association_rules over PyEclat output

[catarinarurbano]

Hi Sebastian,

Hope you are well.

Can the association_rules be used over the output of pyEclat package, which is the frequent itemsets, to generate the association rules?

[rasbt]

In mlxtend, there are two steps for association rules: the frequent itemset generation (currently via apriori, fp-growth, or fp-max), and then the association_rules function you mentioned. Since the association_rules is tailored to the output format of apriori, fp-growth, and fp-max, the pyEclat outputs may not be directly compatible.

However, looking at the pyEclat package (which was released about 2 years ago) it seems that the author made it compatible with association_rules (implemented like 5 years ago in mlxtend). I.e., the eclat_class.df_bin output here https://github.com/jeffrichardchemistry/pyECLAT/blob/master/examples/example.ipynb looks compatible but I haven't tried.

[catarinarurbano]

Thanks for your quick response, I'm trying to use the association_rules function as the pyEclat library only outputs the frequent itemsets, thus missing the rule generation part. However, the final output for frequent itemsets is get_ECLAT_supports_all. So I think it's not in the same format, unfortunately. So, I followed this Implementation to get a dataframe as close as possible to the needed output format for the association_rules function. Nevertheless, I think it's not the same as the frequent itemset generation since the itemsets aren't in the form (item1, item2,...,itemN). Do you know any way to make it compatible?

In addition, as the Eclat algorithm only uses the support metric, the parameter support_only should be set to True, right?

Update: I found in a post TowardsDataScience that they consider get_ECLAT_supports_all as the rules itself, which doesn't make much sense...

Many thanks for your support on the matter.

[rasbt]

Oh I see now. The output looks slightly different because they use "&" instead of a comma to separate the items. Btw. the support_only=True is a bit misleading. Here, it means that the output only contains the support and other values will not be calculated. It is expected that the input dataframe only contains frequent itemsets and the support.

I just gave it a quick try and the following should work:

from pyECLAT import Example1, Example2
from pyECLAT import ECLAT
import pandas as pd

ex1 = Example1().get()
eclat_instance = ECLAT(data=ex1, verbose=True)

get_ECLAT_indexes, get_ECLAT_supports = eclat_instance.fit(min_support=0.08,
                                                           min_combination=1,
                                                           max_combination=3,
                                                           separator=' & ',
                                                           verbose=True)

df = pd.DataFrame(get_ECLAT_supports.items(),columns=['itemsets', 'support'])
df

new_column = []

for row in df['itemsets']:
    r = row.split('&')
    r = tuple(i.strip() for i in r)
    new_column.append(r)

df['itemsets'] = pd.Series(new_column)

df

from mlxtend.frequent_patterns import association_rules

association_rules(df, metric="confidence", min_threshold=0.1)

Please let me know if it works (and if it does, it would be nice to check the answer as correct -- it might be useful as a future reference for others.)

PS: I also agree with you that the TowardsDataScience article is a bit weird 😆

[catarinarurbano]

Sebastian,

thanks a lot for your help. It solved my issue and clarified my doubts.
My question regarding the support_only=True was just because the ECLAT algorithm only uses the support metric (it lacks other interestingness measures such as lift and confidence), so I think the most correct thing would be not to present the values of these measures in the output.

Thanks again for your support.
Have a great day!

[rasbt]

Yay! Glad to hear that it worked!