Lecture 11 Collaborative Bio-Inspired Algorithms Lecture 11 : Clonal Selection Algorithms Prof Jon Timmis November 9, 2009
Lecture 11 Outline Quick Primer Static Clonal Selection AIRS Dynamic Clonal Selection AISEC Summary
Lecture 11 Quick Primer Quick Primer ◮ Supervised machine learning ◮ When you know the class of an instance before you start ◮ Wish to build a model of that data so you can then classify instances you have not seen before ◮ Many approaches including: ◮ Neural networks, rule induction, Bayesian, ILP . . .
Lecture 11 Quick Primer Learning in the Immune System ◮ Recall the learning and memory capabilities in the clonal selection process ◮ Exploit this local v global search in a learning context ◮ Evolve a set of detectors that can generalise well enough to classify unseen data items
Lecture 11 Static Clonal Selection AIRS Artificial Immune Recognition System (AIRS) ◮ Artificial Immune Recognition System (AIRS) [2] ◮ Uses the concepts of ARB’s (Artificial Recognition Balls) ◮ Resource based competition for survival in order to control the population ◮ One-shot learning system Go to the board and describe algorithm . . .
Lecture 11 Static Clonal Selection AIRS AIRS Results Data Set Accuracy Iris 96 % Ionosphere 95 . 6 % Diabetes 74 . 2 % Sonar 84 . 9 %
Lecture 11 Dynamic Clonal Selection AISEC Continuous Learning ◮ Used when you want to classify changes over time (the notion of what is in a class) ◮ Levels of what you are interested in may change over time or the context of where you are working or what you are doing ◮ Web content mining is a perfect testbed for these ideas ◮ This study looked at email classification of interesting v un-interesting
Lecture 11 Dynamic Clonal Selection AISEC Email Filtering ◮ Dynamic supervised classification algorithm ◮ E-mail classified as interesting and uninteresting ◮ Uses constant feedback from user ◮ Capable of continuous adaptation ◮ This tracks concept drift and can also handle concept shift ◮ Representation: Subject, Sender and Return address (based on existing literature, this is all you really need) ◮ Affinity measure: Proportion of words found in one cell compared to another (very naive)
Lecture 11 Dynamic Clonal Selection AISEC AISEC I Memory cells Naive cells Figure: (1) System is initialised with uninteresting emails Figure: (2) Email classified as uninteresting if high stimualtion
Lecture 11 Dynamic Clonal Selection AISEC AISEC II Stimulation Classification Region Region Figure: (3) Highly stimulated cell reproduces, as in clonal selection Figure: (4) Highest affinity cell rewarded through promotion to memory cell
Lecture 11 Dynamic Clonal Selection AISEC AISEC II Figure: (5) Any cell responsible for incorrect classification is removed Figure: (6) Aged cells (and un-stimulated) die
Lecture 11 Dynamic Clonal Selection AISEC Results ◮ 2268 e-mails (742 uninteresting) received over 6 months [1] ◮ E-mails presented in chronological order ◮ Feedback given after EVERY classification and AISEC run 10 times Technique Accuracy C5 83 . 9 % Naive Baysian (static) 85 % Neural network 85 . 6 % AISEC (static) 86 % Naive Baysian (dynamic) 88 . 05 % 89 . 05 % AISEC (dynamic)
Lecture 11 Dynamic Clonal Selection AISEC Results 99% 97% 95% Classification Accuracy 93% 91% 89% 87% 85% 83% 81% 79% AISEC 77% Bayesian 75% 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 Number of e-mails classified Figure: Classification accuracy over time
Lecture 11 Summary Summary ◮ Learning in the immune system ◮ Static clonal selection: AIRS ◮ Dynamic clonal selection: AISEC ◮ Many other variants, not covered here. Read the supporting literature.
Lecture 11 Summary A. Secker, A. Freitas, and J. Timmis. AISEC an artificial immune system for email classification. In Proceedings of the Congress on Evolutionary Computation , pages 131–139, 2003. A. Watkins, J. Timmis, and L. Boggess. Artificial immune recognition system (AIRS): An immune-inspired supervised learning algorithm. Genetic Programming and Evolvable Machines , 5(1):291–317, 2004.
Recommend
More recommend
