mlxtend

Properties2

Type	Concept
Note created	Feb 17, 2025

mlxtend is a Python library of useful tools for machine learning that includes a lot of helper tools and algorithms that are not present in scikit-learn. One of its most interesting submodules is the one providing association rule inference.

Examples

Association Rule Learning in Python

This is a complete pipeline of association rule learning when using Python, delegating all the heavy-lifting to the mlxtend package.

Input data

The input format expects a set of transactions: sets of items that have been consumed together. For example, this list of sets is a valid starting point for this task:

sports_per_device = (
	attendance
	.groupby('identifier')
	.agg({'sport': set})
	.sport.to_list()
 
# [
#  {'Artistic Gymnastics', '3x3 Basketball'},
#  {'3x3 Basketball'},
#  {'Archery', '3x3 Basketball'},
#  {'Athletics', 'Archery'},
#  {'3x3 Basketball'},
# ...
# ]

Before feeding it to mlxtend’s algorithms, we need to format the data using a TransactionEncoder, which will format the data into a sparse matrix:

from mlxtend.preprocessing import TransactionEncoder
 
te = TransactionEncoder()
transactions = te.fit_transform(sports_per_device)

Frequent item sets

Once we have the data ready, we can compute the frequent item sets using any of the provided algorithms: Apriori, FP-Growth, or FP-Max:

from mlxtend.frequent_patterns import apriori, fpmax, fpgrowth
 
df = pd.DataFrame(transactions, columns=te.columns_)
 
frequent_itemsets = apriori(df, min_support=.0001, use_colnames=True)
## Alternatively:
# frequent_itemsets = fpgrowth(df, min_support=.0001, use_colnames=True)
# frequent_itemsets = fpmax(df, min_support=.0001, use_colnames=True)
 
#      support               itemsets
# 5   0.267259            (Badminton)
# 1   0.098270              (Archery)
# 16  0.068881              (Fencing)
# 32  0.061534           (Water Polo)
# 2   0.053268  (Artistic Gymnastics)

Association rules

Once these items are computed, we can call the assocation_rules function to perform the association rule learning and derive all the necessary metrics:

from mlxtend.frequent_patterns import association_rules
 
association_rules(
	frequent_itemsets,
	metric="confidence",
	min_threshold=0.7
)
 
# antecedents           (Breaking, Cycling BMX Freestyle)
# consequents                            (3x3 Basketball)
# antecedent support                             0.000306
# consequent support                             0.044237
# support                                         0.00023
# confidence                                         0.75
# lift                                          16.954152
# leverage                                       0.000216
# conviction                                     3.823052
# zhangs_metric                                  0.941306

Link to original