Spark-FM

Factorization Machines is a general predictor like SVMs but is also able to estimate reliable parameters under very high sparsity. However, they are costly to scale to large amounts of data and large numbers of features. Spark-FM is a parallel implementation of factorization machines based on Spark. It aims to utilize Spark's in-memory computing to address above problems.

Highlight

In order to meet users' demands, Spark-FM supports various of optimization methods to train the model as follows.

Mini-batch Stochastic Gradient Descent (MLlib)
L-BFGS (MLlib)
Parallel Stochastic Gradient Descent (spark-optim)
Parallel Ftrl (spark-optim)

Examples

Scala API

val spark = SparkSession
  .builder()
  .appName("FactorizationMachinesExample")
  .master("local[*]")
  .getOrCreate()

val train = spark.read.format("libsvm").load("data/a9a.tr")
val test = spark.read.format("libsvm").load("data/a9a.te")

val trainer = new FactorizationMachines()
  .setAlgo(Algo.fromString("binary classification"))
  .setSolver(Solver.fromString("pftrl"))
  .setDim((1, 1, 8))
  .setReParamsL1((0.1, 0.1, 0.1))
  .setRegParamsL2((0.01, 0.01, 0.01))
  .setAlpha((0.1, 0.1, 0.1))
  .setBeta((1.0, 1.0, 1.0))
  .setInitStdev(0.01)
  // .setStepSize(0.1)
  .setTol(0.001)
  .setMaxIter(1)
  .setThreshold(0.5)
  // .setMiniBatchFraction(0.5)
  .setNumPartitions(4)

val model = trainer.fit(train)
val result = model.transform(test)
val predictionAndLabel = result.select("prediction", "label")
val evaluator = new MulticlassClassificationEvaluator().setMetricName("accuracy")
println("Accuracy: " + evaluator.evaluate(predictionAndLabel))
spark.stop()

Requirements

Spark-FM is built against Spark 2.1.1.

Build From Source

sbt package

Licenses

Spark-FM is available under Apache Licenses 2.0.

Contact & Feedback

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

hibayesian ([email protected]).