1. Internal Presentation by : Lei Wang Pervasive and Artificial Intelligenge research group http://diuf.unifr.ch/pai On: An Artificial Immune System for E-mail Classification Andy Secker, Alex Freitas, Jon Timmis Computing Laboratory, University of Kent Canterbury, Kent, UK http://www.cs.kent.ac.uk/~ads3 19/02/2004

2. An Artificial Immune System for E-mail Classification Andy Secker, Alex Freitas, Jon Timmis Computing Laboratory, University of Kent Canterbury, Kent, UK http://www.cs.kent.ac.uk/~ads3 19/02/2004

3. Significance  With the increase in information on the Internet, the strive to find more effective tools for distinguishing between interesting and non-interesting material is increasing.  This paper provides an immune-inspired algorithm called AISEC that is capable of continuously classifying electronic mail as interesting and non-interesting without the need for re-training.  Comparing with a naïve Bayesian classifier, the system proposed in this paper performs as well as the naïve Bayesian system and has a great potential for augmentation.

4. 19/02/2004 AISEC, immunity-inspired system  Immune system Human body constantly under attack. Immune system must adapt and respond The (natural) immune system is: 1.Dynamic 2.Adaptive 3.Robust 4.Etc.  Artificial Immune Systems (AIS) use principles and process from observed and theoretical immunology to solve problems

5. 19/02/2004 Artificial Immune Systems  Engineering framework Representation of individual immune cells Affinity measures Evaluate interaction of individuals with environment and/or each other Algorithms Procedures of adaptation manipulate populations of immune cells  AIS as a classifier AIRS A successful supervised AIS algorithm for classification

6. 19/02/2004 AIS for Web Mining  Web mining, an umbrella term used to describe three quite different types of data mining: Content mining A process of extracting useful information from the text, images and other forms of content that make up the pages The mining of textual data is a common task, often for the purposes of information retrieval Usage mining Structure mining  AISEC research goal To develop a highly adaptive system capable of retrieving interesting information from the internet based on user’s current interests The authors believe AIS may offer a number of advantages

7. 19/02/2004 What is AISEC ?  AISEC isn’t a spam filter It has no methods to penalize false positives (loss of important e-mail) Without a very low false positive rate, a spam filter would not be trusted

8. 19/02/2004 What is AISEC ?  AISEC is A first step towards an AIS for web mining. A study of performance and characteristics of an AIS applied to text mining in a dynamic domain A text classification algorithm capable of continuous adaptation, which may yield a classification accuracy comparable to a Bayesian approach. User behaviour and interaction with e-mail can be similar to web pages Supervised classification algorithm E-mail classified as interesting and uninteresting Uses constant(ish) feedback from user Capable of continuous adaptation This tracks concept drift and can also handle concept shift A specialised AIS algorithm based in part on the immune principle of clonal selection No previously documented algorithm was suited for use in this situation without extensive changes

9. 19/02/2004 Representation  Each cell contains 3 sets of words (+ state) Punctuation is removed from fields Research literature has suggested header information is enough to accurately classify e-mail* A = [,, ] Subject field Title of the E-mail Sender field Sender’s name Return field (Sender’s address) * Diao, Lu & Wu (2000). A Comparative Study of Classification Based Personal E-mail Filtering, PAKDD 2000

10. 19/02/2004 Affinity  Affinity value is proportion of words in one cell found in another More features would require a less naïve distance measure Cosine distance is an obvious choice Resultant value always between 0 and 1 A = [,, ] B = [,, ] affinity(A,B) = 4/9 PROCEDURE affinity (bc1, bc2) IF(bc1 has a shorter feature vector than bc2) bshort ← bc1, blong ← bc2 ELSE bshort ← bc2, blong ← bc1 count ← the number of words in bshort present in blong bs_len ← the length of bshort’s feature vector RETURN count/bs_len

11. 19/02/2004 Clone-Mutation  One mutation takes a word previously used in subject or address and replaces single location Subject, sender and return address libraries are kept separately Usually >1 mutation per cell takes place Subjectlib= free,DVD SenderLib = sales,DVD,com ReturnLib = sales,DVD,com A = [,, ] PROCEDURE clone_mutate(bc1,bc2) aff ← affinity(bc1,bc2) clones ← ∅ num_clones ← | aff * Kl | num_mutate ← | (1-aff) * bc’s feature vector length * Km | DO(num_clones)TIMES bcx ← a copy of bc1 DO(num_mutate)TIMES p ← a random point in bcx’s feature vector w ← a random word from the appropriate gene library replace word in bcx’s feature vector at location p with w bcx’s stimulation level ← Ksb clones ← clones ∪ {bcx} RETURN clones

12. 19/02/2004 The algorithm - classification 1.System is initialised with known uninteresting e- mail 2. E-mail presented for classification. Classified as uninteresting as it stimulates close cells Memory cells Naive cells

13. 19/02/2004 The algorithm – correct classification 3.Highly stimulated cell reproduces 7 times. Less stimulated cell produces only 2 clones but with higher mutation rate 4.Cell with highest affinity is known to be useful therefore rewarded by becoming memory cell. Classification Region Stimulation Region

14. 19/02/2004 The algorithm cont… Cell removal Aged naïve cells deleted. Memory cells placed in already covered areas also deleted. Incorrect classification 5.Any cell responsible for incorrect classification is removed (memory or otherwise)

15. 19/02/2004 Results – Classification accuracy  2268 e-mails (742 uninteresting) received over 6 months  E-mails presented in the order of date received  Feedback given after EVERY classification  AISEC run 10 times, results show mean  C5.0, neural network and C&R tree all run in “Clementine” data mining package  Bayesian algorithm used feedback to update like AISEC C5.083.9% Naïve Bayesian85.0% Neural Network85.6% AISEC 86.0%  1.29 C&R tree87.7% Naïve Bayesian88.05% AISEC 89.09%  0.97 Traditional LearningContinuous Learning

16. 19/02/2004 Results – variation of population size

17. 19/02/2004 User point of view  AISEC runs as a proxy on local machine  Advantages No need to switch e-mail client Can collect mail from multiple locations  AISEC’s user interface would require minimal interaction

18. 19/02/2004 User point of view Local machine Server(s) Collect mail AISEC Classifier Collect mail Store Interesting Uninteresting User interaction Positive user response Negative user response Return mail Client

19. 19/02/2004 Results cont…  Standard measures of quality Precision is the proportion of positive documents retrieved compared with the total number of positive documents Recall is the proportion of positive documents actually classified as positive PrecisionRecall Naïve Bayesian93.93%67.76% AISEC 82.20%  2.6381.13%  4.71

20. 19/02/2004 Results – variation of time between user feedback

21. 19/02/2004 Conclusion  AISEC has produced promising results and appears robust Interesting note: Typical accuracy similar to published results from other AIS for text classification (both traditional and continuous learning) Use a larger training set and optimise (the many) parameters Detect when there are the optimum number of cells  AISEC has been useful providing some evidence AIS applied to this domain would be possible  Research on adaptive systems for retrieval of interesting information, not necessarily purely accurate information

22. 19/02/2004 Questions & Discussions An Artificial Immune System for E-mail Classification Andy Secker, Alex Freitas, Jon Timmis Computing Laboratory, University of Kent http://www.cs.kent.ac.uk/~ads3c More information: