Dynamic rules extraction in big data context for knowledge capitalization systems

Abstract

Big data is particularly challenging, when focusing on pattern mining to find rules that describe the hidden behavior over time. Traditional formalism for rules extraction paradigms, has been extended to a high abstraction level, and improved with the automatic choice of the feature space dimension. This paper presents a novel comprehensive theory of large-scale learning with β random walk, and variational autoencoder. The new theory has the following components: 1. Rethinking learning theory; it validates the two bounds context, the local and the global by which the knowledge behavior is caught. 2. Hidden features extraction; large scale variational autoencoder provides a complete decentralization of the latent distribution resided in the latent space. Thus, as a result a new representation of the high dimensionality is replaced by a more relevant low dimensionality distribution. 3. Rules construction, the optimal bound of pattern recognition is achieved by a high abstraction level. In that sense, the proposed theory provides a new understating of the benefit of the hidden features, and gives concrete response to the diversity of rules in the big data context. The results show that the extracted rules are solid by achieving high accuracy, as well as, a high precision.