Classifying Attributes with Gametheoretic Rough Sets Nouman Azam and JingTao Yao Department of Computer Science University of Regina CANADA S4S 0A2 azam200n@cs.uregina.ca jtyao@cs.uregina.ca http://www.cs.uregina.ca/~azam200n http://www.cs.uregina.ca/~jtyao Rough Sets • Sets derived from imperfect, imprecise, and incomplete data may not be able to be precisely defined. • Sets have to be approximated. • Rough sets introduces a pair of sets for such approximation. – Lower approximation – Upper approximation Visualizing Rough Sets • Let • Lower approximation. • Upper approximation. • Positive Region. • Boundary. • Negative region. Probabilistic Rough Sets • Defines the approximations in terms of conditional probabilities. – Introduces a pair of threshold denoted as (α, β) to determine the rough set approximations and regions – Lower approximation – Upper approximation – The three Regions are defined as A Key Issue in Probabilistic Rough Sets • Two extreme cases. – Pawlak Model: (α, β) = (1,0) • Large boundary. Not suitable in practical applications. – Two-way Decision Model: α = β • No boundary: Forced to make decisions even in cases of insufficient information. • Determining Effective Probabilistic thresholds. • The GTRS model. – Finds effective values of thresholds with a gametheoretic process among multiple criteria. Game-theoretic Rough Set Model Utilities for Criterion C1 (α1, β1) 0.5 (α2, β2) 0.7 (α3, β3) 0.9 Rankings based on C1 (α4, β4) 0.6 (α5, β5) 0.3 (α6, β6) 0.2 1 2 3 4 5 6 Game-theoretic Rough Set Model (α1, β1) Utilities for Criterion C2 Rankings based on C2 0.7 1 (α2, β2) 0.1 • Dilemma: (α3, β3) 0.5 – Ranking of C1 vs C2 (α4, β4) 0.6 – Which pair to select (α5, β5) 0.8 (α6, β6) 0.3 2 3 4 5 6 Game Theory for Solving Dilemma • Game theory is a core subject in decision sciences. – Prisoners Dilemma. • A classical example in Game Theory. A Game-theoretic Rough Set Approach • Obtaining Probabilistic threshold with GTRS. – An (α, β) pair is determined with game-theoretic equilibrium analysis. C1 C2 Attribute Types in Rough Sets • Reduct. – A minimal set of attribute set having the same classification ability as the entire attribute set. – Generally there may exist multiple reducts. • Core attribute. – An attribute appearing in every reduct. • Reduct attribute. – An attribute appearing in at least one reduct. • Non-reduct attribute. – An attribute that does not appear in any reduct. Limitations of Existing Methods • For classifying attributes we need to find most, if not all, reducts. • Existing methods for finding multiple reducts. – Commonly involve an iterative process. – Each iteration involves a sub-iterative process for searching a single reduct. – An attribute may be processed multiple times in different iterations of these methods. A GTRS Based Approach • We try to find an additional mechanism for classifying attributes. – Processing each attribute once to avoid extensive computations. • A GTRS based solution – Interpreting the classification of a feature as a decision problem within a game. Attribute Classification with GTRS • Formulating problems with GTRS model requires to, – – – – Identify the players. Identify the strategies of players. Determine the payoff functions. Implement a competition. Players and Strategies • Players were selected as measures of an attribute importance. – Each measure analyzes an attribute for its importance. – A case of two player game was considered. • Two strategies were formulated for each player. – Accepting an attribute, denoted as – Rejecting an attribute, denoted as J T Yao Incorporating Game Theory in Feature Selection for TC 13 Payoff Functions • Let represents a particular measure. – The value of be given as, corresponding to an attribute A, may • Notation for a payoff function. – Payoff of measure , performing action j, given action k of his opponent is denoted as, • The payoff functions of a player in four different situations of a game are calculated as, Obtaining Attribute Classification • The game may result in three possible outcomes. – Both players choose to select – One of the players choose to select – None of the players choose to select. • Attribute classification: An attribute is considered as, – core, when both players choose to select. – reduct, when one of the players choose to select. – Non-reduct, when none of the players select. Attribute Classification Algorithm A Demonstrative Example • Core = {e} • Reduct = {a,c,e} • Non-reduct = {b,d,f} The Measures in the Game • Conditional Entropy. • Attribute Dependency. Payoff Tables • The bold cell represents Nash equilibrium. – None of the players can achieve a higher payoff given their opponents chosen action. – The attribute is classified as core, since both measures choose to select, i.e. core = {e}. Payoff Tables (Cont.) • The actions of players classify the above attributes as reduct attributes. • Equilibrium analysis for attribute b, d, f suggest their classification as non-reduct attributes. Conclusion • Limitations of existing approaches. – Extensive computation due to multiple processing of individual attributes. • GTRS based method. – Interprets the classification of attributes as a game among multiple measures of attribute importance. • Importance of the method. – Each attribute is processed only once in obtaining the classification of attributes.