DS2014: Feature selection in hierarchical feature spaces
Download as PPTX, PDF2 likes1,162 views
The document describes a proposed approach called SHSEL for hierarchical feature selection in machine learning. SHSEL exploits the hierarchical structure of feature spaces, where more specific features imply more general ones. It initially selects ranges of similar features in each branch based on relevance similarity. It then prunes the set further by selecting only the most relevant remaining features. The authors evaluate SHSEL on real and synthetic datasets compared to other feature selection methods, finding it achieves comparable or improved accuracy while significantly reducing the feature space.
![Problem Statement
• Each instance is an n-dimensional binary feature vector (v1,v2,…,vn),
where vi ∈ {0,1} for all 1≤ vi ≤n
• Feature space: V={v1,v2,…, vn}
• Hierarchic relation between two features vi and vj can be denoted as
vi < vj, where vi is more specific than vj
• For all hierarchical features, the following implication holds:
vi < vj → (vi = 1 → vj = 1)
• Transitivity between hierarchical features exists:
vi < vj ˄ vj < vk → vi < vk
• The problem of feature selection can be defined as finding a
projection of V to V’, where V’ ⊆ V and p(V’) ≥ p(V), where p is a
performance function:
푝: 푃 푉 → [0,1]
10/12/2014 Petar Ristoski, Heiko Paulheim 8](https://image.slidesharecdn.com/featureselectioninhierarchicalfeaturespaces-141012094559-conversion-gate02/85/DS2014-Feature-selection-in-hierarchical-feature-spaces-8-320.jpg)
![Hierarchical Feature Selection Approach
(SHSEL)
• Exploit the hierarchical structure of the feature space
• Hierarchical relation : vi < vj → (vi = 1 → vj = 1)
• Relevance similarity:
– Relevance (Blum et al.) : A feature vi is relevant to a target class C if
there exists a pair of examples A and B in the instance space such that
A and B differ only in their assignment to vi and C(A) ≠ C(B)
• Two features vi and vj have similar relevance if:
1 − 푅 푣푖 − 푅 푣푗 ≥ 푡, 푡 → [0,1]
• Goal: Identify features with similar relevance, and select the most
valuable abstract features, without losing predictive power
10/12/2014 Petar Ristoski, Heiko Paulheim 21](https://image.slidesharecdn.com/featureselectioninhierarchicalfeaturespaces-141012094559-conversion-gate02/85/DS2014-Feature-selection-in-hierarchical-feature-spaces-21-320.jpg)
DS2014: Feature selection in hierarchical feature spaces
- 1. 1 Feature Selection in Hierarchical Feature Spaces 10/12/2014 Petar Ristoski, Heiko Paulheim
- 2. Motivation: Linked Open Data as Background Knowledge • Linked Open Data is a method for publishing interlinked datasets using machine interpretable semantics • Started 2007 • A collection of ~1,000 datasets – Various domains, e.g. general knowledge, government data, … – Using semantic web standards (HTTP, RDF, SPARQL) • Free of charge • Machine processable • Sophisticated tool stacks 10/12/2014 Petar Ristoski, Heiko Paulheim 2
- 3. Motivation: Linked Open Data as Background Knowledge 10/12/2014 Petar Ristoski, Heiko Paulheim 3
- 4. Example: the Auto MPG Dataset • A well-known UCI dataset – Goal: predict fuel consumption of cars • Hypothesis: background knowledge → more accurate predictions • Used background knowledge: – Entity types and categories from DBpedia (=Wikipedia) • Results: M5Rules down to almost half the prediction error – i.e. on average, we are wrong by 1.6 instead of 2.9 MPG Attribute set Linear Regression M5Rules RMSE RE RMSE RE original 3.359 0.118 2.859 0.088 original + direct types 3.334 0.117 2.835 0.091 original + categories 4.474 0.144 2.926 0.090 original + direct types + categories 2.551 0.088 1.574 0.042 10/12/2014 Petar Ristoski, Heiko Paulheim 4
