You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Jul 13, 2025. It is now read-only.
Apparently, the order of question the AI is suggesting based on XAI is not good. Also, because we didn't do any logical verification (see #Description for more info), the questions and hence user experience feels bad.
In this issue, we want to fix the problem of some of the most important questions were not being asked at all
Description
Some questions such as travel history or invitation letter are not being asked at all. In #43 it was supposed to happen by actually retrieving data from the real cases through our dashboard, but currently, the data is not available. Yet, these questions are so important that needs to be included.
For inclusion of these questions (which we used to call them "others"), we can take two approaches:
Directly including the parameters into the pretrained models i.e., inserting a tree into another tree ([Transfer Learning] Manual insertion of "other" features and their importance to FLAML models #17): This requires heavily studying the structure of the model trained which practically might change for each run, hence we have to cover more than 10 ML models. This might result in tone of bugs, and it is way harder to debug. (these efforts could be moved to gathering data, and it is worth much more)
Manipulating the output values of the current model Avoid 50% result #88:, including XAI, prediction probability, aggregated XAI, and so on: This is way easier to investigate, implement and debug. Particularly, since all values of importance of these features, and their effect on the final prediction are given, there is no further processing required, and we can just manually change the final output of our model by just manipulating values on the fly via given heuristics.
Based on this information, I suggest following the second approach.
Intro
Apparently, the order of question the AI is suggesting based on XAI is not good. Also, because we didn't do any logical verification (see #Description for more info), the questions and hence user experience feels bad.
In this issue, we want to fix the problem of some of the most important questions were not being asked at all
Description
Some questions such as travel history or invitation letter are not being asked at all. In #43 it was supposed to happen by actually retrieving data from the real cases through our dashboard, but currently, the data is not available. Yet, these questions are so important that needs to be included.
For inclusion of these questions (which we used to call them "others"), we can take two approaches:
Based on this information, I suggest following the second approach.
Tasks