Gradient scarcity with Bilevel Optimization for Graph Learning