pareto | Spatial Containers , Pareto Fronts , and Pareto Archives | Machine Learning library

Remove the line

One thing you could do would be to add a separate scaled y-axis for the geom_line like this

Change path to outfile in your table.to_csv() and see how it works. Namely,

The intuition behind the efficient greedy solution to this problem lies in the fact that a point i is dominated by point j iff x[i] > x[j] and y[i] > y[j], which implies that j must come before i when the points are ordered by either coordinate. Hence, if we traverse the points in increasing order of their x-coordinates, then the point j (if any) that dominates point i must have been traversed before point i is traversed. In other words, it is impossible for the dominating point j to come after the dominated point i in this ordering.

Thus, with this traversal order the domination problem (i.e. checking if a point is dominated by some other point) boils down to checking if we have already seen a point with a lower y-coordinate as the traversal order already enforces the x-coordinate condition. This can easily be done by checking each point's y-coordinate to the lowest (minimum) y-coordinate we have seen so far -- if the minimum y-coordinate is less than the current point i's y-coordinate then the point j with the minimum y-coordinate dominates i as x[j] < x[i] because j was seen before i.

Sorting by the x-coordinate takes O(n log n) time and checking each point (while maintaining the partial minimum y-coordinate) takes O(n) time, making the entire algorithm take O(n log n) time.

If you can get columns K and L below to show your distinct combinations, which shouldn't take more than several minutes, then the rest is pretty easy.

Column M is simply concatenating your results to show your data how you desire.

Disclaimer: Im a co-creator of the package

These probabilistic models of latent attrition are usually applied on customer cohorts because it is assumed that cohorts substantially differ from each other. Hence, you fit one separate model on each cohort. Most commonly the definition of cohorts refers to the join-date (=first transaction) but any other (further) definition is possible, such as by channel or by business segment. See also Fader and Hardie (2010) about why to cohort-wise application is important: http://www.brucehardie.com/papers/022/fader_hardie_mksc_10.pdf

But regardless of your exact cohort definition, all customers are required to have made their first transaction during the estimation period: The model is fit on the transaction data that is present in the estimation period. For all customers that have made a transaction in the estimation period, the future number of transactions are predicted for the prediction horizon which you specify as if you were standing at the end of the estimation period. All customers therefore need to have made their first transaction in the estimation period in order for the model to "know" that they exist. The model can simply not make a prediction for a customer it does not know it exists (=did not make a transaction in the estimation period).

The package could simply remove the customers that do not make their first transaction in the estimation period and only make predictions for the ones that do. However, we believe that user should be aware what happens and therefore consciously prepare the data him/herself.

I should now be able to set the estimation.split to something valid, which is what exactly

You have to specify the estimation end to a date by which all the customers in your data have already made their first transaction. If this is not the case in your data, you should split your data into cohorts defined by first transaction.

Say you have customers transactions from 2015-01-01 until 2020-01-01 and would like to split at 2017-01-01. Then you could define the first cohort as all customers that made their first transaction (=joined) between 2015-01-01 to 2015-12-31 and the second cohort from 2016-01-01 to 2016-12-31. You would create 2 separate clvdata objects for each cohort and then also fit 2 separate models. Note that you cannot create a third cohort from 2017-01-01 to 2018-01-01 with the estimation split at 2017-01-01, rather for this 3rd cohort you would have to define a later split date, say 2019-01-01.

Other cohorting windows such as 1month, 3month, 6month etc are also customary but depend on your data. Make sure to choose an estimation period long enough for the model to actually see the repeat-purchase patterns per customer (check mean interpurchase time in summary(clvdata)). For this reason, the estimation period is commonly longer than the cohorting window, ie the estimation end for a cohort is after max(customer_join_date). You might also be interested in my more in-depth answer about data preparation and cohort-wise analysis here:

https://github.com/bachmannpatrick/CLVTools/issues/101

https://github.com/bachmannpatrick/CLVTools/issues/146

However, I would also like to use the model for prediction, which requires a holdout period

After you have successfully fit the model, you can always make a prediction, also without holdout period. However, you have to specify the prediction.end argument to tell how far ahead you want to predict (number of periods or exact date). You do not have to specify the prediction.end if your data has a holdout period, because it then defaults to the holdout period. The same applies for prediction.end in plot(). To make your final CLV prediction, its actually customary to fit the model on all data in the cohort (ie without holdout period)

Keep rows in the data where the current row is "Abandoned" and the next row is not "Abandoned" for each ID. Find difference in time between every 2 rows to get time required for the caller to make another call to service after it was abandoned, take average of each of the duration to get average time.

I had to change your code a bit to make it run. CallerId and ClusterName are not part of callerData. So first run this: