'Interactive Classification: get more from less by Ilze Birzniece, LV

314 views

Published on

Published in: Technology, Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
314
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
3
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

'Interactive Classification: get more from less by Ilze Birzniece, LV

  1. 1. INTERACTIVE CLASSIFICATION WITH INCLAS: GET MORE FROM LESS M.sc.ing. ILZE BIRZNIECE
  2. 2. Predictive analytics Data mining Classification Artificial intelligence Business intelligence Machine learning DATA 2
  3. 3. CLASSIFICATION  Learning from the past experience E.g. weather prediction, medical diagnostic, organizing documents, credit scoring etc.  Formalized data (attributes, classes)  Experience (training examples)  No. Income Has loan (Eur/month) Client (month) ... Outcome 1 1200 yes 23 .. Untrustful 2 900 no 50 .. Trustful 1700 yes 2 .. ? .. x  Various methods and tools for classification task 3
  4. 4. WHY NEW APPROACH? Inappropriateness of automatic classification methods for every domain where machine learning techniques could be applied to  Practical need to help experts in area of curricula comparison  Experts do not trust fully automated solutions Poor performance Complex and hard to formalize domain Small training base Incompleteness of classifier 4
  5. 5. INTERACTIVE CLASSIFICATION Automatic classification Manual classification Interactive (semiautomatic) classification 5
  6. 6. INTERACTIVE APPROACH Interactive approach includes:  Featuring classifier with ability to detect unclassified and uncertainly classified objects •Asking for the help of human •Updating classifier •Using transparent classifier 6
  7. 7. FEATURES OF INCLAS (INTERACTIVE CLASSIFICATION SYSTEM)  Dealing with       o multi-label class membership o semi-structured and unstructured data o small initial training base many classes with similar probability to appear o o various confidence thresholds single-label structured sufficient two classes traditional Involving expert in order to achieve better results 7
  8. 8. INCLAS MODEL 8
  9. 9. EXPERIMENTAL RESULTS  Number of misclassified objects can be (significantly) reduced if an interactive classification system is applied Medical diagnostics* Study course comparison 0,06 0,367 0,267 0,366 (Partly) Correct Misclassified 0,09 0,85 UnClassified 9 * From Computational Medicine Center's 2007 Medical Natural Language Processing Challenge
  10. 10. VISION InClaS: get more from less knowledge data Broadening application areas of InClaS  Extending current prototype  10
  11. 11. TARGETED ADVERTISING   Lowering advertising costs Adressing right audience Setting confidence threshold  Consulting with expert  11
  12. 12. DOCUMENT (ARTICLE, PICTURE ETC.) ORGANIZATION   Multiple categories for each object Limited amount of categorized data Multi-label classification  Overall approach for using and improving weak classifiers  12
  13. 13. RECOMMENDATIONS OF INCLAS APPLICATION  The use of the interactive classification system is feasible in areas where: Human-expert is available  Problem domain is defined by the attributes which are comprehensible for the expert   The interactive classification approach is more appropriate in areas where at least one of the following statements holds: It is essential to receive a correct classification for as much objects as possible, and it is acceptable to invest the expert’s work and time to achieve it  It is hard to extract or define domain features  Only a small initial learning set is available  13
  14. 14. LOOKING FOR COOPERATION Ilze Birzniece ilze.birzniece@rtu.lv Summary of Doctoral Thesis Development of Interactive Inductive Learning Based Classification System's Model 14

×