at times real-time Select a node to get detailed stats
Comments Capturing individual movements pretty innovative Curved user paths and reorienting the layout based on user paths Overall graph viz not too clear Good tips for creating a web log mining viz
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
Clickstream Visualizer Aggregate nodes using an icon (e.g. all the checkout pages) Edges represent transitions –Wider means more transitions
Customer Segments Collect –Clickstream –Purchase history –Demographic data Associates customer data with their clickstream (scary...) Different color for each customer segment
Filtering Using the mouse or table control, can filter by –Edge weight –Node selection Example: select checkout nodes and see if users are exiting from nodes
Layout Using third party Tom Sawyer package 1.Hierarchical from higher-out degree to higher-in degree –Mirrors actual flow of site users –The default 2.Circular –Puts related nodes into circles –Shows relationships between groups of pages
Layout (cont) Aggregation based on file system path (good idea?)
Initial Findings Gender shopping differences (intriguing...)
Initial Findings (cont) Checkout process analysis Newsletter hurting sales
Comments Visualizing clickstreams with demographic data Grouping pages by type Best use of color Icons an interesting way of reducing complexity
Readings “Visualizing the Evolution of Web Ecologies” Chi et al., Xerox PARC, 1998 “Visualizing Association Rules for Text Mining” Wong, Whitney, & Thomas, Pacific Northwest, 1999 “VISVIP: 3D Visualization of Paths through Web Sites” Cugini & Scholtz, National Institute of Standards and Technology, 1999 “Case Study: E-Commerce Clickstream Visualization Brainerd & Becker, Blue Martini Software, 2001 “What Did They Do? Understanding Clickstreams with the WebQuilt Visualization System” Waterson et al., UC Berkeley, 2002
System Design Log data with proxy Infer actions Aggregate data Layout graph Display interactive visualization
Capturing Interaction Typical HTTP request… Client BrowserWeb Server
Capturing Interaction (cont) WebQuilt captures interaction with a proxy –Proxies have typically been used for caching and firewalls WebQuilt Log Proxy Client BrowserWeb Server
Capturing Interaction (cont) If a page says: Change it to:
Capturing Interaction (cont) Pros: –Don’t need access to servers –Can analyze sites without permission from the server –Can gather clickstreams from a variety of devices including PDAs, phones, desktop computers Cons: –No access direct to the client
Visualization Interactive, zoomable directed graph Nodes = web pages Edges = aggregate traffic between pages Java-based SATIN toolkit for gesturing & zooming interaction Image rendering of web pages: JacoZoom Java callable wrappers around an ActiveX component MSIE window
Directed graph Nodes: visited pages –Color marks entry and exit nodes Arrows: traversed links –Thicker: more heavily traversed –Color Red/yellow: Time spend before clicking Blue: optimal path chosen by designer
Controls Slider: Zoom in and out Checkboxes: Filter paths to display
Pages Zooming in shows page thumbnails Arrows –Originate from actual links or the Back button –Translucent & don’t cover details
Layout Layout system flexible… 1.Edge-weighted depth-first traversal –Most visited path along top –Recursively place less followed paths below 2.Grid positioning –Organizes distance between nodes –Avoid overlapping nodes
Interaction Selecting nodes Zooming in and out Navigational gestures
Inferring & Aggregating Take log files and infer actions, such as when the back button is pressed –Can infer back button pressed, but not combinations of back and forward –Extensible framework to add other inferred actions Aggregate information, preserving individual paths
Running a WebQuilt Remote Usability Test Recruit users Design and distribute tasks (via ) Auto-collect! Watch and wait as users perform tasks and proxy logs data Visualize, analyze Use the results to change design
Pilot Usability Study Edmunds.com PDA web site Visor Handspring equipped with a OmniSky wireless modem 10 users asked to find… –Anti-lock brake information on the latest Nissan Sentra model –The Nissan dealer closest to them.
In the Lab vs. Out in the Wild Comparing in-lab usability testing with WebQuilt remote usability testing 5 users were tested in the lab 5 were given the device and asked to perform the task at their convenience All task directions, demographic data, and follow up questionnaire data was presented and collected in web forms as part of the WebQuilt testing framework.
Classifying Usability Issues Lab: Tester observations, participant comments and questionnaire data Remote: WebQuilt visualization and questionnaire data Four categories of issues Browser Device Test design Site design Six severity levels 0 indicates comment 1-5 where 1 is a very minor issue and 5 is a critical issue
Findings
WebQuilt methodology is promising for uncovering site design related issues. 1/3 of the issues were device or browser related. Browser and device issues can not be captured automatically with WebQuilt unless they cause an interaction with the server can be revealed via the questionnaire data.
Testing Concerns What to do when problems with running the test occur? Understanding user motivation is still ambiguous: Curiosity vs. confusion? Gathering qualitative feedback on mobile devices is difficult –PDA input difficult –Phones have potential for audio
Comments Zooming/filtering great for showing overview and page-level details –Can put screenshots directly into the viz Layout in relation to intended path Study compares remote usability tests to traditional tests - promising Proxy logging very cool
Future Work Expanded mobile device interaction capture, specifically net-enabled cell phones Improve filtering capabilities, integrating questionnaire and demographic data Clever algorithms to simplify graph layout Improved quantitative reporting Improved controls/interaction More rigorous evaluation with designers and usability experts
Concluding Comments Many incremental improvements in web log/data mining viz (using a graph, using demographic data, etc.) Would be really good to see a study of usability engineers and web developers comparing the tools themselves