Preliminaries
Broadly, the statistical learning approaches can be separated between parametric and non-parametric, and this is also the case for machine learning (ML) approaches. The taxonomy is well-studied for such cases.
There is another fundamental separation between the sub-fields of ML, that is, symbolic and non-symbolic learning. Symbolic learning is the process of learning a logical description that represents the theory underlying a certain phenomenon, such as decision trees and rule-based classifiers. Non-symbolic learning, on the other hand, is the process of learning a non-logical description that represents that phenomenon, such as deep neural networks and naive Bayes.
I'm trying to define a formal taxonomy of symbolic learning. There are at least two dimensions, for both supervised and unsupervised learning as follows.
- Language bias is composed of two parts, namely, conceptual bias and logical bias; the former is the set of propositional letters, and the latter is the grammar of the logic that the learner uses for learning.
- Search bias is the way the learner searches through the hypothesis space including the search algorithms, the strategies, and the search heuristics.
Question(s)
Are there any other dimensions that I'm missing out on? For example, in the supervised setting, decision trees are known to be divide-and-conquer approaches and rule-based algorithms are separate-and-conquer.
Moreover, in the unsupervised setting, association rules mining can be compared with other (symbolic) approaches, or it doesn't make sense?
References
