search

kunzaatko / pm_mode

1 stars 0 forks source link

Feature based on ratio vs difference #28

Closed Many98 closed 3 years ago

Many98 commented 3 years ago

@kunzaatko @srameon1 @KholDani Consider example. Lets have 2 teams , team A, team B and consider characteristics (without loss of generality) related to 5 played matches in past (or 5 months).

Team A: goals_scored=10, goals_conceded=2 Team B: goals_scored=20, goals_conceded=4

Goals scored somehow represents attack strength of team (the more the better). Similarly goals conceded represents defensive strength of team (the less the better). So in example above has team B relatively higher attack strength but has slightly weaker defensive strengths. I would intuitively say that team B has higher chance to win if there would be sometime match between them.

Now lets look at features: difference:(GS-GC): team(A)=10-2=8; team(B)=20-4=16 ratio:(GS/GC): team(A)=10/2=5, team(B)=20/4=5 ratio:(GS/[GS+GC], implemented by @srameon1 ): team(A)=10/(10+2)=5/6, team(B)=20/(20+4)=5/6

Both ratio characteristics treat team A and B as equivalent in terms of some overall strength. On the other hand only difference taken into account what is intuitively hidden in GS, GC features that team B is more probable to win. Also ratio characteristics are prone to have some extreme values as infty when for instance GC=0 or (GC+GS)=0 etc.

What do you think about that?

Many98 commented 3 years ago

We agreed to test them and make a decision based on that.

kunzaatko commented 3 years ago

I am working on that ;)