A Hybrid Method for Rating Prediction Using Linked Data Features and Text Reviews

The presentation for our entry to the Linked Data Mining Challenge 2016 organized by Know@LOD workshop at ESWC 2016

Transcript

1. A Hybrid Method for Rating Prediction Using Linked Data Features and Text Reviews Semih, Y., Emir M., Pasquale, M., Erdogan, D., Halife, K. Linked Data Mining Challenge 2016 - Know@LOD ESWC 2016, Heraklion, Crete, Greece
2. “ What makes a good/bad album of music? Can Linked Open Data help with the classification of music albums as “good” or “bad”?
3. 1. Hyphoteses What is our intuition?
4. Bands vs. Singers Bands are more successful than single artists  rdf:type of dbo:artist
5. Music genres Some genres are more popular than others  dbo:genre http://hpo.org/two-things-you-need-to-know-about-genre-hopping/
6. Language Albums in English are more likely to be popular  dbo:language
7. Runtime Longer albums tend to be more popular  dbo:runtime
8. Reviews Words used for good albums differ from the ones used for bad albums http://www.youtube.com
9. Award winners Albums of award winning artists are likely to be more successful  # awards of dbo:artist
10. 2. Datasets and Method
11. Datasets ◎ Training dataset: 1,280 album URIs ◎ Test dataset: 320 album URIs ◎ DBpedia ◎ Metacritic.com
12. Method
13. 3. Experiments and Results
14. Experimental Setup ◎ Python 3.5.1 ◎ Beautiful Soup Library 4.4.0 (Web Scraping) ◎ scikit-learn Library 0.17 (Data Mining) ◎ Jena Fuseki (LD Caching)
15. Different Classifiers and Different Feature Sets Feature Set Linear SVM KNN RBF SVM Dec. Tree Rand. Forest AdaBoost Naïve Bayes LD 76.64% 60.47% 48.05% 72.66% 53.91% 75.00% 76.41% LDA 54.53% 52.58% 54.69% 54.45% 48.91% 54.53% 52.89% LD+LDA 76.72% 60.23% 48.05% 72.66% 52.34% 75.00% 76.41% TEXT 85.00% 50.00% 47.27% 67.27% 52.81% 78.91% 68.44% LD+LDA+TEXT 87.81% 52.81% 47.27% 72.03% 52.58% 82.50% 77.19% + = 90% test set
16. Thanks! Any questions? Find our solution in GitHub https://github.com/semihyumusak/KNOW2016