Query Result Clustering for Objectlevel Search

1
Query Result Clustering for Object-level Search

Seung-won Hwang (POSTECH) Joint work w/ Jongwuk
Lee (POSTECH)Zaiqing Nie, Ji-rong Wen (MSRA)
2
Outline

• Motivation
• Observation
• Preliminaries
• Algorithm
• Experiments

3
Motivation (1 / 4)

• Given a query, search engines retrieve relevant
  results.
• Personalization maximize the satisfaction of a
  particular user.
• Diversification minimize the dissatisfaction of
  varying user intents.

Canon 5D
Good!!
Canon 5D
Good??
4
Motivation (2 / 4)

• Query result organization
• Provide end-users with a succinct overview of
  relevant results.
• e.g., a topic hierarchy or topic terms on a map

Canon 5D
Good!!
Canon 5D
Good!!
5
Motivation (3 / 4)

• Document-level search
• Documents as an information unit
• e.g., Microsoft Live Search
• Object-level search
• Web objects as an information unit
• e.g., Microsoft Libra, Product Search
• More concise results for object queries

6
Motivation (4 / 4)

• Our goal

Query result clustering for object-level search
Visualize a graph for object-level
summarization. Center a query object
Nodes relevant objects Edges relationships
between objects
Users can easily recognize relevant objects, then
drill-down their interests.
7
Motivation (4 / 4) - Demo
8
Why Challenging Object-level Search

• Documents
• A vector of frequencies (homogeneous)
• Well-agreed similarity

• Objects
• A vector of values
• (heterogeneous)
• Different similarity

Compare objects with different importance between
features.
Compare TF-IDF vectors between Docs.
???
???
Sensor size, Optical zoom, Resolution, Weight,
9
Observation (1 / 2)

• Feature-based similarity
• Depend on data-specific and intent-specific
  characteristics.
• Need to use a measure to identify both a relevant
  feature set and the corresponding distance.

DSLR cameras Sensor size, Optical zoom
Compact camerasResolution, Weight
10
Observation (2 / 2)

• Exploit the intuition of subspace clustering to
  identify relevant objects on different subspaces.

AB subspace
BC subspace
11
Preliminaries (1 / 4)

• Challenging issue on subspace clustering
• Expensive to enumerate all possible 2d - 1
  subspaces
• Hard to select a desirable subspace and distance
  among all subspaces

F1 Sensor size, F2, Optical zoom,F3 Weight
???
???
???
???
???
12
Preliminaries (2 / 4)

• Possible solution
• Introduce parameters to save the enumeration cost
  of all subspaces.
• Rmin The minimum distance on a feature
• dmin The minimum number of subspaces

Rmin 0.5, dmin 2
v
Feature-based similarity matrix
(Darker cells indicate closer pairs.)
13
Preliminaries (2 / 4)

• Possible solution
• Introduce parameters to save the enumeration cost
  of all subspaces.
• Rmin The minimum distance on a feature
• dmin The minimum number of subspaces

Rmin 0.5, dmin 2
Feature-based similarity matrix
(Darker cells indicate closer pairs.)
14
Preliminaries (3 / 4)

• Problem
• Clustering results heavily depends on parameters
  Rmin and dmin.
• It is hard to find desirable parameter settings.
• Our solution
• Exploit co-occurrence as votes reflecting wisdom
  of crowds.

15
Preliminaries (4 / 4)

• Co-occurrence similarity
• Pros Presented as ground-truth reflected from
  creators intuition.
• Cons Include inconsistent meanings for different
  characteristics.
• Need to use a complementary measure to
  disambiguate different characteristics.

(1, 2) Similar DSLRs
((1, 2), 6) DSLRs and high-end compact cameras
Co-occurrence matrix
16
Preliminaries (4 / 4)

• Co-occurrence similarity
• Pros Presented as ground-truth reflected from
  creators intuition.
• Cons Include inconsistent meanings for different
  characteristics.
• Need to use a complementary measure to
  disambiguate different characteristics.

Co-occurrence matrix
17
Algorithm (1 / 4)

• Co-occurrence similarity
• Provide ground-truth for the relationships
  between objects.
• Do not distinguish different relationships.

• Feature-based similarity
• Disambiguate inconsistent information for
  different relationships.
• Cluster quality heavily depends on parameters.

18
Algorithm (2 / 4)

• Associate co-occurrence with feature-based
  similarity
• Use co-occurrence similarity to determine the
  order of merging clusters.
• Provide a less sensitive property for specific
  parameters.
• Use feature-based similarity to disambiguate
  relationships with different characteristics.
• Only merge objects with consistent relationships.

19
Algorithm (3 / 4)
Co-occurrence matrix
Rmin 0.5, dmin 2
Feature-based similarity matrix
20
Algorithm (3 / 4)
Co-occurrence matrix
Clustering results
Rmin 0.5, dmin 2
Feature-based similarity matrix
21
Algorithm (4 / 4)

• Parameter setting
• Abstracted as a multivariate interpolation
  problem of (Rmin, dmin).
• Basically, use a linear loosening.
• Unnecessary computation
• Improved loosening tuning
• Conservative loosening
• Use single-linkage clustering.
• Estimate dmin.
• Aggressive loosening
• Use the distribution on pair-wise feature-based
  distances.
• Estimate Rmin and dmin as median values.

22
Experiments (1 / 6)

• Real-life user study
• Conduct 32 people (MSRA interns and POSTECH
  students)
• Cameras Canon Powershot SD850, Nikon D80, Nikon
  D2Xs
• Laptops Lenovo Thinkpad R61, T60, T61

HAC only use co-occurrence. HARP only use
feature-based similarity.
23
Experiments (2 / 6)

• Synthetic datasets
• Parameter settings
• Quality metrics
• CE (Clustering Error) An ideal value is
  minimized as 0.
• F1-value, FF1-value An ideal value is maximized
  as 1.

24
Experiments (3 / 6)

• Varying average feature size

Ideal
Ideal
HydraAdaptive Use aggressive loosening with
Hydra.
25
Experiments (4 / 6)

• Varying cluster size

26
Experiments (5 / 6)

• Varying sparsity

Note When sp 0, co-occurrence exists for all
possible pairs.
27
Experiments (6 / 6)

• Efficiency for cardinality and dimensionality

28
Q A
Thank you!!
28