Market Basket Analysis Using Association Rules in R

learn to use association rules in R, refer to the book: Machine Learning with R

- Basic idea:

Apriori property: all subsets of a frequent itemset must also be frequent.to reduce the association rule search space.

Whether or not an association rule is deemed interesting is determined by two statistical measures: support and confidence measures.

Measuring rule interest –support and confidence:

The support of an itemset or rule measures how frequently it occurs in the data:

  

A rule's confidence is a measurement of its predictive power or accuracy:

- Practice:

Step 1 – collecting data

install "arules" R package and use the data “Groceries 

Step 2 – exploring and preparing the data

a)  summary the Groceries:

library(arules)
library(Matrix)
summary(Groceries)

outcome:

transactions as itemMatrix in sparse format with

 9835 rows (elements/itemsets/transactions) and

 169 columns (items) and a density of 0.02609146

most frequent items:

 whole milk  other vegetables       rolls/buns            soda           yogurt          (Other)

      2513        1903             1809             1715             1372            34055

element (itemset/transaction) length distribution:

sizes

1    2    3    4    5    6    7    8    9   10   11   12   13   14   15   16   17   18   19 

2159 1643 1299 1005  855  645  545  438  350  246  182  117  78   77   55   46   29   14   14

20   21   22   23   24  26   27   28   29   32

9    11    4    6   1    1    1    1    3    1  

 Min.    1st Qu.  Median  Mean  3rd Qu.  Max.

 1.000   2.000   3.000   4.409   6.000   32.000

includes extended item information - examples:

       labels   level2           level1

1   frankfurter sausage meat and sausage

2   sausage sausage meat and sausage

3   liver loaf sausage meat and sausage

b) inspect

inspect(Groceries[1:5])

outcome:

items                                                               

[1] {citrus fruit,semi-finished bread,margarine,ready soups}            

[2] {tropical fruit,yogurt,coffee}                                      

[3] {whole milk}                                                        

[4] {pip fruit,yogurt,cream cheese ,meat spreads}                       

[5] {other vegetables,whole milk,condensed milk,long life bakery product}

c) To examine a particular item

itemFrequency(Groceries[, 1:3])

d) Plot fuction

1) Support>=0.10 frequency items plot

itemFrequencyPlot(Groceries, support = 0.1)

2) TOP 20 frequency items plot

c) plotting the sparse matrix

image(Groceries[1:5])

image(sample(Groceries,100))

Step 3 – training a model on the data

a) train model

Groceriesrules<-apriori(Groceries,parameter = list(support=0.01,confidence=0.25, minlen=2))
inspect(Groceriesrules[1:3])

Outcome:

b) Set confidence as 0.25 then

Groceriesrules<-apriori(Groceries,parameter =
list(support=0.006,confidence=0.25, minlen=2))

outcome:

See the rules:

Groceriesrules

Outcome:

set of 463 rules

Step 4 – evaluating model performance

a) See the summary of rules:

summary(Groceriesrules)

outcome:

Lift: how much more likely one item or itemset is purchased relative to its typical rate of purchase.

e.g. if lift(milk → bread) is greater than one, it implies that the two items are found together more often than one would expect by chance.

b) rule mining

inspect(Groceriesrules[1:3])

outcome:

The first rule can be read in plain language as, "if a customer buys pot plants, they will also buy whole milk." With support of 0.007 and confidence of 0.400, we can determine that this rule covers 0.7 percent of the transactions andis correct in 40 percent of purchases involving potted plants. The lift value tells ushow much more likely a customer is to buy whole milkrelative to the average customer, given thathe or she bought a potted plant. Since we know that about 25.6 percent of the customers bought whole milk (support), while 40 percent of the customers buying a potted plant bought whole milk (confidence), we can compute the lift value as 0.40 / 0.256 = 1.56, which matches the value shown.

Step 5 – improving model performance

To sort rules due to different criteria:

a) find the rules with high support, confidence, or lift

inspect(sort(groceryrules, by = "lift")[1:5])

outcome:

The first rule, with a lift of about 3.96, implies that people who buy herbs are nearly four times more likely to buy root vegetables than the typical customer.

Order by desc:

 parameterdecreasing = FALSE

b) Taking subsets of association rules

investigate whether berries are often purchased with other items:

berryrules <- subset(Groceriesrules, items %in% "berries")
inspect(berryrules)

Outcome:

• When you need “berries” only appear in the left or right side, replace “items” with “lhs/rhs”.
• Chose rules contain “berries” or “yogurt”, replace “%in% "berries"” with “items %in%c("berries", "yogurt”)”

• Additional operators are available for partial matching (%pin%) and complete matching (%ain%)
• “confidence > 0.50” would limit you to the rules with confidence greater than 50 percent
• Matching criteria can be combined with the standard R logical operators such as and (&), or (|), and not (!)
  

Saving association rules to a file or data frame

a) Save rules as .csv format

setwd("C:\\myprojectR\\Groceries")
write(Groceriesrules, file = "Groceriesrules.csv",
       sep = ",", quote = TRUE, row.names = FALSE)

b) Save rules as R data frame

groceryrules_df <- as(Groceriesrules, "data.frame")

see the results:

str(groceryrules_df)

You might choose to do this if you want to perform additional processing on the rules or need to export them to another database