Free cookie consent management tool by TermsFeed Policy Generator

Changes between Initial Version and Version 1 of Ticket #2700, comment 45


Ignore:
Timestamp:
07/12/17 13:18:51 (7 years ago)
Author:
abeham
Comment:

Legend:

Unmodified
Added
Removed
Modified

  • Ticket #2700, comment 45

    initial v1  
    22
    33=== Distance related calculation ===
    4 This should all be moved to HeuristicLab.Common.
     4~~This should all be moved to HeuristicLab.Common~~.
    55
    66The InnerProductDistance needs to be looked at. It does not satisfy point 3 of the conditions mentioned in IDistance. It is not zero-reflexive, because d(x,x) can only be 0 if x is equal to the null vector. Inner product is not a measure for distance, because it gets larger the more similar two vectors are (if normalized with the product of the vectors' lengths). Also, the description mentions angular distance and speaks of normalization (there is no normalization). It is not clear to me why it is only defined for vectors with all non-negative dimensions. A distance measure based on inner product would probably look something like `1 - cosine(alpha)` where `cosine(alpha) = <x,y> / (||x|| * ||y||)` (<x,y> = inner product) to me. This results in a distance measure in the interval [0; 2] related to the angle of two vectors (0 for vectors of identical angle, 1 for vectors perpendicular to each other and 2 for vectors in the exact opposite direction).
    77
    8 Implementation of the distance measures with .Zip() and .Sum() is nice to read, but inefficient. I would recommend to implement efficiently and sacrifice a little bit on readability, especially if it's put in Common.
     8Implementation of the distance measures with .Zip() and .Sum() is nice to read, but inefficient. I would recommend to implement efficiently and sacrifice a little bit on readability~~, especially if it's put in Common~~.
    99
    1010=== Collection and Trees ===
    11 It is mentioned in the TODO of PriorityQueue that this should move to HeuristicLab.Collections. However, it's not a storable class and it's not observable. I would move it to HeuristicLab.Common as well. VantagePointTree and SpacePartitioningTree should also be moved to Common (maybe in a folder Collections).
     11It is mentioned in the TODO of PriorityQueue that this should move to HeuristicLab.Collections. However, it's not a storable class and it's not observable. I would move it to HeuristicLab.Common. VantagePointTree and SpacePartitioningTree should also be moved to Common (maybe all together in a folder Collections).
    1212
    1313=== Static TSNE ===
    14 It uses different default values than the TSNE. I personally think that an eta of 200 is too much, I would go with the same values that are default in the tSNE for regression problems. van der Maarten mentions adaptive learning rate at some points.
     14It uses different default values than the TSNE. I personally think that an eta of 200 is too much, I would go with the same values that are default in the tSNE for data analysis problems. van der Maarten mentions adaptive learning rate at some points.
    1515
    1616=== TSNEAlgorithm ===