How to interpret predictions from a Neural Network

Question

I have trained a Neural Network to make prediction of outcome(Win/Lose) of a hockey game based on a few metrics. The data I have been feeding it looks like this:

Each row represents a team in one game, so two specific rows make a match.

Won/Lost   Home   Away  metric2 metric3 metric4 team1 team2 team3 team4
1          1      0      10      10      10      1     0     0      0
0          0      1      10      10      10      0     1     0      0
1          1      0      10      10      10      0     0     1      0
0          0      1      10      10      10      0     0     0      1




The predictions from the NN looks like this.
[0.12921564 0.87078434]
[0.63811845 0.3618816 ]
[5.8682327e-04 9.9941313e-01]
[0.97831124 0.02168871]
[0.04394475 0.9560553 ]
[0.76859254 0.23140742]
[0.45620263 0.54379743]
[0.01509337 0.9849066 ]

I believe I understand that the first column is for Lost(0), and second is for Won(1), but what I don't understand is: Who won against who? I don't now what to make of these predictions, do they even mean anything to me this way?

score 1 · Answer 1 · answered Dec 29 '18 at 06:04

1

According to the Data Set you show here, it seems that the results of the network would show the probability of wining or losing a team in a match depending on the Race Host. I think you should add one more feature to your data set which shows the rival team in the match if you want your network to show the probability of wining the game against other team and the hosting situation( And if hosting is not important for you then you should remove Home and Away columns).

answered Dec 29 '18 at 06:04

pouyan

3,445
4
26
44

Is there a more effective way to add the away-team in the dataset than as a unitvector? Or will I have to add 30 columns for my 30 teams? – MisterButter Dec 29 '18 at 06:41
you can just create one more column to show the away-team ids and then programmatically convert it to binary columns. – pouyan Dec 29 '18 at 06:50
Alright, I will test that. Thank you for the input! – MisterButter Dec 29 '18 at 06:52

Venkatachalam · Accepted Answer · 2018-12-29T10:37:01.437

Let us take first two rows of your dataset,

Won/Lost   Home   Away  metric2 metric3 metric4 team1 team2 team3 team4
1          1      0      10      10      10      1     0     0      0
0          0      1      10      10      10      0     1     0      0

#predictions 
[0.12921564 0.87078434]
[0.63811845 0.3618816 ]

Team 1 played a game in its home and won the match. Model prediction also aligns with it because it has assigned high probability in the second column, which is the probability of winning as you mentioned.

Similarly Team 2 played a game away and lost the match. Model prediction aligns here as well!

you just mentioned that two specific rows make a match but with available information we cannot say who played with whom. Its just a model to predict the probability of winning for a particular team independently.

EDIT:

Assuming that you have data like this!

gameID          Won/Lost   Home   Away  metric2 metric3 metric4   team1 team2 team3 team4
2017020001         1          1      0      10      10      10      1     0     0      0
2017020001         0          0      1      10      10      10      0     1     0      0

You could transform the data as follows, which can improve the model.

Won/Lost  metric2 metric3 metric4 h_team1 h_team2 h_team3 h_team4 a_team1 a_team2 a_team3 a_team4
1            10      10      10      1       0        0      0         0      1        0      0

Note: won/Lost value would be for home team, which is mentioned by h_team.

I excluded it from the table, but in the dataset I have a column that is game_id, where the two rows that represent a specific match has the same id in rhe form of 2017020001. If I where to add that column, whould I be able to make predictions of team X winning over team Y? — MisterButter, Dec 29 '18 at 06:38
Directly feeding GameID might be helpful, but combining two rows for each game into one, would improve the model. Because the model get to know the details of both the teams play and can make better prediction — Venkatachalam, Dec 29 '18 at 10:34
A will certainly test combining the rows as showed above. Think that will be the best solution and also the easiest way to interpret the predictions that the model makes — MisterButter, Dec 29 '18 at 12:27
do you perhaps happen to know how to make that transformation in python? Can't seem to find any good sources on it. — MisterButter, Dec 29 '18 at 20:55

How to interpret predictions from a Neural Network

2 Answers2

Linked