Add BetaGeoBetaBinomModel
#188
Conversation
ColtAllen
added
enhancement
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
| frequency: array_like | ||
| The number of purchases of customers. | ||
| recency: array_like | ||
| The time of the last, i.e. xth, purchase. |
There was a problem hiding this comment.
We can perhaps make a point that recency is discrete, i.e. int-like.
There was a problem hiding this comment.
Yeah that docstring still requires revision. Could be a good idea to add a validation check for this as well.
| """ | ||
|
|
||
| _model_name = "BG/BB" # Beta-Geometric Beta-Binomial Model | ||
| _params = ["alpha", "beta", "gamma", "delta"] |
There was a problem hiding this comment.
There was a problem hiding this comment.
Oops yes #182, sorry. And sounds good.
I see that the logp utilizes scan, which I'm not super familiar with. I see that @ricardoV94 is giving some good pointers below
| "ContNonContract", | ||
| "ParetoNBD", | ||
| ] | ||
| __all__ = ["ContContract", "ContNonContract", "ParetoNBD", "BetaGeoBetaBinom"] |
There was a problem hiding this comment.
If you add a comma after the last item black will let it stay one item per line, which I think is better for git diffs
| def _B_iter(ix): | ||
| return betaln(alpha + x, beta + t_x - x + ix) + betaln( | ||
| gamma + 1, delta + t_x + ix | ||
| ) | ||
|
|
||
| iter_result, _ = scan( | ||
| fn=_B_iter, | ||
| # outputs_info=pt.zeros(1), | ||
| sequences=pt.arange(t_recent), | ||
| ) |
There was a problem hiding this comment.
If I understand what you're trying to do, it should be something like:
| def _B_iter(ix): | |
| return betaln(alpha + x, beta + t_x - x + ix) + betaln( | |
| gamma + 1, delta + t_x + ix | |
| ) | |
| iter_result, _ = scan( | |
| fn=_B_iter, | |
| # outputs_info=pt.zeros(1), | |
| sequences=pt.arange(t_recent), | |
| ) | |
| def _B_iter(ix, alpha, beta, t_x, x, gamma, delta): | |
| return betaln(alpha + x, beta + t_x - x + ix) + betaln( | |
| gamma + 1, delta + t_x + ix | |
| ) | |
| iter_result, _ = scan( | |
| fn=_B_iter, | |
| outputs_info=[None], | |
| sequences=pt.arange(t_recent), | |
| non_sequences=[alpha, beta, t_x, x, gamma, delta], | |
| ) |
outputs_info=[None] tells Pytensor it expects a single non-recurrent output
non_sequences passes the constants explicitly to the inner function, which is a bit more clean imho
There was a problem hiding this comment.
Note that your scan can be replaced by relying on broadcasting.
| def _B_iter(ix): | |
| return betaln(alpha + x, beta + t_x - x + ix) + betaln( | |
| gamma + 1, delta + t_x + ix | |
| ) | |
| iter_result, _ = scan( | |
| fn=_B_iter, | |
| # outputs_info=pt.zeros(1), | |
| sequences=pt.arange(t_recent), | |
| ) | |
| n_dims_expr = max([alpha.ndim, beta.ndim, x.ndim, t_x.ndim, gamma.ndim, delta.ndim, t_recent.ndim]) | |
| extra_dims_needed = (n_dims_expr - t_recent.ndim) + 1 | |
| ixs = pt.shape_padright(pt.arange(t_recent), extra_dims_needed) | |
| iter_result = ( | |
| betaln(alpha + x, beta + t_x - x + ixs) | |
| + betaln(gamma + 1, delta + t_x + ixs) | |
| ) |
|
I've decided to split this into two PRs - one for the |
This PR closes #176
The
BetaGeoBetaBinomModelis the only model I'm aware of for discrete, non-contractual modeling. This PR adds a model class as well as a testing dataset and distribution block.For now this is a draft PR because many of the tests are blocked by incorrect
logpcalculations; for reference please see (5) on p4 of the research paper:https://www.brucehardie.com/papers/020/fader_et_al_mksc_10.pdf
I don't think I'm using
pytensor.scan()properly in thelogpexpression, so if someone could help me out with that I'd appreciate.Some other concerns:
purchase_heterogenietyandchurn_heterogenietyvariables for plotting the population purchase and churn rate posterior distributions, but since they won't be called until after the model is fit, they are failingflake8.