/// <summary>
/// Lifts a FiniteDist<A> into a SampleableDist<A>
/// </summary>
public static PrimitiveDist <A> ToSampleDist <A>(this FiniteDist <A> dist)
{
return(new SampleDist <A>(() =>
{
var rand = new MathNet.Numerics.Distributions.ContinuousUniform().Sample();
return dist.Pick(Prob(rand));
}));
}
/// <summary>
/// this is the implementation of algorithm 1 in the paper
/// </summary>
/// <param name="QFeatureNum"></param>
/// <param name="TFeatureNum"></param>
/// <param name="MaxIter"></param>
/// <param name="LearningRate"></param>
/// <param name="Lamda"></param>
/// <param name="StopGap"></param>
public void Train(int QFeatureNum, int TFeatureNum, int MaxIter, double LearningRate, double Lamda, double StopGap
,double decay = 1.0)
{
this.QUESTIONFEATURENUM = QFeatureNum;
this.TAGFEATURENUM = TFeatureNum;
this.learningrate = LearningRate;
this.lamda = Lamda;
this.MAX_Iter = MaxIter;
this.Stop_Gap = StopGap;
MathNet.Numerics.Distributions.ContinuousUniform normal = new MathNet.Numerics.Distributions.ContinuousUniform();
tagSimWeights = (Vector)Vector.Build.Random(TAGFEATURENUM, normal);
questionSimWeights = (Vector)Vector.Build.Random(QUESTIONFEATURENUM, normal);
for (int t = 0; t < this.MAX_Iter; t++)
{
loss = 0;
int incorrect_pair = 0;
DenseVector Deriviate_W = new DenseVector(QUESTIONFEATURENUM);
DenseVector Deriviate_V = new DenseVector(TAGFEATURENUM);
for (int i = 0; i < questions.Count; i++)
{
tagFrequency.Clear();
var neighbours = questionNeighbours[i];
var question = questions[i];
candidates = instanceCandidates[i];
if(t==0)
Init(question, neighbours);
else
{
tagSim = new DenseMatrix(candidates.Count, candidates.Count);
questionsim = new DenseVector(neighbours.Count);
tagSimFeatures = instanceTagSimFeatures[i];
questionSimFeatures = instanceQuestionSimFeatures[i];
tagQuestionFeature = new Dictionary<string, Vector>();
foreach (var candidate in candidates)
{
Vector tmp_v = new DenseVector(QUESTIONFEATURENUM);
for (int k = 0; k < QUESTIONFEATURENUM; k++)
{
for (int j = 0; j < neighbours.Count; j++)
{
if (neighbours[j].RelatedTags.Contains(candidate.Key))
tmp_v[k] = Math.Max(tmp_v[k], questionSimFeatures[k][j]);// We implement function g() by Max function
}
}
tagQuestionFeature.Add(candidate.Key, tmp_v);
}
}
ComputeTagSim();
DerivateForPi = null;
ComputeTagSignificance(neighbours);
ComputeQuestionSim(question, neighbours);
ComputeQuestionTagSimHeriusticlly(question);
ComputeDerivateofH();
RankingTags(question, neighbours);
foreach(var postag in question.RelatedTags)
{
if (!tagScore.ContainsKey(postag))
continue;
foreach(var negtag in question.UnRelatedTags)
{
if(tagScore[postag]<tagScore[negtag])
{
incorrect_pair += 1;
loss += -tagScore[postag] + tagScore[negtag];
Deriviate_W += DenseVector.OfVector(ComputeDeriviateOfW(postag, negtag));
Deriviate_V += DenseVector.OfVector(ComputeDeriviateOfV(postag, negtag));
}
}
}
for (int i2 = 0; i2 < QUESTIONFEATURENUM; i2++)
{
questionSimWeights[i2] -= (learningrate *(Math.Pow(decay,t)) * Deriviate_W[i2]);
}
for (int i2 = 0; i2 < TAGFEATURENUM; i2++)
{
tagSimWeights[i2] -= learningrate * (Math.Pow(decay, t)) * (Deriviate_V[i2]);
}
Deriviate_W = new DenseVector(QUESTIONFEATURENUM);
Deriviate_V = new DenseVector(TAGFEATURENUM);
}
if(lastLoss-loss<StopGap)
{
// break;
}
else
{
lastLoss = loss;
}
Console.Write("loss:{0}\tIncorrect_Pair:{1}\t", loss,incorrect_pair);
for (int i = 0; i < QUESTIONFEATURENUM; i++)
{ Console.Write("QuestionFeature{1}:{0}\t", questionSimWeights[i],i); }
for (int i = 0; i < TAGFEATURENUM; i++)
{ Console.Write("TagFeature{1}:{0}\t", tagSimWeights[i],i); }
Console.WriteLine();
}
PrintResult();
}
private void Init(Question question, List<Question> neighbours)
{
featureExtractor = new RandomFeatureExtractor();
MathNet.Numerics.Distributions.ContinuousUniform normal = new MathNet.Numerics.Distributions.ContinuousUniform();
tagSimWeights = (Vector)Vector.Build.Random(TAGFEATURENUM,normal);
questionSimWeights = (Vector)Vector.Build.Random(QUESTIONFEATURENUM,normal);
tagSimFeatures = featureExtractor.ExtractTagSim(candidates.Keys.ToList());
questionSimFeatures = featureExtractor.ExtractQuestionSim(question, neighbours);
tagSim = new DenseMatrix(candidates.Count, candidates.Count);
questionsim = new DenseVector(neighbours.Count);
}
public double[] GetVentSizes()
{
MathNet.Numerics.Distributions.ContinuousUniform uniformDistribution = new MathNet.Numerics.Distributions.ContinuousUniform(0.8, 1.5);
return(uniformDistribution.Samples().ToArray());
}
private Matrix RandomTagFeature(int rowcount, int columncount)
{
MathNet.Numerics.Distributions.ContinuousUniform normal = new MathNet.Numerics.Distributions.ContinuousUniform();
return (Matrix)Matrix.Build.Random(rowcount, columncount, normal);
}
private Vector RandomQuestionFeature(int count)
{
MathNet.Numerics.Distributions.ContinuousUniform normal = new MathNet.Numerics.Distributions.ContinuousUniform();
return (Vector)Vector.Build.Random(count, normal);
}
