Inductive Logic Programming (ILP), a form of machine learning, stands out due to its emphasis on explainability and its symbolic representation. ILP combines logic programming and inductive reasoning to learn interpretable models represented as logical rules. This symbolic representation enables ILP to provide transparent and understandable explanations. Unlike black-box models generated by many machine learning approaches, the logical rules produced by ILP can be easily interpreted and reasoned about by humans. ILP's symbolic nature allows it to incorporate background knowledge and subject expertise and facilitate the discovery of high-level data concepts and relationships. 

The major contribution of this thesis is a novel ILP framework called Meta Inverse Entailment (MIE), along with an efficient implementation called PyGol. The inspiration for MIE comes from combining the advantages of Inverse Entailment and Meta-Interpretive Learning. Additionally, MIE addresses a significant issue encountered in traditional ILP systems, namely user-defined declarative bias, such as mode declarations and metarules. Declarative bias refers to the assumptions and constraints imposed on the search space of the hypothesis language. Specifying mode declarations can be challenging, as it requires domain knowledge and expertise to select appropriate modes and avoid conflicts or inconsistencies. To overcome this challenge, we introduce the concepts of related literals and the bottom clause of relevant literals (BCRL), which do not rely on declarative bias and serve as an alternative to the bottom clause as defined in Inverse Entailment. Metarules are another form of user-defined declarative bias that defines the structure of the hypothesis space in Meta-Interpretive Learning. In this research, we replace metarules with Meta Theory (MT), which can be automatically generated from background knowledge. In MIE, a novel concept of hypothesis space, a double-bounded hypothesis set, is defined by incorporating meta theory and the bottom clause of relevant literals. 

MIE supports both inductive and abductive learning. It also has the capability to perform predicate invention and learning recursive rules. The performance of MIE has been thoroughly evaluated from various perspectives. In the inductive framework, extensive evaluations were conducted using benchmark datasets, including Mutagenesis, Carcinogenesis, Alzheimer’s, DssTox, and datasets related to Phase Transition, among others.  On the other hand, the abductive framework of MIE was utilised to gain insights into microbial interaction. The accuracy achieved in both inductive and abductive frameworks were comparable with the state-of-the-art ILP systems, with the added advantage of superior speed. Furthermore, PyGol was successfully applied to several new real-world applications, including plant disease detection from leaf images and neurodegenerative disease prediction from fundus images.