Lattice Structure Oriented Machine Learning

Traditional machine learning algorithms are often designed for certain types of data. For instance, the algorithm ID3 is for nominal data, and the algorithm BP is for numeric data. In many real cases, however, the Non-Match problems between datasets and algorithms are unavoidable. Generally, attribute transformation is an available solution of accommodating datasets to algorithms. Such learning scheme is called Algorithm Oriented Learning (AOL). Nevertheless, it was testified by experiments that AOL is not an efficient scheme. Some important data information would be damaged in attribute transformation, which will affect the learning.By exploring deeply the nature of the information loss in the attribute transformation under a structure-based opinion of attribute taxonomy and algorithm taxonomy, a Structure Oriented Learning (SOL) scheme is raised in this dissertation. A mid-structure is introduced by this scheme, and the mid-structure plays a role of link between datasets and algorithms, since SOL requires that both datasets and algorithms should be transformed to a consistent type of the mid-structure.The selection of mid-structure is very important in SOL scheme. On the one hand, the mid-structure should be general, which means it can be defined on the most of attribute sets. On the other hand, it also requires a capability of description, which makes it possible to explain other structures reasonably. Lattice is that a suitable structure, simple but complex in the same time.A SOL using"lattice"as a mid-structure is the main subject of this dissertation.Before discussing the instances of SOL, this dissertation will deeply analyze the normal procedure of'learning', propose a 6-tuple representation of Machine Learning Machine (MLM). Based on the notion of MLM, formal definitions of"learning"and related concepts are provided.Now the first half of this dissertation has answered the following three questions,"What is learning?","What is the Structure Oriented Learning?","What is the lattice oriented learning?".And the second half will answer"How to learn under the lattice oriented learning scheme?"by two instances.The first instance is lattice-based rule induction. At first, the Cognition-based Rule with Exception (RE) Learning (CBREL) framework is proposed, together with two algorithms of this framework, algorithm CBREL-CBL and algorithm CBREL-ID3 algorithm. Then, these two algorithms are converted to lattice algorithms by different methods. Finally, a large number of experiments are presented to compare SOL and AOL. The experimental results show that mode"lattice embedded + lattice algorithm"is significantly better than mode"discretization + nominal algorithm", and is not better than mode"encoding + numeric algorithm". This is because the"discretization"is a transformation process of complex structures to simple structures, with information loss, while"encoding"is a transformation process of simple structures to complex structures, without information loss. These analyses are consistent with the opinion of the nature of the information loss in this dissertation.The second instance is semi-lattice based LDA learning. LDA is a topical model in which documents are expressed as mixtures of topics, where a topic is a probability distribution over words. LDA model and its extension are referred to as classes of LDA models. Classes of LDA models are generally applicable to unannotated corpus. The semi-lattice oriented leaning scheme makes it possible to apply classes of LDA models to annotated corpus.