Introduction

In the vibrant cityscape of London, a city renowned for its cultural diversity and culinary richness, Indian cuisine stands out as a favorite among locals and tourists alike. The intriguing blend of spices, the variety of flavors, and the traditional cooking techniques make Indian restaurants a staple in London’s gastronomic scene. To thrive in a competitive market, restaurant managers and chefs often seek to understand customer preferences and dining patterns. This is where data-driven decision-making, leveraging techniques such as the Apriori method, becomes invaluable.

The key aim of the project is the applying the Apriori method to transactional data from an Indian takeaway restaurant in London to reveal frequent itemsets and strong association rules among the dishes ordered. The method will help to get insights into customer behavior, enabling the restaurant to tailor its menu, manage inventory efficiently, and design targeted marketing campaigns.

This project aspires to demonstrate the utility of the Apriori method in the culinary business domain, providing a roadmap for similar establishments.

Dataset and analyzing the data

Dataset used in the project consists of 13 397 transactions from the Indian takeaway restaurant located in London1.

Analyzing the transactional data:

summary(data_trans)
## transactions as itemMatrix in sparse format with
##  13397 rows (elements/itemsets/transactions) and
##  248 columns (items) and a density of 0.02247554 
## 
## most frequent items:
##    Pilau Rice    Plain Naan Plain Papadum   Garlic Naan  Onion Bhajee 
##          4721          3752          3596          2628          2401 
##       (Other) 
##         57576 
## 
## element (itemset/transaction) length distribution:
## sizes
##    1    2    3    4    5    6    7    8    9   10   11   12   13   14   15   16 
##  338  605 1611 2560 2421 1984 1377  935  617  331  227  128  101   63   39   19 
##   17   18   19   20   21   22   24   25   26   27   35   45   46   58   60 
##   10    8    4    3    5    2    1    1    1    1    1    1    1    1    1 
## 
##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##   1.000   4.000   5.000   5.574   7.000  60.000 
## 
## includes extended item information - examples:
##       labels
## 1 Aloo Chaat
## 2  Aloo Gobi
## 3 Aloo Methi
## 
## includes extended transaction information - examples:
##   transactionID
## 1           630
## 2           647
## 3           648

The most frequently ordered dishes are Pilau Rice, Plain Naan, Plain Papadum, Garlic Naan, and Onion Bhajee. The first three dishes account for 90% of all transactions. If all the baskets contained all possible products, 3,322,456 orders would have been made (13,397 * 248). Since the density is 2.25%, 74,674 orders were placed (3,322,456 * 0.0225).

itemFrequencyPlot(data_trans, topN=15, type="relative", main="Item frequency", col="#56B4E9")

The top 5 dishes are the ones mentioned above. The most frequently ordered is Pilau Rice. Also, popular dishes include: Plain Rice, Chicken Tikka Masala, Mango Chutney and others.

Association rules

The Apriori algorithm, a cornerstone in the field of data mining, provides a systematic approach to uncovering relationships between items in large transaction databases. Let’s apply the Apriori algorithm with default values (confidence = 0.8, support = 0.1).

apriori_rules <- apriori(data_trans)
## Apriori
## 
## Parameter specification:
##  confidence minval smax arem  aval originalSupport maxtime support minlen
##         0.8    0.1    1 none FALSE            TRUE       5     0.1      1
##  maxlen target  ext
##      10  rules TRUE
## 
## Algorithmic control:
##  filter tree heap memopt load sort verbose
##     0.1 TRUE TRUE  FALSE TRUE    2    TRUE
## 
## Absolute minimum support count: 1339 
## 
## set item appearances ...[0 item(s)] done [0.00s].
## set transactions ...[248 item(s), 13397 transaction(s)] done [0.00s].
## sorting and recoding items ... [13 item(s)] done [0.00s].
## creating transaction tree ... done [0.00s].
## checking subsets of size 1 2 done [0.00s].
## writing ... [0 rule(s)] done [0.00s].
## creating S4 object  ... done [0.00s].

