Sequential pattern mining aims to discover hidden patterns and relationships in collections of sequential data; it has been successfully applied in a number of data mining tasks including text mining (e.g., finding frequent phrases), market basket analysis (e.g., finding frequent sequences of product sales), and web usage mining (e.g., finding frequent sequences of page visits). In particular, in the context of text mining, frequent phrases (commonly referred to as n-grams) are widely used in applications such as machine translation, speech recognition and information extraction.

MG-FSM is a scalable, general-purpose frequent sequence mining algorithm built for MapReduce. It takes as input a collection of sequences (e.g., a text collection or Web usage logs) and mines frequent sequences subject to a number of constraints such as minimum frequency, maximum length, or proximity constraints (position-based or temporal). A detailed description of MG-FSM can be found here.

###Contributors [Iris Miliaraki] 4, [Klaus Berberich] 5, [Rainer Gemulla] 6, [Kaustubh Beedkar] 7 and [Dhruv Gupta] 8.

Given a collection of input sequences (a sequence database), MG-FSM finds frequent subsequences that

• Occur in at least σ ≥ 1 sequences (support threshold).
• Have length at most λ ≥ 2 (length threshold).
• Have gap at most γ ≥ 0 between consecutive items (gap threshold).

In additition to these constraints, MG-FSM also supports other types of constraints. Please refer to the available command line options for details.

1. Prerequisites for building MG-FSM - Java JDK 1.6 or 1.7 - Maven 3.2 or higher

2. Check out the MG-FSM source code to some directory, we call it here as MGFSM_HOME

3. Compiling:

      $ cd MGFSM_HOME
      $ mvn clean install
   

4. Set the environment variables:

      $ export JAVA_HOME=/location of Java
      $ export HADOOP_HOME=/location of Hadoop
   

After following the above methods, you can run the MG-FSM algorithm with appropriate arguments. Also, note that you can run the algorithm over HDFS data (in distributed mode) or local file system data (in sequential mode).

The input sequence file(s) should be of the following format:

  s1  Central Park is the best place to be on a sunny day in New York
  s2  Monday was a sunny day in New York
  s3  It was a sunny and beautiful New York City afternoon

That is, each line represents a sequence, and has the format:

  SequenceId item1 item2 item3 ... itemN

In this case, the first token of each line (whitespace delimited) becomes the sequence id, and all additional text of the line is interpreted as a sequence of items. For example, the above sequence database contains 3 sequences (s1, s2 and s3).

The sequential version of the algorithm runs locally on a single machine. Following example illustrates how to execute in such a mode:

  $ MGFSM_HOME/bin/mgfsm -i /path/to/input/dir/ -o /path/to/output/dir/ -s σ -g γ -l λ -m s

When executing the “-(s)equential” mode the input data file(s) should be in “.txt” format (i.e. “.txt” extension should be present in the name) for proper execution.

To execute the algorithm on a Hadoop cluster, issue the “distributed” mode to the algorithm. Following example illustrates:

  $ MGFSM_HOME/bin/mgfsm -i /path/to/input/dir/ -o /path/to/output/dir/ -s σ -g γ -l λ -m d

Note: The current version of MG-FSM is tested with Hadoop 2.5.0 and Hadoop 2.6.0.

Frequent sequences are written to a file in the output directory, where each line in the file has the following format:

  sequence <tab> #frequency

Assume the following input sequences are in a file(s) (.txt) in a directory called SAMPLE_INPUT/

  s1 Central Park is the best place to be on a sunny day in New York
  s2 Monday was a sunny day in New York
  s3 It was a sunny and beautiful New York City afternoon

Run the following command to find frequent all subsequences with support (σ = 3), maximum length (λ = 3) and maximum gap (γ = 2).

  $ MGFSM_HOME/bin/mgfsm -i SAMPLE_INPUT/ -o SAMPLE_OUTPUT/ -s 3 -g 2 -l 3 -m s

Sample output:

  sunny New York      3
  a sunny New         3
  New York            3
  sunny New           3
  a sunny             3

Additional Notes

• -(k)eepFiles and -(r)esume options are mutually exclusive.
• -(i)nput is optional as -(r)esume can also be used for pointing to the input source.
• -(i)nput and -(r)esume are mutually exclusive.
• -(o)utput is the only mandatory option. It points to the location where the translated frequent sequences will be written.