I've just started to examine the subject. Excuse me, if my questions sound stupid.

I have some statistics of football matches (injuries, corner kicks, yellow cards) with time references (a certain minute of the match). I want to make the model of neural network, which would bind football match to one of the patterns, depending on the data provided. I have a few questions:

How to create a data set with a time reference?

How to normalize the data, should i use "Min-max normalization" or something else?

What type of neural network architecture is optimal to solve this task?


closed as too broad by Franck Dernoncourt, mdewey, Michael Chernick, gung, John Jan 20 '17 at 23:13

Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. Avoid asking multiple distinct questions at once. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • 1
    $\begingroup$ What pattern are to trying to predict using NN? $\endgroup$ – Ujjwal Kumar Jan 20 '17 at 12:22
  • $\begingroup$ For example, teams play with approximately equal skill. So, I need to determine on the basis of data, which pattern is closer to this match: 1. One of the teams is putting pressure and we are expect a goal. 2. The game is equal and the probability of goals is small. $\endgroup$ – Sid007 Jan 20 '17 at 13:05
  • $\begingroup$ How would you quantify "pressure"? I think you should start with a simpler problem, say just predicting win or loss. This would give you ideas about how to develop features for modelling. Then try to move on to harder problems like the one you're mentioning $\endgroup$ – Ujjwal Kumar Jan 20 '17 at 14:14
  • $\begingroup$ I will quantify "pressure" by live data of match (shots on target, shots off target, dangerous attack, ball safe). But predicting win or loss is not simpler, because this is the same problem, but I need to know from which team expect goal. $\endgroup$ – Sid007 Jan 20 '17 at 15:34

Lets formulate the problem first, so that it can be addresses in an easier manner.

  1. Lets say we discretise match-time, by cutting it into windows of (say) 10 minutes each. Our goal is to be able to predict whether the next 10-minute window would contain a goal by our team or not.

Now that the target is defined, lets generate some features to predict the event.

"Frequency" variables.

  1. So, now we have some events happening some number(or zero) of times, in every of the 9 windows of time, events like attempt, foul, penalty, goal (for both teams). Count of such events in time-window would be their "frequency".

"Recency" variables:

  1. Time passed since last time goal, foul, penalty etc happened.

Apart from frequency and recency, one can also get some continuous features like how much % of time, for this window, did "our" team keep the ball in their control? We can also create cumulative versions of frequency and continuous variables. Some other features like how much time to half-time, how much till full-time, can also be created.

Now using all these features, we can try to predict our stated goal.

Note that NNs require a lot of data to work with large number of variables, and they often overfit. This problem can easily be solved by simple classification algorithms (given it is solvable).


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