Automatic Construction of Multifaceted Browsing Interfaces

1
Automatic Construction of Multifaceted Browsing
Interfaces

• Wisam Dakka Columbia University
• Panagiotis G. Ipeirotis NYU
• Kenneth R. Wood MSR Cambridge

2
Why Guided Navigation?

• Typical search for a product name

• Multifaceted hierarchies are superior than
  single, monolithic hierarchies
• Allow users to browse across multiple dimensions
• Expose the contents of the underlying collection
  and can help users more quickly locate items of
  interest

3
Roadmap

• ?Identify manually the dimensions/facets that can
  be used to browse a collection
• Type, country, price, grape, diet
• ?A technique for extracting facets
• ? Create manually the hierarchies for each
  dimension
• The countries hierarchy
• ?An efficient construction algorithm
• Ranking categories within hierarchies
• How to show the best categories first?
• ?Ranking schemes
• ?Extensive experiments

4
Extracting Important Navigational Facets
Motivation

• Many collections with metadata organized across
  different facets
• Corbis royalty-free collection
• A set of 36,820 annotated images
• Each image has a title, a free-text description,
  and a set of associated keywords
• Total of 65,521 keywords, mainly assigned to 14
  out of the 38 facets
• And many others, like the wine collection we used
  early
• The task for a given image with its metadata,
  extract a set of proper facets or dimensions
• Idea classify keywords in the appropriate facets
• Cat and dog under animal
• Mountain and fields under topographic feature

How do we extract such facets?
5
Basic Idea for Extracting Important Navigational
Facets

• Given a collection of objects and associated
  metadata, where each object is assigned to a list
  of facets (dimensions)
• Train a classifier that given an object and its
  metadata, it projects a list of facets
• Run the classifier on a new set of objects with
  no assigned facets, to identify the frequently
  used facets
• Use the discovered facets for the guided
  navigation

6
The Classifier Straightforward Approach

• Classifying keywords in appropriate facets
• Cat or dog -gt animal
• ? Cannot generalize

????
7
The Classifier Expansion Using WordNet

• Capturing the meanings of other words using
  hypernyms
• Cat feline, carnivore, mammal, animal, living
  being, object, entity
• ? can generalize ? cannot disambiguate

Animal ?
Computer Device ?
Fields
Topographic
feline, carnivore, mammal, animal, living being,
object, entity
Hypernyms
Mountain
Topographic
Animal
Dog
Hypernyms
Hypernyms
8
The Classifier Capturing the Context

• Keywords associated with the same object give
  valuable clues
• ? can disambiguate

Animal
Computer Device
feline, carnivore, mammal, animal, living being,
object, entity
9
Building the Classifier Text Classification
Problem

• We can map our problem to be a classical text
  classification problem
• Each example
• Represents a keyword in an object
• Has a list of assigned classes (facets) to the
  keyword
• Has three vector representations
• The keyword it self
• The expansion from WordNet using hypernyms
• The context - other keywords assigned to same
  object and their hypernyms

10
Efficient Hierarchy Construction

• Once we have identified the facets, we need to
  navigate within each facet
• The subsumption algorithm (Croft and Sanderson,
  SIGIR1999)
• Improved version of the subsumption algorithm
• For the best values of the different params, the
  algorithm runs 3 time faster than the original
  subsumption algorithm
• Good integration with relational databases
• Extensive set of experiments
• Details in paper

11
Ranking Methods

• Ranking categories is important difficult
• Important limited cognitive ability to
  understand presented info
• Difficult lack of explicit user goals while
  browsing
• Maximize Coverage maximizes the number of
  objects that are covered by the displayed, top-k
  categories
• Frequency-based and set-cover schemes
• ? Structure and user effort to find items
• Structure considers the structure of the
  underlying hierarchy and the respective effort
  that the user has to put to locate items of
  interest
• Merit-based

12
Ranking Methods Maximize Coverage

• Frequency-based Ranking (Baseline)
• Users see first categories with the greatest
  wealth of information
• Low ranked categories represent only a small
  fraction of the collection
• An easy schema to implement
• Is it optimal?

13
Ranking Methods Maximize Coverage

• Set-cover Ranking
• Maximizing the cardinality of the top-k ranked
  categories
• A well-known NP-complete problem
• The optimal solution is unnecessary expensive,
  and generates non-monotonic ranking
• A greedy algorithm for approximating the
  set-cover problem

14
Ranking Methods Structure

• Merit-based Ranking
• Ranks higher categories that enable users to
  access their contents with the smallest cost, on
  average
• We start by defining the cost function T(Ci) the
  time to reach an object starting from node Ci in
  the hierarchy
• The time for reading the category headings
• The time spending on correcting mistakes
• The time for browsing the correct sub-tree
• Now let us define the Merit score based on T(Ci)

15
Merit-based Ranking

• The metric is similar to the F1-measure
• o(C) number of distinct objects classified
  under C
• Can be computed very efficiently in a bottom-up
  fashion
• ? Favors categories with low cost and large
  number of objects
• Using the merit of each category, we can rank
  categories appropriately, putting first
  categories that have good hierarchy structures
  under them and provide access to a large number
  of objects

16
Evaluation Settings

• Datasets
• Corbis royalty-free collection
• XMLTV television programs broadcasted over 261
  channels in NYC
• DMOZ real web pages from Open Directory
• Extracting important navigational facets
• Facet classifier using SVM with linear kernels
  and Ripper
• Efficient hierarchy construction
• See paper
• Ranking categories in a hierarchy
• Frequency-based, set-cover, merit-based

17
Extracting Important Navigational Facets Results
using SVM and Ripper

• Baseline
• 10 (F1) slightly above random classification
• Adding hypernyms 71 (F1)
• Adding associated keywords
• Ripper
• investigate whether rule-based assignments are
  sufficient
• High-level WordNet hypernyms
• 55 (F1), significantly worse than SVM
• Some classes (facets) work well with simple,
  rule-based assignment of terms to facets
• Generic Animals (93.3)
• Action Process Activity (35.9)

SVM with hypernyms and associated keywords
F1 harmonic mean of Precision Recall
18
Ranking Quality of the Generated Hierarchies

• How do the structural properties affect the
  browsing experience?
• Coverage the fraction of reachable objects in
  the hierarchy
• Other properties
• Average path length shorter paths are preferable
• Average branching factor users can decide faster
  which category is best with small branching
  factors
• Can we combine these metrics meaningfully?
• Cost the time to reach an object in the
  hierarchy

19
Coverage of Ranking Methods

• Set-cover consistently covers larger fraction of
  the collection
• As expected, merit-based performs slightly worse
  than the set-cover

20
Cost of Ranking Methods

• Merit-based consistently perform better than the
  other approaches, decreasing by 10-50 the time
  needed to locate items of interest

21
Ranking Conclusions

• Merit-based performs very well and offers fast
  access to the contents of the collection.
• Merit-based rankings are efficient to implement
  on top of relational database systems, while the
  set-cover rankings typically take longer to
  compute

22
Summary

• Automatically constructing multifaceted
  hierarchies
• ?A technique for extracting facets
• ?An efficient construction algorithm
• Ranking categories in hierarchy
• ?Frequency-based, set-cover, merit-based schemes
  
• ?Extensive experiments
• Automatic construction of multifaceted interfaces
  is feasible, and generates high-quality
  hierarchies

23
Future Work

• Exploring different ways of presenting the
  hierarchies to expose the contents of the
  collection in efficient ways
• Integrating better browsing and searching in
  multifaceted databases
• Indexing structures to support concurrent
  searching and browsing

24
Thank you for your time