- 5. Drawbacks • The generated feature sets are rather large – e.g. for dataset of 300 instances, it may generate up to 5,000 features from one source • Increase complexity and runtime • Overfitting for too specific features 10/12/2014 Petar Ristoski, Heiko Paulheim 5
- 6. Linked Open Data is Backed by Ontologies LOD Graph Excerpt Ontology Excerpt 10/12/2014 Petar Ristoski, Heiko Paulheim 6
- 7. HIERARCHICAL FEATURE SPACE 10/12/2014 Petar Ristoski, Heiko Paulheim 7
- 8. Problem Statement • Each instance is an n-dimensional binary feature vector (v1,v2,…,vn), where vi ∈ {0,1} for all 1≤ vi ≤n • Feature space: V={v1,v2,…, vn} • Hierarchic relation between two features vi and vj can be denoted as vi < vj, where vi is more specific than vj • For all hierarchical features, the following implication holds: vi < vj → (vi = 1 → vj = 1) • Transitivity between hierarchical features exists: vi < vj ˄ vj < vk → vi < vk • The problem of feature selection can be defined as finding a projection of V to V’, where V’ ⊆ V and p(V’) ≥ p(V), where p is a performance function: 푝: 푃 푉 → [0,1] 10/12/2014 Petar Ristoski, Heiko Paulheim 8
- 9. Hierarchical Feature Space: Example Josh Donaldson is the best 3rd baseman in the American League. LeBron James NOT ranked #1 after newly released list of Top NBA players “Two things are infinite: the universe and human stupidity; and I'm not sure about the universe.”―Albert Einstein Nineteen-year-old figure skater Yuzuru Hanyu, who won a gold medal in the Sochi Olympics, is among the 684 peo... http://bit.ly/1kb6W5y In his weekly address, President Barack Obama discusses expanding opportunity for hard-working Americans: http://ofa.bo/ccH Barack Obama cracks jokes at Vladimir Putin's expense http://dlvr.it/5Z7JCR I spotted the Lance Armstrong case in 2006 when everyone thought he was God, and now this case catches my attention. 10/12/2014 Petar Ristoski, Heiko Paulheim 9
- 10. Hierarchical Feature Space: Example dbpedia-owl: Basketball_Player dbpedia-owl: Baseball_Player dbpedia-owl: Athlete dbpedia:LeBron_James dbpedia:Josh_Donaldson Josh Donaldson is the best 3rd baseman in the American League. LeBron James NOT ranked #1 after newly released list of Top NBA players 10/12/2014 Petar Ristoski, Heiko Paulheim 10
- 11. Hierarchical Feature Space: Example 10/12/2014 Petar Ristoski, Heiko Paulheim 11
- 12. Hierarchical Feature Space • Linked Open Data – DBpedia, YAGO, Biperpedia, Google Knowledge Graph • Lexical Databses – WordNet, DANTE • Domain specific ontologies, taxonomies and vocabularies – Bioinformatics: Gene Ontology (GO), Entrez – Drugs: the Drug Ontology – E-commerce: GoodRelations 10/12/2014 Petar Ristoski, Heiko Paulheim 12
- 13. RELATED APPROACHES 10/12/2014 Petar Ristoski, Heiko Paulheim 13
- 14. Standard Feature Selection • Wrapper methods – Computationally expensive • Filter methods – Several techniques for scoring the relevance of the features • Information Gain • χ2 • Information Gain Ratio • Gini Index – Often similar results 10/12/2014 Petar Ristoski, Heiko Paulheim 14
- 15. Optimal Feature Selection 10/12/2014 Petar Ristoski, Heiko Paulheim 15
- 16. Standard Feature Selection: Information Gain 10/12/2014 Petar Ristoski, Heiko Paulheim 16
- 17. TSEL Feature Selection • Tree-based feature selection (Jeong et al.) – Select most representative and most effective feature from each branch of the hierarchy • 푙푖푓푡 = 푃(푓|퐶) 푃(퐶) 10/12/2014 Petar Ristoski, Heiko Paulheim 17
- 18. Bottom-Up Hill-Climbing Feature Selection • Bottom-up hill climbing search algorithm to find an optimal subset of concepts for document representation (Wang et al.) 푓 = 1 + α − 푛 α ∗ β ∗ 푖∈퐷 퐷푐푖 , 퐷푐푖⊆ 퐷퐾푁푁푖 푎푛푑 β > 0 10/12/2014 Petar Ristoski, Heiko Paulheim 18
