private void ConvergeFeature(TSvdModel model, int f, List <IRating> ratings, TrainingParameters trainingParameters)
{
var count = 0;
var rmseImprovment = float.MaxValue;
while ((rmseImprovment > trainingParameters.RmseImprovementTreshold || count < trainingParameters.MinEpochTreshold) && count < trainingParameters.MaxEpochTreshold)
{
rmsePrev = rmse;
rmse = TrainFeature(model, f, ratings, trainingParameters);
rmseImprovment = Math.Abs(rmse - rmsePrev) / (rmse + rmsePrev);
count++;
Console.WriteLine("F{0}/P{1}:\trmse = {2}\trmseImpr = {3}", f + 1, count, rmse, rmseImprovment);
}
rmsePrev = rmse;
}
protected TSvdModel TrainModel(List <string> users, List <string> artists, List <IRating> trainRatings, TrainingParameters trainingParameters)
{
var model = GetNewModelInstance();
InitializeNewModel(model, users, artists, trainRatings);
CalculateFeatures(model, users, artists, trainRatings, trainingParameters);
return(model);
}
protected void CalculateFeatures(TSvdModel model, List <string> users, List <string> artists, List <IRating> ratings, TrainingParameters trainingParameters)
{
//init
model.UserFeatures = new float[trainingParameters.FeatureCount, users.Count];
model.ArtistFeatures = new float[trainingParameters.FeatureCount, artists.Count];
ResidualRatingValues = new float[ratings.Count];
model.UserFeatures.Populate(0.1f);
model.ArtistFeatures.Populate(0.1f);
rmsePrev = float.MaxValue;
rmse = float.MaxValue;
//MAIN LOOP - loops through features
for (var f = 0; f < trainingParameters.FeatureCount; f++)
{
Console.WriteLine("Training feature {0} ({1})", f + 1, DateTime.Now);
ConvergeFeature(model, f, ratings, trainingParameters);
CacheResidualRatings(model, f, ratings);
}
}
public TSvdModel TrainModel(List <IUser> trainUsers, List <IArtist> artists, List <IRating> trainRatings, TrainingParameters trainingParameters)
{
return(TrainModel(trainUsers.GetLookupTable(), artists.GetLookupTable(), trainRatings, trainingParameters));
}
protected float TrainSample(TSvdModel model, int r, int f, List <IRating> ratings, TrainingParameters trainingParameters)
{
var e = ratings[r].Value - PredictRatingUsingResiduals(model, r, f, ratings);
var uv = model.UserFeatures[f, ratings[r].UserIndex];
model.UserFeatures[f, ratings[r].UserIndex] += trainingParameters.LRate * (e * model.ArtistFeatures[f, ratings[r].ArtistIndex] - trainingParameters.K * model.UserFeatures[f, ratings[r].UserIndex]);
model.ArtistFeatures[f, ratings[r].ArtistIndex] += trainingParameters.LRate * (e * uv - trainingParameters.K * model.ArtistFeatures[f, ratings[r].ArtistIndex]);
return(e * e);
}
protected float TrainFeature(TSvdModel model, int f, List <IRating> ratings, TrainingParameters trainingParameters)
{
var e = ratings.Select((r, i) => TrainSample(model, i, f, ratings, trainingParameters)).Sum() / ratings.Count;
return((float)Math.Sqrt(e));
}
public static void Main(string[] args)
{
#if !DEBUG
try
{
#endif
if (args.Length == 0)
{
Console.WriteLine("Please enter testing arguments.");
return;
}
var argList = args.Select(arg => arg.ToLower()).ToList();
var mode = argList[argList.IndexOf("-mode") + 1].ToLower();
Console.WriteLine("Performing {0} tests...", mode.ToUpper());
List<IArtist> artists;
List<IUser> trainUsers;
List<IRating> trainRatings;
List<IUser> testUsers;
List<IRating> testRatings;
LoadData(out trainUsers, out trainRatings, out testUsers, out testRatings, out artists);
List<int> selectedModels;
List<string> models;
int numberOfTests;
List<int> ks;
bool performNoContentKnnTests;
bool performContentKnnTests;
