Mahout has a Top K Parallel FPGrowth Implementation. Its based on the paper http://infolab.stanford.edu/~echang/recsys08-69.pdf with some optimisations in mining the data.
Given a huge transaction list, the algorithm finds all unique features(sets of field values) and eliminates those features whose frequency in the whole dataset is less that minSupport. Using these remaining features N, we find the top K closed patterns for each of them, generating a total of NxK patterns. FPGrowth Algorithm is a generic implementation, we can use any Object type to denote a feature. Current implementation requires you to use a String as the object type. You may implement a version for any object by creating Iterators, Convertors and TopKPatternWritable for that particular object. For more information please refer the package org.apache.mahout.fpm.pfpgrowth.convertors.string
- The first argument is the iterator of transaction in this case its Iterator<List<String>>
- The second argument is the output of generateFList function, which returns the frequent items and their frequencies from the given database transaction iterator
- The third argument is the minimum Support of the pattern to be generated
- The fourth argument is the maximum number of patterns to be mined for each feature
- The fifth argument is the set of features for which the frequent patterns has to be mined
- The last argument is an output collector which takes [key, value] of Feature and TopK Patterns of the format [String, List<Pair<List<String>, Long>>] and writes them to the appropriate writer class which takes care of storing the object, in this case in a Sequence File Output format
Running Frequent Pattern Growth via command line
The command line launcher for string transaction data org.apache.mahout.fpm.pfpgrowth.FPGrowthDriver has other features including specifying the regex pattern for spitting a string line of a transaction into the constituent features.
Input files have to be in the following format.
<optional document id>TAB<TOKEN1>SPACE<TOKEN2>SPACE....
instead of tab you could use , or | as the default tokenization is done using a java Regex pattern
You can override this parameter to parse your log files or transaction files (each line is a transaction.) The FPGrowth algorithm mines the top K frequently occurring sets of items and their counts from the given input data
$MAHOUT_HOME/core/src/test/resources/retail.dat is a sample dataset in this format.
Other sample files are accident.dat.gz from http://fimi.cs.helsinki.fi/data/. As a quick test, try this:
The minimumSupport parameter -s is the minimum number of times a pattern or a feature needs to occur in the dataset so that it is included in the patterns generated. You can speed up the process by having a large value of s. There are cases where you will have less than k patterns for a particular feature as the rest don't for qualify the minimum support criteria
Note that the input to the algorithm, could be uncompressed or compressed gz file or even a directory containing any number of such files.
We modified the regex to use space to split the token. Note that input regex string is escaped.
Running Parallel FPGrowth
Running parallel FPGrowth is as easy as adding changing the flag -method mapreduce and adding the number of groups parameter e.g. -g 20 for 20 groups. First, let's run the above sample test in map-reduce mode:
The above test took 102 seconds on dual-core laptop, v.s. 609 seconds in the sequential mode, (with 5 gigs of ram allocated). In a separate test, the first 1000 lines of retail.dat took 20 seconds in map/reduce v.s. 30 seconds in sequential mode.
Here is another dataset which, while several times larger, requires much less time to find frequent patterns, as there are very few. Get accidents.dat.gz from http://fimi.cs.helsinki.fi/data/ and place it on your hdfs in a folder named accidents. Then, run the hadoop version of the FPGrowth job:
OR to run a dataset of this size in sequential mode on a single machine let's give Mahout a lot more memory and only keep features with more than 300 members:
The numGroups parameter -g in FPGrowthJob specifies the number of groups into which transactions have to be decomposed. The default of 1000 works very well on a single-machine cluster; this may be very different on large clusters.
Note that accidents.dat has 340 unique features. So we chose -g 10 to split the transactions across 10 shards where 34 patterns are mined from each shard. (Note: g doesnt need to be exactly divisible.) The Algorithm takes care of calculating the split. For better performance in large datasets and clusters, try not to mine for more than 20-25 features per shard. Stick to the defaults on a small machine.
The numTreeCacheEntries parameter -tc specifies the number of generated conditional FP-Trees to be kept in memory so that subsequent operations do not to regenerate them. Increasing this number increases the memory consumption but might improve speed until a certain point. This depends entirely on the dataset in question. A value of 5-10 is recommended for mining up to top 100 patterns for each feature.
Viewing the results
The output will be dumped to a SequenceFile in the frequentpatterns directory in Text=>TopKStringPatterns format. Run this command to see a few of the Frequent Patterns:
or replace -n 4 with -c for the count of patterns.
Open questions: how does one experiment and monitor with these various parameters?