Many machine-learning researchers have utilized Occam's razor [also frequently spelt as Ockham's razor], preferring less complex classifiers in the belief that doing so is likely to reduce prediction error. I believe that this is misguided and provide philosophical and experimental support for this opinion.
The decision tree grafting software that systematically adds complexity to C4.5 decision trees while reducing prediction error can be downloaded here (requires C4.5 release 6). Decision tree grafting is also implemented in the J48Graft component of the Weka machine learning workbench.
Publications
Webb, G. I.
Decision Tree Grafting From The All Tests But One Partition
Dean, T. (Ed.): Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence (IJCAI 99), pp. 702-707, Morgan Kaufmann, Stockholm, Sweden, 1999.
@inproceedings{Webb99,
title = {Decision Tree Grafting From The All Tests But One Partition},
author = {G. I. Webb},
editor = {T. Dean},
year = {1999},
date = {1999-01-01},
booktitle = {Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence (IJCAI 99)},
pages = {702-707},
publisher = {Morgan Kaufmann},
address = {Stockholm, Sweden},
abstract = {Decision tree grafting adds nodes to an existing decision tree with the objective of reducing prediction error. A new grafting algorithm is presented that considers one set of training data only for each leaf of the initial decision tree, the set of cases that fail at most one test on the path to the leaf. This new technique is demonstrated to retain the error reduction power of the original grafting algorithm while dramatically reducing compute time and the complexity of the inferred tree. Bias/variance analysis reveal that the original grafting technique operated primarily by variance reduction while the new technique reduces both bias and variance.},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
Decision tree grafting adds nodes to an existing decision tree with the objective of reducing prediction error. A new grafting algorithm is presented that considers one set of training data only for each leaf of the initial decision tree, the set of cases that fail at most one test on the path to the leaf. This new technique is demonstrated to retain the error reduction power of the original grafting algorithm while dramatically reducing compute time and the complexity of the inferred tree. Bias/variance analysis reveal that the original grafting technique operated primarily by variance reduction while the new technique reduces both bias and variance.
Webb, G. I.
The Problem of Missing Values in Decision Tree Grafting
Antoniou, G.; Slaney, J. K. (Ed.): Lecture Notes in Computer Science Vol. 1502: Advanced Topics in Artificial Intelligence, Selected Papers from the Eleventh Australian Joint Conference on Artificial Intelligence (AI '98), pp. 273-283, Springer-Verlag, Brisbane, Australia, 1998.
@inproceedings{Webb98,
title = {The Problem of Missing Values in Decision Tree Grafting},
author = {G. I. Webb},
editor = {G. Antoniou and J. K. Slaney},
year = {1998},
date = {1998-01-01},
booktitle = {Lecture Notes in Computer Science Vol. 1502: Advanced Topics in Artificial Intelligence, Selected Papers from the Eleventh Australian Joint Conference on Artificial Intelligence (AI '98)},
pages = {273-283},
publisher = {Springer-Verlag},
address = {Brisbane, Australia},
abstract = {Decision tree grafting adds nodes to inferred decision trees. Previous research has demonstrated that appropriate grafting techniques can improve predictive accuracy across a wide cross-selection of domains. However, previous decision tree grafting systems are demonstrated to have a serious deficiency for some data sets containing missing values. This problem arises due to the method for handling missing values employed by C4.5, in which the grafting systems have been embedded. This paper provides an explanation of and solution to the problem. Experimental evidence is presented of the efficacy of this solution.},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
Decision tree grafting adds nodes to inferred decision trees. Previous research has demonstrated that appropriate grafting techniques can improve predictive accuracy across a wide cross-selection of domains. However, previous decision tree grafting systems are demonstrated to have a serious deficiency for some data sets containing missing values. This problem arises due to the method for handling missing values employed by C4.5, in which the grafting systems have been embedded. This paper provides an explanation of and solution to the problem. Experimental evidence is presented of the efficacy of this solution.
Webb, G. I.
Decision Tree Grafting
Proceedings of the Fifteenth International Joint Conference on Artificial Intelligence (IJCAI 97), pp. 846-851, Morgan Kaufmann, Nagoya, Japan, 1997.
@inproceedings{Webb97,
title = {Decision Tree Grafting},
author = {G. I. Webb},
year = {1997},
date = {1997-01-01},
booktitle = {Proceedings of the Fifteenth International Joint Conference on Artificial Intelligence (IJCAI 97)},
pages = {846-851},
publisher = {Morgan Kaufmann},
address = {Nagoya, Japan},
abstract = {This paper extends recent work on decision tree grafting. Grafting is an inductive process that adds nodes to inferred decision trees. This process is demonstrated to frequently improve predictive accuracy. Superficial analysis might suggest that decision tree grafting is the direct reverse of pruning. To the contrary, it is argued that the two processes are complementary. This is because, like standard tree growing techniques, pruning uses only local information, whereas grafting uses non-local information. The use of both pruning and grafting in conjunction is demonstrated to provide the best general predictive accuracy over a representative selection of learning tasks.},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
This paper extends recent work on decision tree grafting. Grafting is an inductive process that adds nodes to inferred decision trees. This process is demonstrated to frequently improve predictive accuracy. Superficial analysis might suggest that decision tree grafting is the direct reverse of pruning. To the contrary, it is argued that the two processes are complementary. This is because, like standard tree growing techniques, pruning uses only local information, whereas grafting uses non-local information. The use of both pruning and grafting in conjunction is demonstrated to provide the best general predictive accuracy over a representative selection of learning tasks.
