I am currently working on unsupervised feature importance ranking using graph clustering methods, specifically focusing on the Laplacian score as a metric. However, I am struggling to clarify the interpretation of the values obtained from Laplacian scores. I have come across conflicting information regarding whether lower or higher Laplacian scores indicate better quality of information within features. What do lower and higher values of Laplacian scores signify in terms of feature quality? Specifically, does a lower score imply that a feature is more informative and better at maintaining local structure? Conversely, does a higher score suggest that the feature may be less relevant? Is there a specific threshold or range of values that can help assess the quality of information provided by features based on their Laplacian scores?
