Support and Confidence Calculation in R

Association Rule Mining in the field of data mining is a powerful tool for discovering implicit relationships between variables within a dataset. When analyzing shopping basket data, user behavior logs, or medical records, we often need to answer questions such as "If item A was purchased, is there a tendency to purchase B?" The core of this process lies in calculating two key metrics: Support and Confidence. Only by accurately understanding and correctly computing these values can we extract truly valuable business insights from massive amounts of data.
Support is a metric that measures the prevalence of a rule within an entire dataset. In simple terms, it answers the question "How frequent is this pattern?" For a given association rule $X \implies Y$ (i.e., if X occurs, then Y occurs), its support is defined as the proportion of transactions containing both X and Y out of the total number of transactions.

Mathematically, the formula for calculating the support of rule $X \implies Y$ is as follows:
$$ \text{Support}(X \implies Y) = P(X \cup Y) = \frac{\text{Number of transactions containing both } X \text{ and } Y}{\text{Total number of transactions}} $$

In practical applications, support is typically represented as a decimal, which can also be converted to a percentage. A higher support indicates that the rule is very common in the dataset and possesses high statistical stability; conversely, rules with low support may merely be coincidental phenomena lacking generalizable value.

Definition and Calculation Method of Confidence

If support describes the "frequency" of a rule occurring, then confidence describes the "probability" of the condition occurring. Confidence measures: given that X has occurred, what is the probability that Y also occurs? This is a critical metric for evaluating the utility of an association rule, directly reflecting its predictive capability.

For a rule $X \implies Y$, the formula for calculating its confidence is:
$$ \text{Confidence}(X \implies Y) = P(Y | X) = \frac{\text{Number of transactions containing both } X \text{ and } Y}{\text{Number of transactions containing } X} $$

From the formula, it can be seen that the numerator for confidence is the same as the numerator used for calculating support, but the denominator changes. The denominator is no longer the total number of transactions, but rather the number of transactions containing only the conditional part X. This means that even if a rule has high support, if the condition X itself is very common, the probability of Y occurring under the condition of X (confidence) may still be low.

Illustrative Example: Supermarket Sales Data Analysis

To understand these two concepts more intuitively, let us construct a simple supermarket sales dataset. Assume we have one week of sales records, totaling 100 orders, involving three products: Milk, Bread, and Diapers.

The statistical data shows:

Orders containing both Milk and Bread: 30.
Orders containing only Milk: 50 (Note: Logic correction based on total orders. If total orders are 100, Milk buyers are 80, Bread buyers are 40, and simultaneous buyers are 30).

Let us redefine a clear set of data for calculation purposes:

Total Orders: 100
Orders containing Milk: 80
Orders containing Bread: 40
Orders containing both Milk and Bread: 30

Based on this data, we can calculate the following association rules:

Rule 1: If Milk is purchased, then Bread is purchased ($Milk \implies Bread$)

First, calculate the support:
$$ \text{Support}(Milk \implies Bread) = \frac{30}{100} = 0.3 $$
This means that 30% of orders contain both Milk and Bread.

Next, calculate the confidence:
$$ \text{Confidence}(Milk \implies Bread) = \frac{30}{80} = 0.375 $$
This indicates that among all orders containing Milk, 37.5% also contain Bread. Although this probability seems reasonable, considering that a large proportion of people buy Milk (80%), the actual proportion of people buying Bread might be lower than average.

Rule 2: If Diapers are purchased, then Beer is purchased ($Diapers \implies Beer$)

Assume another set of data:

Total Orders: 100
Orders containing Diapers: 60
Orders containing Beer: 10
Orders containing both Diapers and Beer: 5

Calculate the support:
$$ \text{Support}(Diapers \implies Beer) = \frac{5}{100} = 0.05 $$
The support for this rule is low, indicating that this situation is not common in the dataset.

Calculate the confidence:
$$ \text{Confidence}(Diapers \implies Beer) = \frac{5}{60} \approx 0.083 $$
This confidence is very low, meaning people who buy diapers almost never buy beer. However, if we consider late-night hours, the probability of buying both diapers and beer rises significantly; this is why association rule mining requires analyzing rules in conjunction with contextual information.

Comprehensive Application and Best Practices

When building an association rule model, relying solely on intuition often leads to misjudgments. Support and Confidence must be used in combination, and reasonable thresholds should be set. The common practice is:

Minimum Support Threshold: Used to filter out combinations that are too rare in the data, reducing noise interference. For example, we can set the support threshold to 0.05, retaining only rules with a frequency of no less than 5%.
Minimum Confidence Threshold: Used to evaluate the reliability of the rule. For example, we can set the confidence threshold to 0.7, ensuring that when the condition is met, the probability of the result occurring is at least 70%.

Using the arules package or the e1071 package in R language, we can easily implement the above calculation process. In actual business scenarios, rules with high support often correspond to widespread market phenomena, while rules with high confidence are more suitable for precision marketing or risk warning. An association rule only possesses practical guiding significance when both metrics meet the expected standards.

In summary, mastering the logic of calculating Support and Confidence is the first step in conducting effective data mining. They are not only the foundation for evaluating algorithm performance but also the bridge connecting data to business decisions.