Webb, G. I.
Further Experimental Evidence Against The Utility Of Occam's Razor
Journal of Artificial Intelligence Research, vol. 4, pp. 397-417, 1996.
@article{Webb96b,
title = {Further Experimental Evidence Against The Utility Of Occam's Razor},
author = {G. I. Webb},
doi = {10.1613/jair.228},
year = {1996},
date = {1996-01-01},
journal = {Journal of Artificial Intelligence Research},
volume = {4},
pages = {397-417},
publisher = {AAAI Press},
address = {Menlo Park, CA},
abstract = {This paper presents new experimental evidence against the utility of Occam's razor. A systematic procedure is presented for post-processing decision trees produced by C4.5. This procedure was derived by rejecting Occam's razor and instead attending to the assumption that similar objects are likely to belong to the same class. It increases a decision tree's complexity without altering the performance of that tree on the training data from which it is inferred. The resulting more complex decision trees are demonstrated to have, on average, for a variety of common learning tasks, higher predictive accuracy than the less complex original decision trees. This result raises considerable doubt about the utility of Occam's razor as it is commonly applied in modern machine learning.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
This paper presents new experimental evidence against the utility of Occam's razor. A systematic procedure is presented for post-processing decision trees produced by C4.5. This procedure was derived by rejecting Occam's razor and instead attending to the assumption that similar objects are likely to belong to the same class. It increases a decision tree's complexity without altering the performance of that tree on the training data from which it is inferred. The resulting more complex decision trees are demonstrated to have, on average, for a variety of common learning tasks, higher predictive accuracy than the less complex original decision trees. This result raises considerable doubt about the utility of Occam's razor as it is commonly applied in modern machine learning.
Webb, G. I.
Generality Is More Significant Then Complexity: Toward An Alternative To Occams Razor
Zhang, C.; Debenham, J.; Lukose, D. (Ed.): Artificial Intelligence: Sowing the Seeds for the Future, Proceedings of Seventh Australian Joint Conference on Artificial Intelligence (AI'94), pp. 60-67, World Scientific, Armidale,NSW, Australia, 1994.
@inproceedings{Webb94b,
title = {Generality Is More Significant Then Complexity: Toward An Alternative To Occams Razor},
author = {G. I. Webb},
editor = {C. Zhang and J. Debenham and D. Lukose},
year = {1994},
date = {1994-01-01},
booktitle = {Artificial Intelligence: Sowing the Seeds for the Future, Proceedings of Seventh Australian Joint Conference on Artificial Intelligence (AI'94)},
pages = {60-67},
publisher = {World Scientific},
address = {Armidale,NSW, Australia},
abstract = {Occam's Razor is widely employed in machine learning to select between classifiers with equal empirical support. This paper presents the theorem of decreasing inductive power: that, all other things being equal, if two classifiers a and b cover identical cases from the training set and a is a generalisation of b, a has higher probability than b of misclassifying a previously unsighted case. This theorem suggests that, to the contrary of Occam's Razor, generality, not complexity, should be used to select between classifiers with equal empirical support. Two studies are presented. The first study demonstrates that the theorem of decreasing inductive power holds for a number of commonly studied learning problems and for a number of different means of manipulating classifier generality. The second study demonstrates that generality provides a more consistent indicator of predictive accuracy in the context of a default rule than does complexity. These results suggest that the theorem of decreasing predictive power provides a suitable theoretical framework for the development of learning biases for use in selecting between classifiers with identical empirical support},
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
Occam's Razor is widely employed in machine learning to select between classifiers with equal empirical support. This paper presents the theorem of decreasing inductive power: that, all other things being equal, if two classifiers a and b cover identical cases from the training set and a is a generalisation of b, a has higher probability than b of misclassifying a previously unsighted case. This theorem suggests that, to the contrary of Occam's Razor, generality, not complexity, should be used to select between classifiers with equal empirical support. Two studies are presented. The first study demonstrates that the theorem of decreasing inductive power holds for a number of commonly studied learning problems and for a number of different means of manipulating classifier generality. The second study demonstrates that generality provides a more consistent indicator of predictive accuracy in the context of a default rule than does complexity. These results suggest that the theorem of decreasing predictive power provides a suitable theoretical framework for the development of learning biases for use in selecting between classifiers with identical empirical support