The algorithm has not found any rules with the appropriate support and confidence levels. To generate association rules, we reduce the support coefficient to 0.02. The algorithm has already found 11 rules.

apriori_rules <- apriori(data_trans, parameter = list(supp = 0.02, conf = 0.8))
## Apriori
## 
## Parameter specification:
##  confidence minval smax arem  aval originalSupport maxtime support minlen
##         0.8    0.1    1 none FALSE            TRUE       5    0.02      1
##  maxlen target  ext
##      10  rules TRUE
## 
## Algorithmic control:
##  filter tree heap memopt load sort verbose
##     0.1 TRUE TRUE  FALSE TRUE    2    TRUE
## 
## Absolute minimum support count: 267 
## 
## set item appearances ...[0 item(s)] done [0.00s].
## set transactions ...[248 item(s), 13397 transaction(s)] done [0.00s].
## sorting and recoding items ... [63 item(s)] done [0.00s].
## creating transaction tree ... done [0.00s].
## checking subsets of size 1 2 3 4 done [0.00s].
## writing ... [11 rule(s)] done [0.00s].
## creating S4 object  ... done [0.00s].

The graphic below shows 11 association rules:

plot(apriori_rules, method="graph")

We observe a strong link between such dishes as: Mint Sauce, Mango Chutney and Plain Papadum; Onion Chutney, Mango Chutney and Plain Papadum.

plot(apriori_rules, method = "scatterplot", measure = c("support", "confidence"), shading = "lift")

Confidence coefficient

Confidence measures how often a rule is true when the conditions on the left-hand side (LHS) are satisfied.It indicates the probability that an item or item set on the right-hand side (RHS) of the rule will occur in a transaction when the items on the LHS are present. Formula: Confidence(A → B) = (Support(A ∪ B)) / (Support(A))

confidence <- sort(apriori_rules, by = "confidence", decreasing = TRUE)
inspect(head(confidence, n = 3))
##     lhs                 rhs                support confidence   coverage     lift count
## [1] {Mango Chutney,                                                                    
##      Mint Sauce,                                                                       
##      Onion Chutney}  => {Plain Papadum} 0.02104949  0.9306931 0.02261700 3.467323   282
## [2] {Mango Chutney,                                                                    
##      Mint Sauce,                                                                       
##      Pilau Rice}     => {Plain Papadum} 0.02321415  0.9014493 0.02575203 3.358375   311
## [3] {Mint Sauce,                                                                       
##      Onion Chutney}  => {Plain Papadum} 0.03112637  0.8687500 0.03582892 3.236553   417

According to the highest confidence rates, we got the three important following rules:

  • if a customer orders such a set of dishes as Mango Chutney, Mint Sauce, Onion Chutney, he will buy Plain Papadum with the probability of 93%;
  • if a customer orders such a set of dishes as Mango Chutney, Mint Sauce, Pilau Rice, he will buy Plain Papadum with the probability of 90%;
  • if a customer orders such a set of dishes as Mint Sauce, Onion Chutney, he will buy Plain Papadum with the probability of 86.9%.

Support coefficient

Support measures the frequency of occurrence of an item set in a dataset. Formula: Support(A) = (Number of transactions containing item set A) / (Total number of transactions)

support <- sort(apriori_rules, by = "support", decreasing = TRUE)
inspect(head(support, n = 3))
##     lhs                               rhs             support    confidence
## [1] {Mango Chutney, Mint Sauce}    => {Plain Papadum} 0.04650295 0.8652778 
## [2] {Mint Sauce, Pilau Rice}       => {Plain Papadum} 0.03836680 0.8263666 
## [3] {Mango Chutney, Onion Chutney} => {Plain Papadum} 0.03679928 0.8427350 
##     coverage   lift     count
## [1] 0.05374338 3.223617 623  
## [2] 0.04642830 3.078652 514  
## [3] 0.04366649 3.139633 493

The top three rules based on the support value are:

  • customers who buy Mango Chutney and Mint Sauce also order Plain Papadum in about 4.65% of the transactions;

  • customers who buy Mint Sauce and Pilau Rice also order Plain Papadum in about 3.84% of the transactions;

  • customers who buy Mango Chutney and Onion Chutney also buy Plain Papadum in about 3.68% of the transactions.

