Edinburgh Napier University

  With the explosion of Big Data, fast and frugal reasoning algorithms are increasingly needed to keep up with the size and the pace of user-generated contents on the Web. In many real-time applications, it is preferable to be able to process more data with reasonable accuracy rather than having higher accuracy over a smaller set of data. In this work, we leverage on both random features and random neurons to perform analogical reasoning over Big Data. Due to their big size and dynamic nature, in fact, Big Data are hard to process with standard dimensionality reduction techniques and clustering algorithms. To this end, we apply random projection to generate a multi-dimensional vector space of commonsense knowledge and use an extreme learning machine to perform reasoning on it. In particular, the combined use of random multi-dimensional scaling and randomly-initialized learning methods allows for both better representation of high-dimensional data and more efficient discovery of their semantic and affective relatedness.