- 19. Greedy Top-Down Feature Selection • Greedy based top-down search strategy for feature selection (Lu et al.) – Select the most effective nodes from different levels of the hierarchy 10/12/2014 Petar Ristoski, Heiko Paulheim 19
- 20. PROPOSED APPROACH 10/12/2014 Petar Ristoski, Heiko Paulheim 20
- 21. Hierarchical Feature Selection Approach (SHSEL) • Exploit the hierarchical structure of the feature space • Hierarchical relation : vi < vj → (vi = 1 → vj = 1) • Relevance similarity: – Relevance (Blum et al.) : A feature vi is relevant to a target class C if there exists a pair of examples A and B in the instance space such that A and B differ only in their assignment to vi and C(A) ≠ C(B) • Two features vi and vj have similar relevance if: 1 − 푅 푣푖 − 푅 푣푗 ≥ 푡, 푡 → [0,1] • Goal: Identify features with similar relevance, and select the most valuable abstract features, without losing predictive power 10/12/2014 Petar Ristoski, Heiko Paulheim 21
- 22. Hierarchical Feature Selection Approach (SHSEL) • Initial Selection – Identify and filter out ranges of nodes with similar relevance in each branch of the hierarchy • Pruning – Select only the most relevant features from the previously reduced set 10/12/2014 Petar Ristoski, Heiko Paulheim 22
- 23. Initial SHSEL Feature Selection 1. Identify range of nodes with similar relevance in each branch: – Information 푠 푣, 푣gain: = 1 푠− (푣푖 0.45 , 푣푗 ) = − 1 0.5 − = 퐼퐺 0.95 푣푖 − 퐼퐺(푣푗 ) 푖 푗 – Correlation: t=0.9 푠(푣푖 , 푣푗) = 퐶표푟푟푒푙푎푡푖표푛(푣푖 , 푣푗) s>t 2. If the similarity is greater than a user specified threshold, remove the more specific feature, based on the hierarchical relation 10/12/2014 Petar Ristoski, Heiko Paulheim 23
- 24. Post SHSEL Feature Selection • Select the features with the highest relevance on each path – user specified threshold – select features with relevance above path average relevance 퐼퐺(푣푖)=0.2 AVG(Sp)=0.25 10/12/2014 Petar Ristoski, Heiko Paulheim 24
- 25. EVALUATION 10/12/2014 Petar Ristoski, Heiko Paulheim 25
- 26. Evaluation • We use 5 real-world datasets and 6 synthetically generated datasets • Classification methods: – Naïve Bayes – k-Nearest Neighbors (k=3) – Support Vector Machine (polynomial kernel function) No parameter optimization 10/12/2014 Petar Ristoski, Heiko Paulheim 26
- 27. Evaluation: Real World Datasets Name Features #Instances Class Labels #Features Sports Tweets T DBpedia Direct Types 1,179 positive(523); negative(656) 4,082 Sports Tweets C DBpedia Categories 1,179 positive(523); negative(656) 10,883 Cities DBpedia Direct Types 212 high(67); medium(106); low(39) 727 NY Daily Headings DBpedia Direct Types 1,016 positive(580); negative(436) 5,145 StumbleUpon DMOZ Categories 3,020 positive(1,370); negative(1,650) 3,976 • Hierarchical features are generated from DBpedia (structured version of Wikipedia) – The text is annotated with concepts using DBpedia Spotlight • The feature generation is independent of the class labels, and it is unbiased towards any of the feature selection approaches 10/12/2014 Petar Ristoski, Heiko Paulheim 27
- 28. Evaluation: Synthetic Datasets Name #Instances Class Labels #Features S-D2-B2 1,000 positive(500); negative(500) 1,201 S-D2-B5 1,000 positive(500); negative(500) 1,021 S-D2-B10 1,000 positive(500); negative(500) 961 S-D4-B2 1,000 positive(500); negative(500) 2,101 S-D4-B4 1,000 positive(500); negative(500) 1,741 S-D4-B10 1,000 positive(500); negative(500) 1,621 • Generate the middle layer using polynomial function • Generate the hierarchy upwards and downwards following the hierarchical feature implication and transitivity rule • The depth and branching factor are controlled with parameters D and B 10/12/2014 Petar Ristoski, Heiko Paulheim 28
