Higher-order networks are a general form of network models that include memory nodes used to capture indirect dependencies between entities. When built from sequential or pathway data, random walks performed on these networks usually represent input sequences better than traditional first-order models. Unlike the latter, there are various ways to build higher-order networks that already exist in the literature. We introduce a new variable-order network where nodes can encode sequences of varying length. Nodes are selected by looking at sub-sequences that are unlikely extensions of already detected sub-sequences. Using experiments on real-world datasets, we demonstrate that our method produces smaller and as accurate models compared to the main variable-order model in the literature. We also study the differences achieved when ranking items using a higher-order reformulation of the PageRank metric.