Lift coefficient

Lift quantifies how much more likely item B is purchased when item A is purchased, compared to when item B is purchased without item A. Formula: Lift(A → B) = (Confidence(A → B)) / (Support(B))

lift <- sort(apriori_rules, by = "lift", decreasing = TRUE)
inspect(head(lift, n = 3))
##     lhs                 rhs                support confidence   coverage     lift count
## [1] {Mango Chutney,                                                                    
##      Mint Sauce,                                                                       
##      Onion Chutney}  => {Plain Papadum} 0.02104949  0.9306931 0.02261700 3.467323   282
## [2] {Mango Chutney,                                                                    
##      Mint Sauce,                                                                       
##      Pilau Rice}     => {Plain Papadum} 0.02321415  0.9014493 0.02575203 3.358375   311
## [3] {Mint Sauce,                                                                       
##      Onion Chutney}  => {Plain Papadum} 0.03112637  0.8687500 0.03582892 3.236553   417

The lift values are all above 1, which suggests that the presence of the items on the left-hand side increases the likelihood of Plain Papadum being purchased. This means that these combinations are stronger than random chance and can be used to create targeted promotions. Let’s explain the rules based on the lift coefficient:

  • The combination of Mango Chutney, Mint Sauce, and Onion Chutney has a lift of 3.467, which means that the likelihood of Plain Papadum being purchased when these three items are bought together is 3.467 times greater than the likelihood of Plain Papadum being bought at all. The confidence is very high at 93.06%, indicating that when customers buy these three items, they are very likely to also purchase Plain Papadum. However, this occurs in 2.1% of transactions, showing that while it’s a strong rule, it affects a smaller segment of transactions.

  • The second rule has a slightly lower lift at 3.358, suggesting that buying Mango Chutney, Mint Sauce, and Pilau Rice together increases the chances of Plain Papadum being purchased by 3.359 times compared to its overall likelihood of being bought. The confidence is also high at 90.14%, and the support is 2.32%, indicating that this rule is reliable but slightly less frequent than Rule 1.

  • With a lift of 3.237, the third rule indicates that the purchase of Mint Sauce and Onion Chutney together is 3.237 times more likely to result in the sale of Plain Papadum than chance. It has the highest support at 3.11%, meaning it occurs more frequently in the dataset than the other two rules, and it has a high confidence of 86.88%, making it a reliable rule.

plot(apriori_rules, method = "grouped", control = list(k = 11))

Other interesting Association Rules

The top-15 frequent items also include such popular dishes as Plain Naan, Garlic Naan and Onion Bhajee. Let’s explore the association rules related to these dishes.

rules_naan <- apriori(data=data_trans, parameter=list(supp=0.01, conf = 0.5), appearance=list(default="lhs", rhs="Plain Naan"), control=list(verbose=F)) 

rules_naan_s <- sort(rules_naan, by="support", decreasing=TRUE)

inspect(head(rules_naan_s))
##     lhs                                    rhs          support    confidence
## [1] {Chicken Tikka Masala, Plain Rice}  => {Plain Naan} 0.01978055 0.5567227 
## [2] {Curry}                             => {Plain Naan} 0.01925804 0.5149701 
## [3] {Pilau Rice, Plain Rice}            => {Plain Naan} 0.01858625 0.5928571 
## [4] {Bombay Aloo, Chicken Tikka Masala} => {Plain Naan} 0.01724267 0.5539568 
## [5] {Mango Chutney, Plain Rice}         => {Plain Naan} 0.01694409 0.5550122 
## [6] {Bombay Aloo, Plain Rice}           => {Plain Naan} 0.01388371 0.5068120 
##     coverage   lift     count
## [1] 0.03553034 1.987850 265  
## [2] 0.03739643 1.838767 258  
## [3] 0.03135030 2.116873 249  
## [4] 0.03112637 1.977974 231  
## [5] 0.03052922 1.981743 227  
## [6] 0.02739419 1.809638 186

When ordering left-hand side set of products, customers will buy Plain Naan with a probability of 50-59% (as stated in the 6 rules above). In addition, the first three rules have a high level of support and lift coefficients. It means that these combinations of dishes are the common trends and can be used for promotions or cross-seling.

The graph below shows the rules for the Plain Naan dish.

plot(rules_naan, method="graph")

rules_gnaan <- apriori(data=data_trans, parameter=list(supp=0.01, conf = 0.3), appearance=list(default="lhs", rhs="Garlic Naan"), control=list(verbose=F)) 

rules_gnaan_s <- sort(rules_gnaan, by="support", decreasing=TRUE)

inspect(head(rules_gnaan_s))
##     lhs                             rhs           support    confidence
## [1] {Butter Chicken}             => {Garlic Naan} 0.02231843 0.3051020 
## [2] {Butter Chicken, Pilau Rice} => {Garlic Naan} 0.01142047 0.3355263 
##     coverage   lift     count
## [1] 0.07315071 1.555347 299  
## [2] 0.03403747 1.710444 153

