Is an ordinal class structure useful in classifier learning
read more
Citations
Ordinal Regression Methods: Survey and Experimental Study
Preference Learning and Ranking by Pairwise Comparison
Binary Decomposition Methods for Multipartite Ranking
Citation-based journal ranks: The use of fuzzy measures
Exploitation of pairwise class distances for ordinal classification
References
The Nature of Statistical Learning Theory
C4.5: Programs for Machine Learning
Data Mining: Practical Machine Learning Tools and Techniques
Statistical Comparisons of Classifiers over Multiple Data Sets
Related Papers (5)
Frequently Asked Questions (10)
Q2. What is the degree to which a learner benefits from an ordinal structure?
The degree to which a learner benefits from an ordinal structure depends on itsflexibility: Complex methods producing models with flexible decision boundaries will benefit less than methods producing simple decision boundaries.
Q3. Why have several previous studies resorted to discretized regression data for experimental purposes?
1Due to a lack of ordinal benchmark data, several previous studies, including (Frank and Hall, 2001; Fürnkranz, 2002b), have resorted to discretized regression data for experimental purposes.
Q4. What is the importance of the base learners?
As the authors have furthermore seen, the flexibility of the base learners is also important for the effectiveness of the meta-techniques investigated in this paper.
Q5. What is the definition of a prediction of the probability of a class of x?
Given a query instance x, a prediction Mi(x) is interpreted as an estimation of the probability that the class of x, denoted y(x), is in {yi+1 . . . ym}, that is, an estimation of the probability Pr(y(x) yi).
Q6. What test was used to find differences between pairs of methods?
In case this hypothesis is rejected, a Nemenyi test (Nemenyi, 1963) was applied as post-hoc test to find significant differences between pairs of methods.
Q7. Why is the VOI for the regression data so small?
since the VOI values are obviously smaller than for the regression data, the results confirm their presumption that discretized regression data exhibits an even stronger developed ordinal structure than truly ordinal data.
Q8. What is the statistical (null) hypothesis that r 0.5?
the authors test the statistical (null) hypothesis that r ≤ 0.5 against the (alternative) hypothesis r > 0.5, using a win/loss sign test according to Demšar (2006).
Q9. How can the authors get the probabilities of the individual classes yi?
the probabilities of the individual classes yi can be derived quite elegantly, namely by multiplying the probabilities along the path from the root of the tree to the leaf node for yi.
Q10. What is the difference between EOND and FH?
since bigger meta-classes will usually call for more complex models (decision boundaries), flexible classifiers such as decision trees are advantageous for EOND, and even more so for FH.