switch (mode)
{
case "simple":
case "naive":
#region Simple
var simpleTester = new NaiveTester(trainUsers, artists, trainRatings, testUsers);
simpleTester.Test();
break;
#endregion
case "knn":
#region kNN
ks = argList[argList.IndexOf("-k") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(int.Parse).ToList();
numberOfTests = int.Parse(argList[argList.IndexOf("-t") + 1]);
performNoContentKnnTests = argList.IndexOf("-noco") >= 0;
performContentKnnTests = argList.IndexOf("-co") >= 0;
var sims = new List<ISimilarityEstimator<ISimpleKnnUser>>();
foreach (var argSim in argList[argList.IndexOf("-sim") + 1].ToLower().Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries))
{
switch (argSim)
{
case "pse":
sims.Add(new PearsonSimilarityEstimator());
break;
case "cse":
sims.Add(new CosineSimilarityEstimator());
break;
case "upse":
sims.Add(new UnionPearsonSimilarityEstimator());
break;
case "ucse":
sims.Add(new UnionCosineSimilarityEstimator());
break;
}
}
var rgs = new List<IRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>>();
foreach (var argRg in argList[argList.IndexOf("-rg") + 1].ToLower().Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries))
{
switch (argRg)
{
case "sara":
rgs.Add(new RatingAggregationRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>(new SimpleAverageRatingAggregator<ISimpleKnnUser>()));
break;
case "wsra":
rgs.Add(new RatingAggregationRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>(new WeightedSumRatingAggregator<ISimpleKnnUser>()));
break;
case "awsra":
rgs.Add(new RatingAggregationRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>(new AdjustedWeightedSumRatingAggregator<ISimpleKnnUser>()));
break;
case "frg":
rgs.Add(new FifthsSimpleRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>());
break;
case "ldrg":
rgs.Add(new LinearDescentSimpleRecommendationGenerator<ISimpleKnnModel, ISimpleKnnUser>());
break;
}
}
var knnTrainer = new SimpleKnnTrainer();
timer.Restart();
var knnModel = knnTrainer.TrainModel(trainUsers, artists, trainRatings);
timer.Stop();
Console.WriteLine("Model trained in {0}ms.", timer.ElapsedMilliseconds);
foreach (var k in ks)
{
foreach (var sim in sims)
{
foreach (var rg in rgs)
{
if (performNoContentKnnTests)
{
var knnTester = new KnnTester<SimpleKnnRecommender>
{
K = k,
Sim = sim,
Rg = rg,
TestUsers = testUsers,
SimpleKnnModel = knnModel,
Trainer = knnTrainer,
Artists = artists,
NumberOfTests = numberOfTests
};
knnTester.Test();
}
if (performContentKnnTests)
{
var contentKnnTester = new KnnTester<ContentSimpleKnnRecommender>
{
K = k,
Sim = sim,
Rg = rg,
TestUsers = testUsers,
SimpleKnnModel = knnModel,
Trainer = knnTrainer,
Artists = artists,
NumberOfTests = numberOfTests
};
contentKnnTester.Test();
}
}
}
}
break;
#endregion
case "svd-train":
case "mf-train":
#region SVD/MF Training
var svdFs = argList[argList.IndexOf("-f") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(int.Parse).ToList();
var svdLrs = argList[argList.IndexOf("-lr") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(s => float.Parse(s, CultureInfo.InvariantCulture)).ToList();
var svdKs = argList[argList.IndexOf("-k") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(s => float.Parse(s, CultureInfo.InvariantCulture)).ToList();
var svdRis = argList[argList.IndexOf("-ri") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(s => float.Parse(s, CultureInfo.InvariantCulture)).ToList();
var svdEs = argList[argList.IndexOf("-e") + 1].Split(new[] {'-'}, StringSplitOptions.RemoveEmptyEntries).Select(int.Parse).ToList();
