No description, website, or topics provided.
Switch branches/tags
Clone or download
Vincent Xie
Vincent Xie Merge pull request #17 from Fangrou/master
Upgrade scalastyle to 1.0.0
Latest commit 6610127 Oct 31, 2017


Build Status

A package contains three Spark-based implementations. It includes

  • Factorization Machines (LibFM)
  • Field-Awared Factorization Machine (FFM)
  • Conditional Random Fields (CRF)
  • Adaptive learning rate optimizer (AdaGrad, Adam)

This package can be imported as a dependency in other codes. Then, all functions of LibFM, FFM, CRF and AdaOptimizer in this package can be used.

Build from Source

sbt package

How to Use

IMLLIB package can be either imported directly in Spark-Shell, or be imported as a dependency in other codes.

Use in Spark-Shell

Temporary Use

(1) Run Spark-Shell with IMLLIB package

spark-shell --jars 'Path/imllib_2.11-0.0.1.jar'

Path is the path of imllib_2.11-0.0.1.jar

Permanent Use

(1) On driver node, add following codes into conf/spark-default.conf

spark.executor.extraClassPath    /usr/local/spark/lib/*
spark.driver.extraClassPath      /usr/local/spark/lib/*

(2) Create /usr/local/spark/lib
(3) Copy imllib_2.11-0.0.1.jar to /usr/local/spark/lib
(4) Copy conf/spark-default.conf and /usr/local/spark/lib to all worker nodes
(5) Run Spark-Shell


Use as a denpendency

(1) build from source and publish locally

sbt compile publish-local

(2) Move the whole directory from .ivy2/local to .ivy2/cache
(3) Add following codes into build.sbt when you want to import IMLLIB package as a denpendency

libraryDependencies += "" % "imllib_2.11" % "0.0.1"

How to Import


Test Examples

There are three shell scripts in bin/template for testing LibFM, FFM, CRF and LR with AdaOptimizer respectively. The script runs in a local mode Spark with the data on hadoop. You can first modify the script with necessary changes, such as hostname, port for hadoop, etc. Then run the script to test if the algorithm works.


A Spark-based implementation of Factorization Machines (LibFM)

The code base structure of this project is from spark-libFM, but the optimization method is based on parallel-sgd which has stronger convergence than miniBatch-sgd.


A Spark-based implementation of Field-Awared Factorization Machine with parallelled AdaGrad solver. See

Need to rework the data format the fit FFM, that is to extends LIBSVM data format by adding field information to each feature to have formation like: label field1:feat1:val1 field2:feat2:val2


A Spark-based implementation of Conditional Random Fields (CRF) for segmenting/labeling sequential data.

CRF-Spark provides following features:

  • Training in parallel based on Spark RDD
  • Support a simple format of training and test file. Any other format also can be read by a simple tokenizer.
  • A common feature templates design, which is also used in other machine learning tools, such as CRF++ and miralium
  • Fast training based on Limited-memory BFGS optimizaton algorithm (fitting L2-regularized models) or Orthant-wise Limited Memory Quasi-Newton optimizaton algorithm (fitting L1-regularized models)
  • Support two verbose levels to provide extra information. VerboseLevel1 shows marginal probabilities for each tag and a conditional probability for the entire output; VerboseLevel2 shows marginal probabilities for all other candidates additionally.
  • Support n-best outputs
  • Linear-chain (first-order Markov) CRF
  • Test can run both in parallel and in serial


A Spark-based implementation of Adam and AdaGrad optimizer, methods for Stochastic Optimization. See and Comparing with SGD, Adam and AdaGrad have better performance. Especially in case of sparse features, Adam can converge faster than normal SGD.

Contact & Feedback

If you encounter bugs, feel free to submit an issue or pull request. Also you can mail to: