Using search engines for classification: does it still work?
- 1. USING SEARCH ENGINES FOR CLASSIFICATION: DOES IT STILL WORK? Sten Govaerts, Nik Corthaut, Erik Duval
- 2. • Our problem • Classification using search engines • The setup • The evaluation • Conclusion
- 3. TUNIFY
- 4. TUNIFY
- 5. TUNIFY
- 6. HOW DOES IT WORK? • manually annotated metadata •5 music experts at Aristo Music and different consultants • almost 80,000 songs • but, not enough...
- 7. PROBLEMS • satisfying the music choice of all customers • retail and catering differ from you and me! • new markets • react fast on emerging music trends • adding the full Belgian library catalog
- 8. GENERATE THE METADATA • from different sources: • the audio signal • web sources • the Aristo database • attention metadata • using our metadata generation framework: SamgI
- 9. GENRE... • our master thesis looked at different ways to generate genre...
- 10. ONE APPROACH... • M. Schedl, T. Pohle, P. Knees, G. Widmer, “Assigning and Visualizing Music Genres by Web-based Co-occurrence Analysis”, Proceedings of the 7th International Conference on Music Information Retrieval, 2006, pp. 260-265. • G. Geleijnse, J. Korst, "Web-based Artist Categorization", Proceedings of the 7th International Conference on Music Information Retrieval, 2006, pp. 266 - 271.
- 11. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence
- 12. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence
- 13. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence Artist + Genre + Schema
- 14. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence Artist + Genre + Schema
- 15. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence Artist + Genre + Schema
- 16. CLASSIFICATION WITH SEARCH ENGINES using co-occurrence Artist + Genre + Schema
- 18. Rock: Jazz: Blues: Pop: Country: Metal:
- 19. Rock: Jazz: 0,013 0,013 Blues: Pop: 0,009 0,015 Country: Metal: 0,009 0,005
- 20. RESULTS • master thesis student’s results were much worse • what happened? • did Google search result count change? • has Google Search API different results? • is the student’s implementation correct?
- 21. HOW TO EVALUATE THIS? • re-run the original experiment • evaluate on the same data set: 1995 artists and 9 genres. • different search engines: Google,Yahoo! and Live! Search. • over time: 8 times over a period of 36 days.
- 22. THE DATA SET Blues Country Electronic Folk Jazz Metal Rap Reggae RnB
- 23. THE DATA SET Blues Country Electronic Folk Jazz Metal Rap Reggae RnB 10% 9% 3% 2% 12% 13% 5% 4% 41%
- 24. THE DATA SET Blues Country Electronic Folk Jazz Metal Rap Reggae RnB
- 29. MORE FINE-GRAINED... • 18 artists • more search engines: Google.co.uk/.fr/.be, uk/ fr.search.yahoo.com • twice a day for 53 days • 250,000 queries!
- 30. 2 Pac Rap Alan Lomax Folk Art Pepper Jazz Cradle of Filth Metal David Parsons Electronic Desmond Dekker Reggae Downpour Metal IceT Rap Jerry Butler RnB Joy Lynn White Country Louisiana Red Blues Lou Rawls RnB LTJ Bukem Electronic Peter Tosh Reggae Pinetop Smith Jazz Robert Johnson Blues Roy Rogers Country Steeleye Span Folk
- 32. MAIN SEARCH ENGINE RESULTS
- 33. REGIONAL GOOGLES
- 35. WHAT TO USE? • use Google when it’s stable else rely on Yahoo! • when is it stable? test with a small set • some artists get classified incorrectly on bad days • compare the accuracy achieved with the test set to the average.
- 36. CONCLUSION • still works after 3 years • Google -> Yahoo! -> Live! Search • why does Google fluctuate? •a generic version of an all purpose classifier is implemented in metadata generation framework
- 37. FUTURE WORK • understand the performance differences of regional search engines • use alternative search engines • tweak the genre taxonomy depending on the search engine
- 38. Q & A.
- 39. DEMO METADATA GENERATION • http://ariadne.cs.kuleuven.be/samgi-service/
Editor's Notes
- #8: NOT the Southern African Media and Gender Institute.
- #22: 1. MG is better than MS, a possible explanation is that style is a broader term than genre for music 2. Google outperforms Yahoo! & Live! 3. results fluctuate over time 4. technical issues with Yahoo! only a fraction of the artists are retrieved
- #28: 1. the accuracy is not the exactly the same as for the large data set. but the overall trends are similar. 2. MG schema is still more accurate 3. Yahoo! MG is a very stable 4. Live! is still the worst and Google the best!
- #29: 1. Yahoo! is very stable 2. Live is the worst, Google the best! 3. no noticable differences between Live and Bing. Bing was launched on 3 June. 4. On 29 July, collaboration between Bing and Yahoo
- #30: 1. .com performs best! -> co.uk -> fr -> be 2. fr and be worse: maybe because genres are in english 3. one could also check if local artists are classified better
- #31: correct: light incorrect: dark 1. yahoo most stable 2. google changes most often. 3. changing from correct to incorrect occurs most, but no clear pattern 4. Live seems to struggle with the same artists, one time they do it correctly, the next time wrong.