Research Colloquium Presentation by N'Dah Jean Kouagou

3 years ago

ML on Knowledge Graphs

Concept Learning using refinement operators

Here, I gave a brief introduction to concept learning using refinement operators. In particular, I gave an example of search tree of a concept learning algorithm. I also argued that current concept learning approaches do not scale up to the sizes of knowledge graphs.

Prediction of concept lengths for fast concept learning

I implemented four neural network architectures--MLP, CNN, LSTM, GRU, for the prediction of concept (in $\mathcal{ALC}$) lengths. The length predictors were evaluated on three benchmark knowledge graphs--Carcinogenesis, Moral Reasoner, Semantic Bible, which can be downloaded from SmartDataAnalytics. From the results of the evaluation, recurrent neural networks perform best at the prediction of concept lengths with an F-measure of up to $72\%$.

Integration of concept length predictors into CELOE

The pre-trained length predictors are used to predict the length of the goal concept, which allows the Class Expression Learning for Ontology Engineering (CELOE) algorithm to only test concepts of length at most the predicted length. This new approach appears to be at least $5\times$ faster than traditional approaches while preserving the same solution quality.

The slides of the presentation can be found here