var svdBbs = argList[argList.IndexOf("-bb") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(int.Parse).ToList();
var svdTypes = argList[argList.IndexOf("-type") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).ToList();
var minEpoch = svdEs[0];
var maxEpoch = svdEs[1];
var svdBasic = svdTypes.Contains("basic");
var svdBias = svdTypes.Contains("bias");
foreach (var svdF in svdFs)
{
foreach (var svdLr in svdLrs)
{
foreach (var svdK in svdKs)
{
foreach (var svdRi in svdRis)
{
foreach (var svdBb in svdBbs)
{
var trainingParameters = new TrainingParameters(svdF, svdLr, svdK, svdRi, minEpoch, maxEpoch, svdBb);
var filename = string.Format(CultureInfo.InvariantCulture, "F{0}-LR{1}-K{2}-RI{3}-E{4}-{5}", svdF, svdLr, svdK, svdRi, minEpoch, maxEpoch);
if (svdBasic)
{
var basicSimpleSvdBiasBinsTrainer = new SimpleSvdBiasBinsTrainer();
var model = basicSimpleSvdBiasBinsTrainer.TrainModel(trainUsers, artists, trainRatings, trainingParameters);
basicSimpleSvdBiasBinsTrainer.SaveModel(string.Format(CultureInfo.InvariantCulture, @"D:\Dataset\models\basicSvd-{0}-BB{1}.rs", filename, svdBb), model);
new SimpleSvdTrainer().SaveModel(string.Format(CultureInfo.InvariantCulture, @"D:\Dataset\models\basicSvd-{0}.rs", filename), model);
}
if (svdBias)
{
var biasSimpleSvdBiasBinsTrainer = new BiasSimpleSvdBiasBinsTrainer();
var model = biasSimpleSvdBiasBinsTrainer.TrainModel(trainUsers, artists, trainRatings, trainingParameters);
biasSimpleSvdBiasBinsTrainer.SaveModel(string.Format(CultureInfo.InvariantCulture, @"D:\Dataset\models\biasSvd-{0}-BB{1}.rs", filename, svdBb), model);
new BiasSimpleSvdTrainer().SaveModel(string.Format(CultureInfo.InvariantCulture, @"D:\Dataset\models\biasSvd-{0}.rs", filename), model);
}
}
}
}
}
}
break;
#endregion
case "svd":
case "mf":
#region SVD/MF Prediction
selectedModels = new List<int>();
models = Directory.GetFiles(@"D:\Dataset\models\", "*svd*.rs", SearchOption.TopDirectoryOnly).ToList();
if (argList.Contains("-all"))
selectedModels.AddRange(models.Select((t, i) => i));
else
{
for (var i = 0; i < models.Count; i++)
{
var modelName = Path.GetFileName(models[i]);
modelName = modelName != null ? modelName.Remove(modelName.Length - 3) : "UnnamedModel";
Console.WriteLine("{0}) {1}", i + 1, modelName);
}
Console.WriteLine("Enter numbers of models for testing (separate with comma):");
string line;
while ((line = Console.ReadLine()) == null)
{}
selectedModels.AddRange(line == "all" ? models.Select((t, i) => i) : line.Split(new[] {' ', ','}).Select(part => int.Parse(part) - 1));
}
for (var i = 0; i < selectedModels.Count; i++)
{
var selectedModel = selectedModels[i];
if (selectedModel >= models.Count)
continue;
var modelFile = models[selectedModel];
var testName = Path.GetFileName(modelFile);
testName = testName != null ? testName.Remove(testName.Length - 3) : "SvdTest";
if (modelFile.Contains("basic"))
{
if (modelFile.Contains("BB"))
{
var rs = new SimpleSvdRecommendationSystem<ISvdBiasBinsModel>(new SimpleSvdBiasBinsTrainer(), new SimpleSvdBiasBinsRecommender());
var model = new SvdBiasBinsModel();
rs.Recommender.LoadModel(model, modelFile);
new SvdTester<ISvdBiasBinsModel>(testName, rs, model, testUsers, testRatings, artists).Test();
}
else
{
var rs = new SimpleSvdRecommendationSystem<ISvdModel>(new SimpleSvdTrainer(), new SimpleSvdRecommender());
var model = new SvdModel();
rs.Recommender.LoadModel(model, modelFile);
new SvdTester<ISvdModel>(testName, rs, model, testUsers, testRatings, artists).Test();
}
}
