google / lifetime_value Goto Github PK

In lifetime_value/notebooks/kaggle_acquire_valued_shoppers_challenge/regression.ipynb , the output units=3 and activation=relu in the last layer. While in the paper, section 3 mentioned that "three activation functions are sigmoid, identity, and softplus".

In the function to calculate ZILN loss,

The classification loss and regression loss is combined together, however, the scale of two loss is different meaning, the binary crossentropy loss is often much smaller than regression loss. Shouldn't we compensate for this difference in scale?

In addition, why do we take the negative of regression loss?

Ref:

Thank you for making the code for the loss function, metrics and example notebooks open source.

From the last paragraph in section 3:

Another key advantage of the ZILN loss is that it provides a full prediction distribution. We obtain not only the probability of returning but also the value distribution of LTV for returning customers. In addition to mean LTV prediction, the uncertainty of LTV predictions can be assessed using quantiles of a lognormal distribution as in general quantile regression.

To me it is unclear both from checking the code examples and the paper, how the LTV prediction distributions are calculated, can you get them both for a singel predictions and a group of prediction for example?

Thank you!

sklearn is being actively deprecated in favor of installing scikit-learn. This means that installing this library will cause failures if the dependencies are not changed to scikit-learn in the setup.py file.

Reference: https://pypi.org/project/sklearn/

First of all, thank you for making the code for the loss function, metrics & example notebooks open source.

I am looking forward to trying this approach out. I have one question that I am a bit confused about.

The paper mentions that the model structure, in combination with the loss function used, allows for one model to capture both objectives: churn prediction and remaining lifetime value prediction.

First, it is capable of predicting the churn probability and LTV value simultane-
ously. It reduces the engineering complexity of building a two-stage model (Vanderveld et al., 2016)
— a binary classification model to predict repeat purchase propensity, followed by a regression model
to predict the LTV of returning customers predicted in stage 1.

Indeed, the loss function returns a composite of both churn & remaining value losses:

  return classification_loss + regression_loss

However, the notebooks are split up in two, one for each separate objective.

Can one trained model indeed perform both tasks: the notebooks are split up to illustrate a comparison between the churn and remaining value tasks to two separate models?

Or are did I completely misinterpret the paper? In which it is meant that the type of model can perform both tasks, but you need two trained models: one fine-tuned for each task.

Thanks, and I am looking forward to trying this out. :)