Hey Eeghor,

Hope you're doing well!

First thank you for this package it has been really useful for our company. But since your last commits, we now encounter an error when importing the library which returns us:
TypeError: 'ABCMeta' object is not subscriptable

Do you know how can we handle this error ? The code is running on a google colab with Python 3.7.12.

Thank you in advance,
Thibault

I saw that you reference articles in the read me/comments. For example, http://wnzhang.net/share/rtb-papers/data-conv-att.pdf.
How can you validate an attribution model like this?
The authors suggest that "a good MTA model should have a high degree of accuracy in correctly classifying a user as positive (with a conversion action) or negative (without a conversion action)".
What does this mean in practice? Attribution models look at all purchases, and attribute those purchases to different channels. They do not predict number of purchases. How do these authors treat these interpretive models as predictive models for this form of validation?

Which way would be suited best to compare the models? I´m looking for a way to compare all models and only use the best performing one i.e. the one with the highest model accuarcy. I was thinking of AIC to compare the dynamic models.
Is there a way to calculate the model fit within the functions or would you recommend calculating them manually?

Hi,
Maybe I haven't understood the entire thing, but in the method logistic_regression, we have:

                # expand the original data set into a feature matrix
		lists_ = []
		flags_ = []

		for i, row in enumerate(self.data.itertuples()):

			for _ in range(row.total_conversions):
				lists_.append({c: 1 for c in row.path})
				flags_.append(1)

			for _ in range(row.total_null):
				lists_.append({c: 1 for c in row.path})
				flags_.append(0)

When I do the inspection with:

print(lists_[:10], end = '\n')
print(flags_[:10], end = '\n')

I got:

# lists_[:10]
[{'e': 1, 't': 1, 'a': 1, ' ': 1, '>': 1, 'i': 1, 'o': 1, 'l': 1, 'p': 1, 'h': 1}, 
{'e': 1, 't': 1, 'a': 1, ' ': 1, '>': 1, 'i': 1, 'o': 1, 'l': 1, 'p': 1, 'h': 1}, 
{'e': 1, 't': 1, 'a': 1, ' ': 1, '>': 1, 'i': 1, 'o': 1, 'l': 1, 'p': 1, 'h': 1}, 
{'e': 1, 't': 1, 'a': 1, ' ': 1, '>': 1, 'i': 1, 'o': 1, 'l': 1, 'p': 1, 'h': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}, 
{'i': 1, 'o': 1, 't': 1, 'a': 1, ' ': 1, '>': 1}]

# flags_[:10]
[1, 0, 0, 0, 1, 1, 0, 0, 0, 0]

Finally by calling show(), I got:

linreg
a  0.449081
g  0.003764
h  0.454263
k  0.085471
n  0.003918
t  0.000376
z  0.003128

I am wondering, shouldn't I get something like:

linreg
alpha  0.449081
beta  0.003764
...

rather than spliting each alpha to a, l, p, h and a, etc?

Thank you.
Yours,
Xiaotian

Hi, very good job! I saw that you implemented ROI. Choosing which touchpoints belong to which 'abc' channels. Was it random? Like, why is 'alpha' on channel 'b'?

Hi,
In the mta.py, there is a function:

            def show(self):
		"""
		show simulation results
		"""
		res = pd.DataFrame.from_dict(mta.attribution)
		print(res)

in which the
res = pd.DataFrame.from_dict(mta.attribution)
should actually be
res = pd.DataFrame.from_dict(self.attribution)

Yours,
Xiaotian

Hi,
You did an amazing job by implementing MTA.
but I don't get the calculation behind the total conversion value and total null value.

So can you please tell me, how they have calculated or any equations?

Thank you.
Divyesh