I did some research and according to Wikipedia an "enrichment factor" in terms of bioinformatics is...
Where possible, it's best to use explanations provided in the original source. This is because definitions could be different, depending on who is writing. Here's some descriptive information from the paper that may be more informative:
The enrichment factor (EF) is another metric we use and is very
popular for use in evaluation of virtual screening performance. EF
denotes the quotient of true actives among a subset of predicted
actives and the overall fraction of actives.... To evaluate the
maximum achievable enrichment factor (EFmax), the maximum number of
actives among the percentage of tested compounds is divided by the
fraction of actives in the entire dataset. To more consistently
quantify enrichment at 0.1% of all compounds tested, we report the
ratio of EF/ EFmax. This normalized metric is useful because EF values
are not directly comparable across different datasets due to the
maximum possible EF being constrained by the ratio of total to active
compounds and the evaluated fraction as shown in the equation above.
So for the purpose of this paper, they're saying that this ratio is an appropriately normalised substitute for a plain enrichment factor. High enrichment factor == good; low enrichment factor == bad.
Given the information you've provided, I'd agree that the curves in this figure appear to suggest that the multitask DNN performs worse with smaller numbers of active compounds (where Naive Bayes or Logistic regression would be a better choice). It's hard to tell from the graph (and legend) what the curves represent, and the clipping at EF/EFmax = 0 doesn't help.
The curves have a strange shape, suggesting the number of active compounds is dependent on the EF/EFmax ratio, rather than the other way round (which would make more sense to me). If I consider the curves to be something like a smoothed mean of the active compound number, then any place where there are fewer dots (i.e. low EF/EFmax for Multitask DNN) will have more associated error. For Multitask DNN, I see a large gap between 0.0 and 0.2, with lots of outliers on the 0.0 line (for all methods), so presumably the 0.0 outliers are skewing the curves.
Given that this is a preprint, the generation of the curves isn't specified, and the other figures (especially Figure 3) seem to tell a different story (i.e. that multi-task DNN performs better at all scales), I'd recommend contacting the authors to ask / clarify what the curves represent, or ignoring the curves entirely.