3
$\begingroup$

Google Analytics allows me to collect data about every web-session. For simplicity, let's assume for each user, we collect the number of pages and time spent on site for each session:

user_id visit_id page_views time_spent result
1       1        10         100        0
1       2        31         510        0
1       3        1          10         1

How would you model this data? What I would like the ML algorithm to do:

  1. Extract as much information as possible
  2. Have a flexible number of inputs (e.g. the number of sessions can go to infinity)

What I can think of:

  1. Aggregate the data per user e.g. average page_views or total page_views and feed it into a general algorithm e.g. random forrest (but I lose information with aggregation)
  2. Use LSTM and feed at most last 3 visits (will also lose information, but would this perform better than aggregation?)

Goal: To build a predictive model to analyse all user sessions and make a prediction whether the person will convert or not.

$\endgroup$
  • $\begingroup$ Your goal is so vague and unclear. The title and introduction is about Google Analytics data but you end with a goal about predicting whether or not a person will convert or not. What is 'result'? What does 'extract as much information as possible' mean? Is your page a site to convert a person's ideology or religion? If so what are the people doing on your site? Are they filling out forms or taking surveys? Is the data dynamic? $\endgroup$ – Brian O'Donnell Mar 7 '18 at 22:48
1
$\begingroup$

I understand that in your example you are interested in modelling the outcome of the 'result' column.

One easy model I would suggest is to model it using the Bernoulli distribution (https://en.wikipedia.org/wiki/Bernoulli_distribution) with the probability of success p.

Then you can model p with something like this

x = a + b * log(page_views) + c * log(time_spent) + e

p = exp(x) / (1+exp(x))

where e is normally distributed, e ~ N(0, sigma^2) (or simply centered around zero).

a, b and c are parameters that you can estimate.

I.e. the probability of success (conversion) is modeled as a sigma function of a certain variable that depends on page_views and time_spend. You can also add squares (and higher powers) of page_views and time_spend in the equation of x (i.e. upto a certain threshold page_views can have a positive effect on conversion and then a negative one, then again a positive effect).

Also reading about logistic regression should put you on the right track: https://en.wikipedia.org/wiki/Logistic_regression

$\endgroup$
0
$\begingroup$

I would model this data as a 3d-tensor (user,timestep,features) [organization depends on which DL Framework you use] for your input data. Also for the output data is a 3d-tensor (user,timestep,result) appropriate.

The next step would ne tp train a LSTM or CNN model to predict the result (what requires a lot of data)(Would be my first choice). If you have less data, try out logistic regression as supposed by the other answer.

Good luck!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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