Mining Unexpected Rules by Pushing User Dynamics

1
Mining Unexpected Rules by Pushing User Dynamics

• Ke Wang
• Yuelong Jiang
• Laks V.S. Lakshmanan

2
Unexpected Rules

• Unexpectedness user finds the rules surprising
• Existing approaches
• Syntax distance (B. Liu, W. Hsu, AAAI96)
• Logical contradiction (B. Padmanabhan, A.
  Tuzhilin, KDD98)
• Both by direct comparison between rules

3
Our approach Data Violation

• Knowledge rules Ui
• The data rule r
• unexpected to the user who links owning
  house at BeverlyHill to movie stars and well
  paid
• Each tuple that satisfies r but violates Ui is an
  evidence for unexpectedness of r

4
Three Issues

• Knowledge Dynamics
• User decides the best knowledge to apply given a
  scenario (i.e., a tuple) --- modeling
• Knowledge Push
• Push user knowledge right from the start of
  search --- rule mining
• Unexpectedness Dynamics
• Adjust the unexpectedness of remaining rules by
  what has been presented so far --- rule selection

5
Rule Representation

• Knowledge rules and data rules
• Domain values in data rules, and fuzzy terms
  (such as High, Low) in knowledge rules.
• Match degree measures the match between a domain
  value (i.e., Primary) and a fuzzy term (i.e., Low)

Target attribute
6
Main Ideas

• Preference model the user specifies the best
  knowledge rules for each tuple
• e.g., U1 and U2 for those owning a house at
  BeverlyHill
• Violation model we measure the unexpectedness of
  r by the violation of satisfying tuples to
  their best knowledge rules.

7
The Preference Model

• User specifies covering knowledge for each tuple
• d (covering depth) best knowledge rules that
  match the tuple
• Ways to specify best
• Explicit enumeration (not scalable)
• Rank by preference max strength, best match,
  min violation, etc.

8
The Violation Model

• For a tuple t and a knowledge rule U
• Body match degree, bm(t,U), in 0,1
• Head match degree, hm(t,U), in 0,1
• Violation of U by t
• Violation of t, v(t), is aggregated v(t,U) over
  the covering knowledge U of t.

if bm(t, U) ? ? otherwise
9
The Mining Problem

• Unexpectedness Support of r
• Unexpectedness Confidence of r
• Unexpectedness of r
• Problem Find all data rules r above specified
  thresholds for Usup and Ustr.

10
The Mining Algorithm

• Three Phases
• Violation Phase
• Rule Phase
• Final Phase

11
Violation Phase

• Compute and store v(t) for all tuples t in the
  database T, pruning all t with v(t) 0 get new
  database T
• prunes the data consistent with the user
  knowledge, very effective.

12
Rule Phase

• Generate all rules r with Usup(r) above threshold
  using T
• Usup(r) is anti-monotone
• Usup(r) decreases as the body b(r) grows
• independent of preference model and violation
  function v(t)
• Any frequent itemset algorithms can be applied in
  this phase

13
Final Phase

• Compute sup(r) and sup(b(r)) for rules produced
  in rule phase
• Output rules r with Ustr(r) above threshold.

14
The Selection Problem

• Display a specified number k of rules to the
  user, in the order of unexpectedness
• See-and-Know Assumption
• After seeing rules R, user is interested in only
  rules that are unexpected with respect to

15
The Selection Algorithm

• At each step,
• greedily select the most unexpected rule (until k
  rules are selected or there is no rule to select)
• add the selected rule to user knowledge
• for each matching tuple, update the violation
  values to reflect the new covering knowledge.

16
Experiment Dataset

• KDD-CUP-98 Dataset
• Target Attribute
• NK97 donation amount in 1997 campaign
• five scales c0, c1, c2, c3, c4, in increasing
  order.
• 23 non-target attributes
• Their meanings are easier to understand than
  other attributes

17
User Knowledge

• Observation People tend to remain unchanged in
  donation behaviors
• Four knowledge rules

18
Efficiency of Mining

• Three Algorithms
• UMINE(NULL), without user knowledge
• UMINE-Unpruned, without tuple pruning
• UMINE-Pruned, pruning those tuples with vt 0

19
Interestingness of Rules
Ui(x,y) Ui covers x tuples with total violation y
20
Effectiveness of Selection
21
Conclusion

• A new approach for finding interesting rules by
  modeling user knowledge
• Violation of covering knowledge by satisfying
  tuples
• Model human user as a dynamic entity in applying
  knowledge and interpreting presented rules.
• Push user knowledge in data preparation, mining,
  and rule selection. This benefits both search and
  quality.