In this blog post I will give an introduction to the **discovery of periodic patterns in data**. Mining **periodic patterns** is an important data mining task as patterns may periodically appear in all kinds of data, and it may be desirable to find them to understand data for taking strategic decisions.

For example, consider customer transactions made in retail stores. Analyzing the behavior of customers may reveal that some customers have periodic behaviors such as buying some products every week-end such as *wine* and *cheese*. Discovering these patterns may be useful to promote products on week-ends, or take other marketing decisions.

Another application of **periodic pattern mining** is stock market analysis. One may analyze the price fluctuations of stocks to uncover periodic changes in the market. For example, the stock price of a company may follow some patterns every month before it pays its dividend to its shareholders, or before the end of the year.

In the following paragraph, I will first present the problem of discovering** frequent periodic patterns, **periodic patterns that appear frequently. Then, I will discuss an extension of the problem called **periodic high-utility pattern mining **that aims at discovering profitable patterns that are periodic. Moreover, I will present open-source implementations of these algorithms.

**The problem of mining frequent periodic patterns**

The problem of discovering **periodic patterns** can be generally defined as follows. Consider a database of transactions depicted below. Each transaction is a set of items (symbols), and transactions are ordered by their time.

Here, the database contains seven transactions labeled T1 to T7. This database format can be used to represent all kind of data. However, for our example, assume that it is a database of customer transactions in a retail store. The first transaction represents that a customer has bought the items *a* and * c* together. For example, *a *could mean an apple, and *c* could mean *cereals*.

Having such a database of objects or transactions, it is possible to find **periodic patterns**. The concept of **periodic patterns** is based on the notion of **period**.

A **period** is the time elapsed between two occurrences of a pattern. It can be counted in terms of time, or in terms of a number of transactions. In the following, we will count the length of periods in terms of number of transactions. For example, consider the itemsets (set of items) {*a,c*}. This itemset has five periods illustrated below. The number that is used to annotate each period is the period length calculated as a number of transactions.

The first period of {*a,c*} is what appeared before the first occurrence of {*a,c*}. By definition, if {*a,c*} appears in the first transaction of the database, it is assumed that this period has a length of 1.

The second period of {*a,c*} is the gap between the first and second occurrences of {*a,c*}. The first occurrence is in transaction T1 and the second occurrence is in transaction T3. Thus, the length of this period is said to be 2 transactions.

The third period of {*a,c*} is the gap between the second and third occurrences of {*a,c*}. The second occurrence is in transaction T3 and the third occurrence is in transaction T5. Thus, the length of this period is said to be 2 transactions.

The fourth period of {*a,c*} is the gap between the third and fourth occurrences of {*a,c*}. The third occurrence is in transaction T5 and the fourth occurrence is in transaction T6. Thus, the length of this period is said to be 2 transactions.

Now, the fifth period is interesting. It is defined as the time elapsed between the last occurrence of {*a,c*} (in T6) and the last transaction in the database, which is T7. Thus, the length of this period is also 1 transaction.

Thus, in this example, the **list of period lengths** of the pattern {*a,c*} is: 1,2,2,1,1.

Several algorithms have been designed to discover periodic patterns in such databases, such as the *PFP-Tree, MKTPP, ITL-Tree, PF-tree*, and *MaxCPF* algorithms. For these algorithms, a pattern is said to be periodic, if:

(1) it appears in at least *minsup* transactions, where *minsup *is a number of transactions set by the user,

(2) and the pattern has no period of length greater than a *maxPer* parameter also set by the user.

Thus, according to this definition if we consider *minsup* = 3 and *maxPer* =2, the itemset {*a,c*} is said to be periodic because it has no period of length greater than 2 transactions, and it appears in at least 3 transactions.

This definition of a **periodic pattern** is, however, too strict. I will explain why with an example. Assume that *maxPer* is set to 3 transactions. Now, consider that a pattern appears every two transactions many times but that only once it appears after 4 transactions. Because the pattern has a single period greater than *max**Per, *this pattern would be automatically be deemed as non periodic although it is in general periodic. Thus, this definition is too strict.

Thus, in a recent paper, we proposed a solution to this issue. We introduced two new measures called the *average periodicity *and the *minimum periodicity*. The idea is that we should not discard a pattern if it has a single period that is too large but should instead look at how periodic the pattern is on average. The designed algorithm is called PFPM**. **It discovers periodic patterns using a more flexible definition of what is a **periodic pattern**. **A pattern is said to be periodic pattern if:**

(1) the average length of its periods denoted as avgper(X) is not less than a parameter *minAvg* and not greater than a parameter *maxAvg*.