- 29. Evaluation: Synthetic Datasets • Depth = 1 & Branching = 2 1 1 1 1 0 1 0 1 1 0 1 0 0 0 0 1 0 0 1 0 0 10/12/2014 Petar Ristoski, Heiko Paulheim 29
- 30. Evaluation: Synthetic Datasets Name #Instances Class Labels #Features S-D2-B2 1,000 positive(500); negative(500) 1,201 S-D2-B5 1,000 positive(500); negative(500) 1,021 S-D2-B10 1,000 positive(500); negative(500) 961 S-D4-B2 1,000 positive(500); negative(500) 2,101 S-D4-B4 1,000 positive(500); negative(500) 1,741 S-D4-B10 1,000 positive(500); negative(500) 1,621 • Generate the middle layer using polynomial function • Generate the hierarchy upwards and downwards following the hierarchical feature implication and transitivity rule • The depth and branching factor are controlled with parameters D and B 10/12/2014 Petar Ristoski, Heiko Paulheim 30
- 31. Evaluation: Approach • Testing all approaches using two classification methods – Naïve Bayes, KNN and SVM • Metrics for performance evaluation – Accuracy: Acc V′ = 퐶표푟푟푒푐푡푙푦 퐶푙푎푠푠푓푖푒푑 퐼푛푠푡푎푛푐푒푠 (푉′) 푇표푡푎푙 푁푢푚푏푒푟 표푓 퐼푛푠푡푎푛푐푒푠 – Feature Space Compression: 푐 푉′ = 1 − |푉′| |푉| – Harmonic Mean: 퐻 = 2 ∗ 퐴푐푐 푉′ ∗푐 푉′ 퐴푐푐 푉′ +푐 푉′ • Results calculated using stratified 10-fold cross validation – Feature selection is performed inside each fold • Parameter optimization for each feature selection strategy 10/12/2014 Petar Ristoski, Heiko Paulheim 31
- 32. • Classification accuracy when using different relevance similarity threshold on the cities dataset 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Relevance Similarity Threshold Accuracy Compression H. Mean Evaluation: SHSEL IG 10/12/2014 Petar Ristoski, Heiko Paulheim 32
- 33. Evaluation: Classification Accuracy (NB) 10/12/2014 Petar Ristoski, Heiko Paulheim 33 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Sports Tweets T Sports Tweets C StumbleUpon Cities NY Daily Headings original initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% S_D2_B2 S_D2_B5 S_D2_B10 S_D4_B2 S_D4_B5 S_D4_B10 original initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown
- 34. Evaluation: Feature Space Compression (NB) 10/12/2014 Petar Ristoski, Heiko Paulheim 36 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Sports Tweets T Sports Tweets C StumbleUpon Cities NY Daily Headings initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% S_D2_B2 S_D2_B5 S_D2_B10 S_D4_B2 S_D4_B5 S_D4_B10 initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown
- 35. Evaluation: Harmonic Mean (NB) 10/12/2014 Petar Ristoski, Heiko Paulheim 39 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% Sports Tweets T Sports Tweets C StumbleUpon Cities NY Daily Headings initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown 100.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00% S_D2_B2 S_D2_B5 S_D2_B10 S_D4_B2 S_D4_B5 S_D4_B10 initialSHSEL IG initialSHSEL C pruneSHSEL IG pruneSHSEL C SIG SC TSEL Lift TSEL IG HillClimbing GreedyTopDown
- 36. Conclusion & Outlook • Contribution – An approach that exploits hierarchies for feature selection in combination with standard metrics – The evaluation shows that the approach outperforms standard feature selection techniques, and other approaches using hierarchies • Future Work – Conduct further experiments • E.g. text mining, bioinformatics – Feature Selection in unsupervised learning • E.g. clustering, outlier detection • Laplacian Score 10/12/2014 Petar Ristoski, Heiko Paulheim 43
- 37. 44 Feature Selection in Hierarchical Feature Spaces 10/12/2014 Petar Ristoski, Heiko Paulheim