AutoML Add Recommendation Task #4246

LittleLittleCloud · 2019-09-25T20:57:50Z

What's already be done in this PR

added Recommendation task and experiment in AutoML
added MatrixFactorization as MatrixFactorizationExtension
added a new Column Purpose (LabelFeature) and it's corresponding TransformerExtension (LabelCategorical) so that AutoML can construct the pre-process pipeline for MatrixFactorizationExtension correctly
added a new recommendation example (with rating only) in AutoML.Example, and you can play with that!

What's need to be done (Feel Free to CRUD)

figuring out how to accelerate and properly presenting the training process. Seems that MatrixFactorization requires more time to train a round, and the algorithm for sweeping params requires to train many rounds to find out the best parameter. It's time costy and customers might not like that.
Corresponding CodeGen part
Test case!
Better Naming and code style
Enable support for multiple feature trainers in AutoML (it requires some refactor works and shouldn't be done in this PR. But it's important)

…odeGen (dotnet#4043) * add CodeGen Library * rename namespace to ML.CodeGen.* * cancel delay sign in CodeGen * update based on comment * remove useless nuget package * add ComsumeModel class * use consumeModel in CodeGen * use different annotation for different target * target to netstandard2.0 * remove console output * adjust output result * remove useless variable * update CodeGen name to CodeGenerator * rebase to latest branch * fix bug in Normalize function * fix test * rename features to featuresList * move enum out of class * remove useless items * remove NLog.config * update generated CSharp file * change wording and delete useless file * using Uppercase in comment

…arning into dotnet-features/automl

src/Microsoft.ML.Recommender/MatrixFactorizationPredictor.cs

src/Microsoft.ML.Recommender/MatrixFactorizationTrainer.cs

src/Microsoft.ML.AutoML/API/AutoCatalog.cs

src/Microsoft.ML.AutoML/ColumnInference/ColumnPurpose.cs

eerhardt · 2019-09-25T21:07:18Z

using System;

All product code needs a copyright.

Refers to: src/Microsoft.ML.AutoML/API/RecommendationExperiment.cs:1 in 9695ffe. [](commit_id = 9695ffe, deletion_comment = False)

eerhardt · 2019-09-25T21:08:21Z

mlnet.csproj cannot go into master yet.

Refers to: src/mlnet/mlnet.csproj:1 in 9695ffe. [](commit_id = 9695ffe, deletion_comment = False)

src/Microsoft.ML.AutoML/ColumnInference/ColumnPurpose.cs

justinormont · 2019-09-26T01:03:12Z

src/Microsoft.ML.AutoML/ColumnInference/ColumnPurpose.cs

@@ -13,6 +13,7 @@ internal enum ColumnPurpose
        TextFeature = 4,
        Weight = 5,
        ImagePath = 6,
-        SamplingKey = 7
+        SamplingKey = 7,
+        LabelFeature,   // CategoricalFeature that requires ValueToKey converter, better naming?


I think we'd want to use HashToKey (name be off) instead of the mentioned ValueToKey as the ValueToKey will map future unseen values to NA in your test dataset; and as a lesser issue is slow by taking a full pass of the dataset.

src/Microsoft.ML.AutoML/TransformInference/TransformInference.cs

test/Microsoft.ML.AutoML.Tests/TrainerExtensionsTests.cs

docs/samples/Microsoft.ML.AutoML.Samples/RecommendationExperiment.cs

src/Microsoft.ML.AutoML/API/RecommendationExperiment.cs

src/Microsoft.ML.AutoML/Utils/UserInputValidationUtil.cs

src/Microsoft.ML.AutoML/TrainerExtensions/SweepableParams.cs

src/Microsoft.ML.AutoML/API/RecommendationExperiment.cs

src/Microsoft.ML.AutoML/API/ColumnInference.cs

docs/samples/Microsoft.ML.AutoML.Samples/RecommendationExperiment.cs

- remove Label column for recommendation: not required to set up

docs/samples/Microsoft.ML.AutoML.Samples/RecommendationExperiment.cs

src/Microsoft.ML.AutoML/Experiment/SuggestedTrainer.cs

src/Microsoft.ML.AutoML/TransformInference/TransformInference.cs

docs/samples/Microsoft.ML.AutoML.Samples/RecommendationExperiment.cs

test/Microsoft.ML.AutoML.Tests/AutoFitTests.cs

src/Microsoft.ML.AutoML/TransformInference/TransformInference.cs

eerhardt

This is looking really good, @maryamariyan. Thanks for the great work!

LGTM

Trains Recommendation models able to predict rating for existing users Conflicts: pkg/Microsoft.ML.AutoML/Microsoft.ML.AutoML.nupkgproj src/Microsoft.ML.AutoML/Microsoft.ML.AutoML.csproj test/Microsoft.ML.AutoML.Tests/AutoFitTests.cs test/Microsoft.ML.AutoML.Tests/ColumnInferenceTests.cs test/Microsoft.ML.AutoML.Tests/ColumnInformationUtilTests.cs test/Microsoft.ML.AutoML.Tests/Microsoft.ML.AutoML.Tests.csproj test/Microsoft.ML.AutoML.Tests/TrainerExtensionsTests.cs test/Microsoft.ML.AutoML.Tests/TransformInferenceTests.cs test/Microsoft.ML.AutoML.Tests/UserInputValidationTests.cs

Trains Recommendation models able to predict rating for existing users

LittleLittleCloud and others added 10 commits August 25, 2019 13:07

pack codegen into mlnet

7e0f6d0

pack codegen into mlnet (dotnet#4179)

22edabb

Merge branch 'features/automl' of https://github.com/dotnet/machinele…

50e0dcd

…arning into dotnet-features/automl

add MatrixFactorization Trainer

09c56f7

add RecommendationExperiment and other functions

15c58f1

some refactor in MatrixFactorization, plus fix small bugs

ac57d9a

add LabelFeautre ColumnPurpose and some update

c07948f

Merge branch 'u/xiaoyun/recommendation'

f182a20

add missing Native dll

9695ffe

LittleLittleCloud requested review from a team as code owners September 25, 2019 20:57