The values of perplexity

[vu-minh]

What are the range values of perplexity should we use in the experiments? Note: Perplexity should not exceed the number of instances in the dataset.

• linear-scale values: perps = range(1, N//2)
• log-scale values: perps=[???]

[vu-minh]

The value of perplexity should be integer or decimal?

• perplexity is interpreted as an approximation of number of neighbors of each data point, so should it be integer?

• perplexity should not exceed the number of instances, should we limit the maximum value to N-1 or to N//2:

• [ ] perps = range(1, N//2)

• [ ] perps = log_scale(0.1, N-1, num=N//2)

• test with some close decimal values of perp, e.g. 10.2, 10.5, 11.0 to see the differences in the embedding results.

[vu-minh]

[vu-minh]