- Shouldn't an index have only one value in the feature axis?
Yes, that's correct. On the graph given as example this is not visible because there are too many row indexes (50000). As a consequence it's impossible to distinguish a particular index from its neighbors, but if the X axis was stretched long enough one would see a single feature value for every index.
- One horizontal line should mean that the feature values for all indexes have been uniformized, not randomized?
I think there could be two different confusions here:
- An horizontal line means that a single feature value is distributed uniformly across the indexes, which is equivalent to saying that the indexes are random for this feature value. In other words, the chance that this feature value appears at a particular index is the same as at any other index. This is what the author means: the order (indexes) is random for any feature value.
- The values for all the features have not been uniformized, this can be seen from the fact that vertically the density of the points is different around the middle (say 0.4-0.6) and the extremes (say 0-0.2 and 0.8-1). Of course this would be more visible with a standard histogram, which would show a kind of peak in the middle but with two high bars at the extremes for 0 and 1 (it can be seen from the continuous lines for these two features values that they appear much more frequently).
One may also note on this graph that there is some kind of underlying discrete distribution of the values: very clearly for values 0 and 1, but also from all the white horizontal lines which show that some values seldom exist in the data.