Some of the popular combinations are Butter Chicken and Garlic Naan, Butter Chicken with Pilau Rice and Garlic Naan. When customers order Butter Chicken or Pilau Rice, the staff can offer Garlic Naan.

rules_bhajee <- apriori(data=data_trans, parameter=list(supp=0.01, conf = 0.4), appearance=list(default="rhs", lhs="Onion Bhajee"), control=list(verbose=F)) 

rules_bhajee_c <- sort(rules_bhajee, by="confidence", decreasing=TRUE)

inspect(head(rules_bhajee_c))
##     lhs               rhs          support    confidence coverage  lift    
## [1] {Onion Bhajee} => {Pilau Rice} 0.07942077 0.4431487  0.1792192 1.257544
##     count
## [1] 1064

In almost 8% of all transactions, customers buy Onion Bhajee with Pilau Rice. In addition, the lift coefficient is higher than 1. These results indicate that it is a frequently ordered combination of two dishes and can also be used for promotions.

Let’s analyze the most popular dish Pilau Rice as both a left-hand side product and a right-hand side product.

rules_rice_lhs <- apriori(data=data_trans, parameter=list(supp=0.01, conf = 0.3), appearance=list(default="rhs", lhs="Pilau Rice"), control=list(verbose=F)) 

rules_rice_lhs_s <- sort(rules_rice_lhs, by="support", decreasing=TRUE)

inspect(head(rules_rice_lhs_s))
##     lhs             rhs             support   confidence coverage  lift    
## [1] {Pilau Rice} => {Plain Papadum} 0.1332388 0.3780979  0.3523923 1.408614
## [2] {Pilau Rice} => {Plain Naan}    0.1291334 0.3664478  0.3523923 1.308449
##     count
## [1] 1785 
## [2] 1730

In 13% of transactions, a customer buys Plain Papadum or Plain Naan after having Pilau Rice in their basket. But the likelihood of such situations is not very high (38% and 37%).

rules_rice_rhs <- apriori(data=data_trans, parameter=list(supp=0.01,conf = 0.6), appearance=list(default="lhs", rhs="Pilau Rice"), control=list(verbose=F)) 

rules_rice_rhs_s <- sort(rules_rice_rhs, by="support", decreasing=TRUE)

inspect(head(rules_rice_rhs_s))
##     lhs                                      rhs          support    confidence
## [1] {Chicken Tikka Masala, Plain Naan}    => {Pilau Rice} 0.04030753 0.6257242 
## [2] {Chicken Tikka Masala, Plain Papadum} => {Pilau Rice} 0.03687393 0.6159601 
## [3] {Peshwari Naan, Plain Papadum}        => {Pilau Rice} 0.02396059 0.6056604 
## [4] {Bombay Aloo, Chicken Tikka Masala}   => {Pilau Rice} 0.02015377 0.6474820 
## [5] {Korma, Plain Papadum}                => {Pilau Rice} 0.01970590 0.6226415 
## [6] {Onion Bhajee, Peshwari Naan}         => {Pilau Rice} 0.01910876 0.6139089 
##     coverage   lift     count
## [1] 0.06441741 1.775647 540  
## [2] 0.05986415 1.747938 494  
## [3] 0.03956110 1.718710 321  
## [4] 0.03112637 1.837390 270  
## [5] 0.03164888 1.766899 264  
## [6] 0.03112637 1.742118 256

Analyzing Pilau Rice as a right-hand side product, we notice that the most frequently ordered combination is with Chichen Tikka Masala and Plain Naan (540 transactions).

The graph below shows the rules for the Pilau Rice dish.

plot(rules_rice_rhs, method="graph")

Conclusion

The Apriori method was used to study the behavior of customers of an Indian takeaway restaurant in London. Comparing the association rules by confidence, support, and lift coefficients, we came to the following conclusions:

In addition, there are many trendy combinations of Pilau Rice with Chicken Tikka Masala; Mango Chutney and Plain Papadum, etc. Curry with Plain Naan; Bombay Aloo, Chicken Tikka Masala with Plain Naan; Butter Chicken with Garlic Naan; or Onion Bhajee with Pilau Rice are interesting association rules with a high lift indicator. The restaurant may design the menu to highlight these combinations that lead to a higher lift coefficient, encouraging customers to try these combinations. By leveraging these rules, restaurants can potentially increase their average order value and improve customer satisfaction by offering popular item combinations.

Dataset source


  1. Food Orders of Indian takeaway restaurant in London↩︎