else if (modelFile.Contains("bias"))
{
if (modelFile.Contains("BB"))
{
var rs = new SimpleSvdRecommendationSystem<IBiasSvdBiasBinsModel>(new BiasSimpleSvdBiasBinsTrainer(), new BiasSimpleSvdBiasBinsRecommender());
var model = new BiasSvdBiasBinsModel();
rs.Recommender.LoadModel(model, modelFile);
new SvdTester<IBiasSvdBiasBinsModel>(testName, rs, model, testUsers, testRatings, artists).Test();
}
else
{
var rs = new SimpleSvdRecommendationSystem<IBiasSvdModel>(new BiasSimpleSvdTrainer(), new BiasSimpleSvdRecommender());
var model = new BiasSvdModel();
rs.Recommender.LoadModel(model, modelFile);
new SvdTester<IBiasSvdModel>(testName, rs, model, testUsers, testRatings, artists).Test();
}
}
}
break;
#endregion
case "sbk":
#region SVD boosted kNN
#region CLI argument parsing
numberOfTests = int.Parse(argList[argList.IndexOf("-t") + 1]);
performNoContentKnnTests = argList.IndexOf("-noco") >= 0;
performContentKnnTests = argList.IndexOf("-co") >= 0;
ks = argList[argList.IndexOf("-k") + 1].Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries).Select(int.Parse).ToList();
var sbkSims = new List<ISimilarityEstimator<ISvdBoostedKnnUser>>();
foreach (var argSim in argList[argList.IndexOf("-sim") + 1].ToLower().Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries))
{
switch (argSim)
{
case "pse":
sbkSims.Add(new PearsonSvdBoostedKnnSimilarityEstimator());
break;
case "cse":
sbkSims.Add(new CosineSvdBoostedKnnSimilarityEstimator());
break;
}
}
var sbkRgs = new List<IRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>>();
foreach (var argRg in argList[argList.IndexOf("-rg") + 1].ToLower().Split(new[] {','}, StringSplitOptions.RemoveEmptyEntries))
{
switch (argRg)
{
case "sara":
sbkRgs.Add(new RatingAggregationRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>(new SimpleAverageRatingAggregator<ISvdBoostedKnnUser>()));
break;
case "wsra":
sbkRgs.Add(new RatingAggregationRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>(new WeightedSumRatingAggregator<ISvdBoostedKnnUser>()));
break;
case "awsra":
sbkRgs.Add(new RatingAggregationRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>(new AdjustedWeightedSumRatingAggregator<ISvdBoostedKnnUser>()));
break;
case "frg":
sbkRgs.Add(new FifthsSimpleRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>());
break;
case "ldrg":
sbkRgs.Add(new LinearDescentSimpleRecommendationGenerator<ISvdBoostedKnnModel, ISvdBoostedKnnUser>());
break;
}
}
#endregion
#region Model selection
selectedModels = new List<int>();
models = Directory.GetFiles(@"D:\Dataset\models\", "*svd*.rs", SearchOption.TopDirectoryOnly).Where(m => !m.Contains("BB")).ToList();
if (argList.Contains("-all") || models.Count == 1)
{
selectedModels.AddRange(models.Select((t, i) => i));
}
else
{
for (var i = 0; i < models.Count; i++)
{
var modelName = Path.GetFileName(models[i]);
modelName = modelName != null ? modelName.Remove(modelName.Length - 3) : "UnnamedModel";
Console.WriteLine("{0}) {1}", i + 1, modelName);
}
Console.WriteLine("Enter numbers of models for testing (separate with comma):");
string line;
while ((line = Console.ReadLine()) == null)
{ }
selectedModels.AddRange(line == "all" ? models.Select((t, i) => i) : line.Split(new[] { ' ', ',' }).Select(part => int.Parse(part) - 1));
}
#endregion
#region Testing
foreach (var k in ks)
{
foreach (var sbkSim in sbkSims)
{
foreach (var sbkRg in sbkRgs)
{
for (var i = 0; i < selectedModels.Count; i++)
{
var selectedModel = selectedModels[i];
if (selectedModel >= models.Count)
continue;
var modelFile = models[selectedModel];
