Modelos mixtos en spark. Intento 1
A los que nos dedicamos a esto siempre echamos de menos un lme4
en python
o en Spark
. En Julia
afortunadamente tenemos MixedModels.jl
.
Total que buscando alguna posible solución para poder usar esto en spark me encuentro con dos posibles soluciones.
Ambos repos llevan un tiempo sin actualizarse así que no sé yo.
photon-ml
es de linkedin y tiene buena pinta, al menos el tutorial, que tienes que bajarte un docker y tal, funciona. Aunque la sintaxis es rara. Aún tengo que probarlo más y probar a crear el jar del proyecto ya que no está en maven central y tal (y no me funcionó)
Ejemplo de sintaxis de photon-ml
// Define another feature shard for our random effect coordinate, and create a new mapping
// with both our 'global' and 'perUser' shards.
val perUserFeatureShardId = "perUser"
val perUserFeatureShard = Set("genreFeatures", "movieLatentFactorFeatures")
val mixedFeatureShardBags = Map(
-> globalFeatureShard,
globalFeatureShardId -> perUserFeatureShard)
perUserFeatureShardId
// Since we have a new shard, re-read the training and validation data into a new DataFrame
// (and a new index map for the new feature shard).
val (mixedInputData, mixedFeatureShardInputIndexMaps) = dataReader.readMerged(
Seq("/data/movielens/trainData.avro"),
,
mixedFeatureShardBags)
numPartitionsval mixedValidateData = dataReader.readMerged(
Seq("/data/movielens/validateData.avro"),
,
mixedFeatureShardInputIndexMaps,
mixedFeatureShardBags) numPartitions
Donde mixedInputData es un dataframe de spark con esta pinta.
.show()
mixedInputData
+----+--------+------+-------+------+------+--------------------+--------------------+
| uid|response|userId|movieId|weight|offset| global| perUser|
+----+--------+------+-------+------+------+--------------------+--------------------+
|null| 4.0| 1| 1215| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.5| 1| 1350| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.0| 1| 2193| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.5| 1| 3476| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 5.0| 1| 4993| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.0| 3| 1544| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.0| 7| 440| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.0| 7| 914| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.0| 7| 1894| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.0| 7| 2112| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.0| 7| 3524| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.0| 7| 3911| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 5.0| 11| 256| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 5.0| 11| 1200| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 4.5| 11| 48394| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 1.0| 11| 56003| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 0.5| 11| 64508| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 5.0| 14| 471| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.0| 14| 2018| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
|null| 3.5| 14| 6936| null| null|(51,[0,1,2,3,4,5,...|(51,[0,1,2,3,4,5,...|
Donde las columna global y perUser son iguales, pero una se usa para estimar la parte de los efectos fijos y la otra para los aleatorios.
Y luego sigue con
// A 'RandomEffectDataConfiguration' requires an identifier field to use for grouping data from the
// same entity, in addition to the fields that a 'FixedEffectDataConfiguration' requires. It also has
// some additional optional parameters not covered in this tutorial.
val perUserRandomEffectId = "userId"
val perUserDataConfig = RandomEffectDataConfiguration(
,
perUserRandomEffectId,
perUserFeatureShardId,
numPartitions= IndexMapProjection)
projectorType
// A 'RandomEffectOptimizationConfiguration' is defined much like a
// 'FixedEffectOptimizationConfiguration'. The options below are varied from those above primarily
// for variety and demonstration.
val perUserOptimizerConfig = OptimizerConfig(
= TRON,
optimizerType = 1e-3,
tolerance = 4)
maximumIterations val perUserRegularizationContext = L2RegularizationContext
val perUserRegularizationWeight = 1
val perUserOptimizationConfig = RandomEffectOptimizationConfiguration(
,
perUserOptimizerConfig,
perUserRegularizationContext)
perUserRegularizationWeight
// Assign a coordinate ID to the random effect configurations we defined above. This time, we have
// multiple coordinates and need to determine the update sequence. In general, it's recommended to
// order coordinates from least to most granular, i.e. those that correlate most with the response to
// those that correlate least.