(2) the pattern has no period greater than a maximum *maxPer*.

(3) the pattern has no period smaller than a minimum *minPer.*

This definition of a **periodic pattern** is more flexible than the previous definition and thus let the user better select the periodic patters to be found. The user can set the minimum and maximum average as the main constraints for finding periodic patterns and use the minimum and maximum as loose constraints to filter patterns having periods that vary too widely.

Based on the above definition of periodic patterns, **the problem of mining all periodic patterns** **in a database** is to find all periodic patterns that satisfy the constraints set by the user. For example, if *minPer* = 1, *maxPer* = 3, *minAvg* = 1, and *maxAvg* = 2, the 11 periodic patterns found in the example database are the one shown in the table below. This table indicates the measures of support (frequency), minimum, average and maximum periodicity for each pattern found:

As it can be observed in this example, the *average periodicity* can give a better view of how periodic a pattern is. For example, considers the patterns {*a,c*} and {*e*}. Both of these patterns have a largest period of 2 (called the maximum periodicity), and would be considered as equally periodic using the standard definition of a periodic pattern. But the average periodicity of these two patterns is quite different. The average periodicity measure indicates that on average {*a,c*} appears with a period of 1.4 transactions, while {e} appears on average with a period of 1.17 transaction.

**Discovering periodic patterns using the SPMF open-source data mining library**

An implementation of the PFPM algorithm for discovering periodic patterns, and its source code, can be found in the **SPMF data mining software**. This software allows to run the algorithm from a command line or graphical interface. Moreover, the software can be used as a library and the source code is also provided under the GPL3 license. For the example discussed in this blog post, the input database is a text file encoded as follows:

3 153 5 1 2 43 5 2 43 1 43 5 13 5 2

where the numbers 1,2,3,4,5 represents the letters *a*,*b*,*c,d,e. *To run the algorithm from the source code, the following lines of code need to be used:

String output = "output.txt"; String input = "contextPFPM.txt"; int minPeriodicity = 1; int maxPeriodicity = 3; int minAveragePeriodicity = 1; int maxAveragePeriodicity = 2; // Applying the algorithm AlgoPFPM algorithm = new AlgoPFPM(); algorithm.runAlgorithm("input.txt", "output.txt", minPeriodicity, maxPeriodicity, minAveragePeriodicity, maxAveragePeriodicity);

The output is then a file containing all the **periodic patterns** shown in the table.

2 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 2 5 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 2 5 3 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 2 3 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 4 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 4 3 #SUP: 3 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.75 1 #SUP: 4 #MINPER: 1 #MAXPER: 2 #AVGPER: 1.41 3 #SUP: 4 #MINPER: 1 #MAXPER: 2 #AVGPER: 1.45 #SUP: 5 #MINPER: 1 #MAXPER: 2 #AVGPER: 1.1666666666666667 5 3 #SUP: 4 #MINPER: 1 #MAXPER: 3 #AVGPER: 1.4 3 #SUP: 6 #MINPER: 1 #MAXPER: 2 #AVGPER: 1.0

For example, the eighth line represents the pattern {*a*,*c*}. It indicates that this pattern appears in four transactions, that the smallest and largest periods of this pattern are respectively 1 and 2 transactions, and that this pattern has an average periodicity of 1.4 transactions.

**Related problems**

Note that there also exists extensions of the problem of discovering periodic patterns. For example, another algorithm offered in the SPMF library is called PHM. It is designed to discover “**periodic high-utility itemsets**” in customer transaction databases. The goal is not only to find patterns that appear periodically, but also to discover the patterns that yields a high profit in terms of sales.

**Conclusion**

In this blog post, I have introduced the problem of discovering **periodic patterns** in databases. I have also explained how to use open-source software to discover periodic patterns. Mining periodic patterns is a general problem that may have many applications.

There are also many research opportunities related to periodic patterns, as this subject as not been extensively studied. For example, some possibility could be to transform the proposed algorithms into an incremental algorithm, a fuzzy algorithm, or to discover more complex types of periodic patterns such as periodic sequential rules.

Also, in this blog post, I have not discussed about time series (sequences of numbers). It is also another interesting problem to discover periodic patterns in time series, which requires different types of algorithms.

==

**Philippe Fournier-Viger** is a full professor and the founder of the open-source data mining software SPMF, offering more than 110 data mining algorithms. If you like this blog, you can tweet about it and/or subscribe to my twitter account **@philfv **to get notified about new posts.

Pingback: An introduction to frequent pattern mining - The Data Mining & Research Blog

Pingback: Brief report about the 16th Industrial Conference on Data mining 2016 (ICDM 2016) - The Data Mining & Research Blog