I assume you describe the dataset elsewhere in the paper (including data collection methods, or, if you're using and analyzing a data set published elsewhere, at minimum the data analysis methods) like Massimo Ortolano suggests. If feasible, I would suggest clearly labeling the legend entries to make the data sources clear, e.g. instead of "Group 1", "Group 2", and "Group 3" you might have "Isaev (2003)", "Someone else (2012)", and something like "This work" (or "this study"). This way the fact that the data comes from different sources gets communicated clearly even if the figure is displayed without the caption, which might well happen in talks etc.
Often more informative labels are useful, e.g. if there are datasets collected using different techniques or multiple datasets from the same source. I would slightly caution against legend entries using numerical references such as "Isaev [3]", where [3] is some entry in the list of references. I've done so in the past, and unless you automatically generate the figure the numbers can easily end up pointing to the wrong references as as citations are changed in the main paper. It is better to include such direct citations in the caption, just as you suggest. However, there is little reason to put something like "[own dataset]" in the caption, especially if you have "This work" in the legend. You may well want to devote some space in the caption contrasting your and others' results, for which phrases such as "our results" may be useful, but generally what isn't explicitly referenced is assumed to be your original work. At least, that's the case if it is a reasonable assumption in light of the rest of the work.