val perUserCoordinateId = "perUser"
val mixedCoordinateDataConfigs = Map(
-> globalDataConfig,
globalCoordinateId -> perUserDataConfig)
perUserCoordinateId val mixedCoordinateOptConfigs = Map(
-> globalOptimizationConfig,
globalCoordinateId -> perUserOptimizationConfig)
perUserCoordinateId val mixedUpdateSequence = Seq(globalCoordinateId, perUserCoordinateId)
// Reset our estimator. The training task hasn't changed, but the data configurations and update
// sequence have. Furthermore, since there are now multiple coordinates, we should try multiple
// passes of coordinate descent.
.setCoordinateDataConfigurations(mixedCoordinateDataConfigs)
estimator.setCoordinateUpdateSequence(mixedUpdateSequence)
estimator.setCoordinateDescentIterations(2)
estimator
// Train a new model.
val (mixedModel, _, mixedModelConfig) = estimator.fit(
,
mixedInputDataSome(mixedValidateData),
Seq(mixedCoordinateOptConfigs)).head
// Save the trained model.
.saveGameModelToHDFS(
ModelProcessingUtils,
scnew Path("output/mixed"),
,
mixedModel,
trainingTask,
mixedModelConfigNone,
)
mixedFeatureShardInputIndexMaps
Y guarda los coeficientes en avro
"avro cat -n 1 ./output/mixed/random-effect/perUser/coefficients/part-00000.avro" #| "jq ." !
{
"variances": null,
"means": [
{
"term": "Drama",
"name": "Genre",
"value": -0.35129547272878503
},
{
"term": "Musical",
"name": "Genre",
"value": -0.2967514108349342
},
{
"term": "",
"name": "7",
"value": -0.13789947075029355
},
{
"term": "",
"name": "14",
"value": -0.13577029316450503
},
{
"term": "",
"name": "8",
"value": -0.12850130065314527
},
{
"term": "",
"name": "26",
"value": -0.11646520581859549
},
{
"term": "",
"name": "15",
"value": -0.09620039918539182
},
{
"term": "",
"name": "6",
"value": 0.08934738779979344
},
{
"term": "Comedy",
"name": "Genre",
"value": 0.08833383209245319
},
{
"term": "",
"name": "2",
"value": -0.08756438537931642
},
more coefficients
"modelClass": "com.linkedin.photon.ml.supervised.regression.LinearRegressionModel",
"lossFunction": "",
"modelId": "7"
}
Lo dicho, no tiene mala pinta y ajusta rápido, me falta probar a crear el jar del proyecto
Por otro lado MomentMixedModels
también parecía prometedora pero al intentar crear el jar con sbt (tampoco está en maven central) peta con (*:update) sbt.ResolveException: unresolved dependency: com.stitchfix.algorithms.spark#sfs3_2.11;0.7.0-spark2.2.0: not found
y viendo el build.sbt
hace referencia a http://artifactory.vertigo.stitchfix.com/artifactory/releases
que parece que ya no existe, así que mi gozo en un pozo. La sintaxis parecía sencilla.
val linearModelFitter = {
new MixedEffectsRegression()
.setResponseCol("Reaction")
.setFixedEffectCols(Seq("Days"))
.setRandomEffectCols(Seq("Days"))
.setFamilyParam("gaussian")
.setGroupCol("Subject")
}
val linearModel = linearModelFitter.fit(sleepstudyData)
println(linearModel.β)
Pues nada, a ver si algún ingenazi con alma de analista se digna a hacer una implementación de lme4
en Spark
, porque, reconozcámoslo Spark-ml
es una ñapa. Lo único que medio funciona bien es usar los algoritmos de h2o sobre spark con sparkling-water
y me falta probar un poco más su implementación de modelos jerárquicos
Hasta otra.