var testName = Path.GetFileName(modelFile);
testName = testName != null ? testName.Remove(testName.Length - 3) : "SvdBoostedKnnTest";
if (modelFile.Contains("basic"))
{
if (performNoContentKnnTests)
{
var sbkRs = new SvdBoostedKnnRecommendationSystem<ISvdBoostedKnnModel>(new SvdBoostedKnnSvdTrainer(),
new KnnTrainerForSvdModels(),
new SvdBoostedKnnRecommender<ISvdBoostedKnnModel>(sbkSim,
sbkRg,
new NewUserFeatureGenerator(),
k));
var sbkModel = sbkRs.KnnTrainerForSvdModels.TrainKnnModel(modelFile, trainUsers);
testName = string.Format("SBK-(k{0}-{1}-{2}-{3}-T{4})-{5}", k, sbkSim, sbkRg, sbkRs.Recommender, numberOfTests, testName);
var sbkTester = new SvdBoostedKnnTester<ISvdBoostedKnnModel>(testName, sbkRs, sbkModel, testUsers, testRatings, artists, numberOfTests);
sbkTester.Test();
}
if (performContentKnnTests)
{
var sbkRs = new SvdBoostedKnnRecommendationSystem<ISvdBoostedKnnModel>(new SvdBoostedKnnSvdTrainer(),
new KnnTrainerForSvdModels(),
new ContentSvdBoostedKnnRecommender<ISvdBoostedKnnModel>(sbkSim,
sbkRg,
new NewUserFeatureGenerator(),
new ContentSimilarityEstimator(),
k));
var sbkModel = sbkRs.KnnTrainerForSvdModels.TrainKnnModel(modelFile, trainUsers);
testName = string.Format("SBK-(k{0}-{1}-{2}-{3}-T{4})-{5}", k, sbkSim, sbkRg, sbkRs.Recommender, numberOfTests, testName);
var sbkTester = new SvdBoostedKnnTester<ISvdBoostedKnnModel>(testName, sbkRs, sbkModel, testUsers, testRatings, artists, numberOfTests);
sbkTester.Test();
}
}
else if (modelFile.Contains("bias"))
{
if (performNoContentKnnTests)
{
var sbkRs = new SvdBoostedKnnRecommendationSystem<IBiasSvdBoostedKnnModel>(new BiasSvdBoostedKnnSvdTrainer(),
new BiasKnnTrainerForSvdModels(),
new SvdBoostedKnnRecommender<IBiasSvdBoostedKnnModel>(sbkSim,
sbkRg,
new BiasNewUserFeatureGenerator(),
k));
var sbkModel = sbkRs.KnnTrainerForSvdModels.TrainKnnModel(modelFile, trainUsers);
testName = string.Format("SBK-(k{0}-{1}-{2}-{3}-T{4})-{5}", k, sbkSim, sbkRg, sbkRs.Recommender, numberOfTests, testName);
var sbkTester = new SvdBoostedKnnTester<IBiasSvdBoostedKnnModel>(testName, sbkRs, sbkModel, testUsers, testRatings, artists, numberOfTests);
sbkTester.Test();
}
if (performContentKnnTests)
{
var sbkRs = new SvdBoostedKnnRecommendationSystem<IBiasSvdBoostedKnnModel>(new BiasSvdBoostedKnnSvdTrainer(),
new BiasKnnTrainerForSvdModels(),
new ContentSvdBoostedKnnRecommender<IBiasSvdBoostedKnnModel>(sbkSim,
sbkRg,
new BiasNewUserFeatureGenerator(),
new ContentSimilarityEstimator(),
k));
var sbkModel = sbkRs.KnnTrainerForSvdModels.TrainKnnModel(modelFile, trainUsers);
testName = string.Format("SBK-(k{0}-{1}-{2}-{3}-T{4})-{5}", k, sbkSim, sbkRg, sbkRs.Recommender, numberOfTests, testName);
var sbkTester = new SvdBoostedKnnTester<IBiasSvdBoostedKnnModel>(testName, sbkRs, sbkModel, testUsers, testRatings, artists, numberOfTests);
sbkTester.Test();
}
}
}
}
}
}
#endregion
break;
#endregion
}
#if !DEBUG
}
catch (Exception e)
{
Console.WriteLine("{0}{1}{1}{2}", e, Environment.NewLine, e.Message);
}
#endif
Console.WriteLine("DONE!");
if (args.Where(arg => arg.ToLower() == "-wait").Count() != 0)
Console.ReadLine();
}
public new TBiasBinsSvdModel TrainModel(List <IUser> trainUsers, List <IArtist> artists, List <IRating> trainRatings, TrainingParameters trainingParameters)
{
var model = TrainModel(trainUsers.GetLookupTable(), artists.GetLookupTable(), trainRatings, trainingParameters);
BiasBinsCalculator.CalculateBiasBins(model, trainRatings, trainUsers, artists, trainingParameters.BiasBinCount);
return(model);
}