Make it possible to find the closest cluster/node for new data

[pulquero]

Given a new data point and its associated lens values, make it possible to find where it should appear on the graph, i.e. find which cluster is the closest. The lens values can be used to lookup the hypercube(s), then find the nearest cluster(s), like get the N nearest points and return which clusters they belong to.

[sauln]

I could see that being very useful. Thanks for the suggestion. If you would like to take a stab at the implementation, that would be great for everyone.

[pulquero]

Here is something functional. With the addition of some extra internal data structures to track cubes->nodes this could probably be made a bit cleaner.

# maybe this should be something like cover.contains() to handle Cover subclasses
def find_hypercubes(lens_values, cover):
  cube_ids = []
  for i, center in enumerate(cover.centers_):
    lower_bounds, upper_bounds = center - cover.radius_, center + cover.radius_
    if np.all(lens_values >= lower_bounds) and np.all(lens_values <= upper_bounds):
      cube_ids.append(i)
  return cube_ids

def find_nodes(cube_ids, kmgraph):
  nodes = {}
  for node, data_ids in kmgraph['nodes'].items():
    if node.startswith(tuple(['cube'+str(i) for i in cube_ids])):
      nodes[node] = data_ids
  return nodes

def assign_nodes(data_values, data, nodes, knn):
  knn_data = []
  knn_node_ids = []
  for node, data_ids in nodes.items():
    node_data = data[data_ids]
    knn_data.append(node_data)
    knn_node_ids.append([node]*len(node_data))
  knn_data = np.vstack(knn_data)
  knn_node_ids = np.concatenate(knn_node_ids)
  knn.fit(knn_data)
  nn_ids = knn.kneighbors([data_values], return_distance=False)
  return np.unique(knn_node_ids[nn_ids])

[sauln]

Looks great! would you like to submit a PR?

[ahassaine]

This would indeed be very useful. I was wondering if this has been implemented or shall I just use the code snippet shared above?
Many thanks!

[pulquero]

Most of it has been merged already except for one outstanding PR, use the code from that branch. Sent from my Galaxy -------- Original message --------From: ahassaine ***@***.***> Date: 16/11/2021 11:15 (GMT+00:00) To: scikit-tda/kepler-mapper ***@***.***> Cc: pulquero ***@***.***>, Author ***@***.***> Subject: Re: [scikit-tda/kepler-mapper] Make it possible to find the closest cluster/node for new data (#208) This would indeed be very useful. I was wondering if this has been implemented or shall I just use the code snippet shared above? Many thanks! —You are receiving this because you authored the thread.Reply to this email directly, view it on GitHub, or unsubscribe.Triage notifications on the go with GitHub Mobile for iOS or Android.