1
$\begingroup$

Say I have 100 points. Each of these points have a date associated with them and have a "yes" label. I can artificially add "no" points to create more data. Now, I want, using these 100 "yes" points + X "no" points, to predict if a given set of features (a new point) will be categorized "yes" or "no".

I can develop a model that does so.

Now, let's imagine that some "rules of the game" can change. Something that in 1967 was a "yes" point, becomes a "no" point after a (unknown) rule change in 1989. Most factors stay the same, but one change causes the point to change classification drastically.

Can I modify my data to add more importance to recent dates (that is, duplicate recent values so that in "importance", 1 value from 2010 equals 3 values from 1990 and 5 values from 1970 for example), because it's a better representation of our prediction for a future point? Or is this a terrible idea?

Basically, to simplify, if my data is 5 points, labeled "1939", "1967", 1980", "1982", "2010", can I artificially modify it to be 12 points, "1939", "1967", 1980", "1980", "1982", "1982", "1982", "2010", "2010", "2010", "2010", "2010"?

Of course, chosing the "importance" of a given more recent point is pretty difficult to begin with.

$\endgroup$

1 Answer 1

0
$\begingroup$

Maybe I misunderstood something, but if you have feature vector with N features, it would be ok to add one more feature reflected the time. Then there could be two instances with the same first N features, but the different (N+1)-th and as a result the answer will differ too: "yes" for one year and "no" for another. However, for some instances answer will stay the same, which is also possible. So, I guess your idea might work, if there's actually some connection and meaningful "pattern" in real-life data.

$\endgroup$
1
  • $\begingroup$ This is a good point. Instead of artificially changing my data my duplicating point, I could indeed artificially adding point with a "date" feature changing for when I know a rule change has happened. There is a pattern in my data, as some obvious rule change happen at a given date. Some rule change might be hidden though (or unknown to me), but I guess that amounts to not knowing a critical feature in my data and thus my model will be less/not accurate by definition, whether this feature is "time" or not. $\endgroup$ Aug 25, 2019 at 15:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.