Association rule is a method for pattern mining. In this article, we perform an association rule analysis of some demo data.

The Problem Defined

Suppose we own a store called KIOSK. Here at KIOSK, we sell 4 different things.

  1. Milk
  2. Croissant
  3. Coffee
  4. Fries

We need to know what items are associated with each other when the customers are buying.

We have collected the following data. Beware that this small amount of data might not be enough for a real-world problem.

INDEX Items
1 croissant, milk
2 coffee, croissant
3 coffee, croissant
4 coffee, croissant, milk
5 coffee, milk
6 fries, milk
7 coffee, croissant, fries
8 croissant, fries
9 croissant, milk
10 croissant, fries, milk
11 coffee, croissant, milk

The Rule, Support, and Confidence

Association rule should contain a rule, support, and confidence.

In this rule, the first part means that if the customers are buying milk they buy croissant too. Support is the probability of the records with both milk and croissant (2) among all the records (5). Confidence is the conditional probability that the records have croissant given milk in the record.

Then we define two thresholds that tell us whether this rule is valid.

  1. Support threshold
  2. Confidence threshold

If the support and confidence are both larger than the thresholds, we say the rule is valid.

Sometimes it is easier to deal with support counts instead of support probability itself.

It is, however, very hard to calculate association rules with many items without a smart idea. Suppose we have 10 items, the total number of possible association rules is 10!=3628800. It is not practical to explore all the possible combinations.

Keywords

  1. itemset: a set that contains items such as milk, coffee, etc.
  2. k-itemset: an itemset with k items
  3. frequent: an itemset is frequent if it’s support count is larger than the threshold (minimum support count). In this article, we define the minimum support count to be 2.

Algorithms

As long as we have the frequent itemsets, it is easy to calculate the association rule. We need some algorithms to calculate the frequent itemsets. For example, we could use some algorithms to

  1. eliminate some of the possible combinations when generating a pool of rules to be calculated,
  2. or enable parallel computations,
  3. or better search algorithms.

Apriori Method

Apriori method is based on a simple rule: Subsets of a frequent k-itemset should also be frequent.

For example, if {coffee, croissant} is a frequent itemset, {coffee} and {croissant} must be frequent itemsets. This rule allows us to calculate less combination.

The following calculations assumed a minimum support count of 2.

Identifying Frequent Itemsets

To find all the frequent 1-itemsets, we simply explore all the possible 1-itemsets.

  1. Construct itemsets and calculate the support count of all the 1-itemsets.

    1-itemset support count
    {coffee} 6
    {croissant} 9
    {fries} 3
    {milk} 7
  2. Drop all 1-itemsets that have support count less than 2. In our example, all the 1-itemsets are frequent.

    1-itemset support count
    {coffee} 6
    {croissant} 9
    {fries} 3
    {milk} 7

To find all the frequent 2-itemsets, we construct the 2-itemsets using frequent 1-itemsets only.

  1. Construct 2-itemsets based on the frequent 1-itemsets. We choose two 1-itemsets from the frequent 1-itemsets and find the union.

    2-itemset support count
    {coffee, croissant} 5
    {coffee, fries} 1
    {coffee, milk} 3
    {croissant, fries} 3
    {croissant, milk} 5
    {fries, milk} 2
  2. Drop itemsets with support count less than 2. We get the frequent 2-itemsets.

    2-itemset support count
    {coffee, croissant} 5
    {coffee, milk} 3
    {croissant, fries} 3
    {croissant, milk} 5
    {fries, milk} 2

To find all the frequent 3-itemsets, we construct 3-itemsets using frequent 2-itemsets.

  1. Construct 3-itemsets by joining 2-itemsets that have 3-2=1 item in common. This is done in two steps. First, we find 2-itemset that has 3-2=1 item in common, e.g., {coffee, croissant}, {coffee, fries}, {coffee, milk}. Secondly, we choose two of them and union, e.g. {coffee, croissant} union {coffee, fries} = {coffee, croissant, fries}.

    3-itemset support count
    {coffee, croissant, fries} 1
    {coffee, croissant, milk} 2
    {coffee, fries, milk} 0
    {croissant, fries, milk} 1
  2. Drop itemsets with support count less than 2.

    3-itemset support count
    {coffee, croissant, milk} 2

Our calculation stops here since any 4-itemset would require at least two frequent 3-itemsets which we have only one.

However, it is straightforward to construct the 4-itemsets if we had frequent 3-itemsets. We find 3-itemsets that have 4-2=2 items in common. Then we choose any two itemsets from them and union.

Generating Association Rules and Calculating the Confidence

Association rules are generated by splitting the itemsets into two subsets without intersections. Suppose we are constructing association rules from {coffee, croissant, milk}. We follow three steps.

  1. Find the subsets. Here we have $C^3_1=3$ combinations.

    1. {coffee}, {croissant, milk}
    2. {coffee, croissant}, {milk}
    3. {coffee, milk}, {croissant}
  2. The rules will be from one of the subsets to another, e.g., {coffee} to {croissant, milk}. We have 6 in total.
  3. The confidence of the rule $\text{ {coffee} } \Rightarrow \text{ {croissant, milk} }$ is

Pattern-growth Approach

The problem with the Priori method is that the search is still querying the table again and again on each itemset build. Pattern-grwoth approach is different since it builds a tree, with the name frequent-pattern tree (FP-tree) with a few table scans.

Simply put, the pattern-growth approach will build a tree to represent the relations between the items. Then the association rules will be calculated based on the tree.

Vertical Data Format

The table

INDEX Items
1 croissant, milk
2 coffee, croissant
3 coffee, croissant
4 coffee, croissant, milk
5 coffee, milk
6 fries, milk
7 coffee, croissant, fries
8 croissant, fries
9 croissant, milk
10 croissant, fries, milk
11 coffee, croissant, milk

becomes vertical data format if we have

Item Item ID
coffee {2,3,4,5,11}
croissant {1,2,3,4,7,8,9,10,11}
fries {6,7,8}
milk {1,4,5,6,9,10,11}

Finding frequent itemsets becomes set operations. For examples, to calculate the support count of {coffee, croissant}, we will find the intersection of the two item id sets associated with coffee and croissant, i.e.,

\begin{equation} \{2,3,4,5,11\} \cap \{1,2,3,4,7,8,9,10,11\} = \{ 2,3,4,11 \} \end{equation}

The Right Way of Using Association Rules

From the previous calculations, we find the association rule

\begin{equation} \text{coffee} \Rightarrow \text{croissant} [ \text{support} = 45\%,\text{confidence} = 83\% ] \end{equation}

We know that the support for croissant is 82%. The fact that the confidence of this association rule is 83% makes this association rule not so interesting. Even without buying coffee, 82% of the records already have croissant in it. Buying coffee first doesn’t really increase the probability of purchasing croissant significantly.

Sometime the confidence might even be lower than probability of purchasing croissant, in which case coffee would decrease the probability of purchasing croissant.

This example shows that we should combine association rules with some other correlation measures. One of such measures is the lift.

Lift is defined as

\begin{equation} \text{Lift}(\text{coffee}, \text{croissant} ) = \frac{ P( \text{coffee} \cup \text{croissant} ) }{ P(\text{coffee}) P(\text{croissant}) } \end{equation}

When we have the lift equals 1, the two events are independent. If lift is greater than 1, the probability of the union of coffee and croissant is larger than the assumed independent probability, which means some items in the two sets are correlated indicating a positive correlation between the two sets. If lift is less than 1, the two sets are negatively correlated.

Another measure is chi-square analysis.