svm - Using machine learning to estimate likelihood of an even occurrence given a stream of data -

i have stream of data (e.g. 3d position) generating system looks like:

(pos1, time1) (pos2, time2) (pos3, time3) ...

i want use machine learning technique estimate likelihood (or detect) of particular event given stream of data. have done:

1. i've tagged data @ every frame yes if event occurred @ frame, otherwise set no.

(pos1, time1, no) (pos2, time2, yes) (pos3, time3, no) ...(posk, timek, yes)...

2. set window length l train model giving l consecutive frames , corresponding tag set tag of last element on window:

(pos1, pos2, pos3, no) (pos2, pos3, pos4, no) (pos3, pos4, pos5, no) ... (posk-2, posk-1, posk, yes) ...

3. finally, trained model set of that.
4. for testing, concatenate l consecutive frames , ask model find corresponding tag set of data (e.g. yes or no).

i realize occurrence of "no" lot more frequent "yes". because system on idle state , have no event. affects on training.

could give me hints: 1) type of machine learning model best fit problem. 2) @ moment classifying output either "yes" or "no" have probability of occurrence of event @ anytime. kind of model suggest?

thanks

i think there 2 questions, here: how build dataset, , predictor use.

for building dataset, @ time point i, make sure choose ℓ instances happening before i (the phrasing in question made seem you're choosing 1 including i). label of outcome should 1 @ i, though. after all, you're attempting predict future based on present, no? predicting present based on present rather easy.

another point how choose ℓ, or whether choose single ℓ. note if choose number of different values of ℓ, multivariate model.

finally, question directly asked predictor use. wide answer without knowing dataset (and playing it). might want read bias-variance tradeoff see why there no "best" predictor problem.

having said that, i'd suggest start logistic regression simple , robust classifier outputs probabilities (as asked).