Forum Discussion
Interpreting Key Influencer Visual
Well, the way I look at this is the following (maybe I am wrong, and if so please correct me). I don't know the math behind this, but this is what I learned about statements as 'is x times more likely to have y in case of z) Statistically, it normalizes other variables around it so that it truly can be compared based on the specific variable you are looking at (in this case, subcategory).
Think of it as this example, a table of festivals:
| FestivalName | Visitors | Weather | Revenue |
| A | 1000 | Sun | 100.000 |
| B | 1100 | Sun | 120.000 |
| C | 20000 | Rain | 1.000.000 |
| D | 500 | Sun | 60.000 |
| E | 18000 | Rain | 950.000 |
If you don't look at the whole dataset (all columns), you might draw statements like 'You have much more visitors when it rains'. Obviously this is wrong, these festivals (C and E) might be just very large festivals that even with rain have a higher number of visitors. So you take into account the Revenue and correct for that. Now a better conclusion will be "You have more revenue per visitor when it is sunny", which suddenly makes sense. But you can't say that by looking at absolute numbers or subsets of the data.
So, when you say 'with a higher volume of projects AND a larger proportion of successful projects', that doesn't mean yet that this give a higher chance on success, if other factors combined (and combined over all other subcategories) are playing a role.
I hope this makes sense, it's late where I am 😉
Kind regards
Djerro123
-------------------------------
If this answered your question, please mark it as the Solution. This also helps others to find what they are looking for.
Keep those thumbs up coming! 🙂
Key Influencer visual do balanced down-sampling to 10K, i.e., if you have 130K TRUE target and 140K FALSE target, it will only take 5K TRUE and 5K False due to the performance concern for interactive visual and platform limitation (cannot afford to train too large input dataset with ML models).
They adjust the lift with the positiveRatio = subsetTrue/totalTrue, and negativeRatio. Yet it is still likely to have the discrepency on lift because sampling is just a subset of the total population.
In particular for sparse features, e.g., your subcategory may have a lot of distinct categries, and Literary Journals is one of them with small portion of input, the sampling may not work that well.