New Metafeatures

[emrysshevek]

This is a list of metafeatures to be implemented for metafeature experiments and other use. This list will be updated with more specific metafeatures as more research is performed.

Metafeatures

• Radius Neighbors Graph based:
  • Number of Nodes (not really necessary since it's just the number of instances)
  • Number of Edges
  • (un)weighted diameter (longest shortest path in the graph)
  • (un)weighted shortest paths (over all pairs of nodes)
  • clustering (fraction of a vertices neighbors that are also neighbors of each other)
  • degree (number of vertices adjacent to a vertex)
  • class change ratio (number of edges connecting vertices of different classes)
• More Model-based:
  • KNN (no longer going to implement)
  • Perceptron ( on full dataset, 1/10 dataset, 1/2 dataset, sqrt dataset)
    • sum of weights
    • distribution of weights
    • distribution of biases
    • number of iterations to convergance
  • Clustering
  • Bayesian Network
• Timing
• DCoL
  • Regression
  • Classification
• Time-series

Summarization

• Meta-metafeatures
• sum

[MichaelMMeskhi]

Regarding DCoL, from what I have experienced, having it wrapped in Python would be best. Topological calculations are slow thus keeping them in C++ might be best. I have done some of the DCoL wrapping last year for personal purposes might give it a try again for this project if I find time.

[emrysshevek]

That would be great if you have time! Which parts have you done so far?

[MichaelMMeskhi]

I've personally wrapped computeNonLinearityLCDistance just for something specific I was trying to do. But I will look back at DCoL and see what to do with it. I was thinking of maybe even using Cython? Not sure how that will affect performance yet. Doing higher level math in C++ is way faster.

[emrysshevek]

I'm not really sure how the DCoL package works, is there a quick/easy way to just wrap a call to get all of the metrics for a given dataset?

[MichaelMMeskhi]

There is a way but if I remember correctly it writes the metrics to a file. Some modifications must be made in C++ before wrapping.

[MichaelMMeskhi]

This is a list of metafeatures to be implemented for metafeature experiments and other use. This list will be updated with more specific metafeatures as more research is performed.

Metafeatures

• Radius Neighbors Graph based:

  • Number of Nodes (not really necessary since it's just the number of instances)
  • Number of Edges
  • (un)weighted diameter (longest shortest path in the graph)
  • (un)weighted shortest paths (over all pairs of nodes)
  • clustering (fraction of a vertices neighbors that are also neighbors of each other)
  • degree (number of vertices adjacent to a vertex)
  • class change ratio (number of edges connecting vertices of different classes)

• More Model-based:

  • KNN

  • Perceptron

    • Perceptron sum of weights on full dataset, 1/10 dataset, 1/2 dataset, sqrt dataset

  • Clustering

  • Bayesian Network

• Regression

• Timing

• DCoL
  • Measures of overlaps in the feature values from different classes.
    • The maximum Fisher's discriminant ratio (F1).
    • The directional-vector maximum Fisher's discriminant ratio (F1v).
    • The overlap of the per-class bounding boxes (F2).
    • The maximum (individual) feature efficiency (F3).
    • The collective feature efficiency (F4).
  • Measures of class separability.
    • The leave-one-out error rate of the one-nearest neighbor classifier (L1).
    • The minimized sum of the error distance of a linear classifier (L2).
    • The fraction of points on the class boundary (N1).
    • he ratio of average intra/inter class nearest neighbor distance (N2).
    • The training error of a linear classifier (N3).
  • Measures of geometry, topology, and density of manifolds.
    • The nonlinearity of a linear classifier (L3).
    • The fraction of maximum covering spheres (T1).
    • The average number of points per dimension (T2).

• Time-series

Summarization

• Meta-metafeatures
• sum

[MichaelMMeskhi]

Suggestion to add a new meta-feature: Concept Variation (Task complexity) as defined here.