Fig 3. ML2Motif.
Our approach proceeds in two steps: First, feature importances are extracted from a given learning machine resulting in a gPOIM (left). Second, a corresponding convex motifPOIM is constructed by a set of proposal motifs (right). Adjusting the proposal motifs such that the distance between gPOIM and the convex motifPOIM becomes minimal, gives us the desired motifs.
