/// <summary>
/// Calculate topic distinctiveness for each topic and return the average TD
/// TD for bad topic is treated as 0
/// </summary>
/// <returns>average TD</returns>
public double ComputeAvgTopicDistinctiveness()
{
var globalDist = this.wordTF.Select(x => (float)x / (float)this.totalTF).ToArray();
DenseVector globalDistVector = new DenseVector(globalDist);
for (int i = 0; i < this.NumTopics; i++)
{
if (this.topicsInfo[i].IsBadTopic)
{
continue;
}
var topicDist = this.topicWords[i];
DenseVector topicDistVector = new DenseVector(topicDist);
var td = 1 - VectorBase.CosineSimilarity(topicDistVector, globalDistVector);
this.topicMetrics[i, (int)MetricType.Distinctiveness] = td;
this.avgTopicDistinctiveness += td;
}
this.avgGoodTopicDistinctiveness = this.avgTopicDistinctiveness / this.topicsInfo.Count(x => !x.IsBadTopic);
this.avgTopicDistinctiveness = this.avgTopicDistinctiveness / this.NumTopics;
return(this.avgTopicDistinctiveness);
}
/// <summary>
/// Given a group of N vectors, compute NxN comparisions and return results in a matrix.
/// The comparison operator can be Distance or Similarity
/// </summary>
/// <param name="func">The comparison function to apply between pairs of vectors: Euclidean Distance or Cosine Similarity</param>
/// <param name="n">Limit comparisons to the top n vectors in the vector group. If less than 1, compute comps for all vectors in vectorGroup.</param>
/// <param name="maxCountItemsToCompare">For each of n vectors in the top, compare to the top maxCountItemsToCompare.
/// If less than 1, compare top n vs all vectors in vectorGroup.
/// If both n and maxCountItemsToCompare are less than 1 a full NxN compas are computed where N = |vectorGroup|
/// </param>
/// <returns>A diagonal matrix of doubles. Each cell encodes either the euclidean Distance or Cosine similarity between each vector pair.
/// The bottom triangule is set to 0.
/// </returns>
public double[][] GenerateSimDistMatrix(VectorFunction func, int n, int maxCountItemsToCompare)
{
if (this.RowCount <= 0)
{
return(null);
}
// Are we being asked to compare top n vectors vs. only maxCountItemsToCompare vector?
int vectorCount = ((maxCountItemsToCompare > 0) && (maxCountItemsToCompare < this.RowCount)) ? maxCountItemsToCompare : this.RowCount;
// Create a diagonal matrix: for document i we need to compute and store only (docCount - (i+1)) similarity or distance scores
// If asked to compare only the top n vectors againts the rest we only need to allocate a matrix of height n.
// Otherwise (n <= 0) we compare all (e.g. vectorCount) vs. all => need to allocate a matrix of height vectorCount
int matrixHeight = (n > 0) ? n : vectorCount - 1;
var funcMatrix = new double[matrixHeight][];
for (int i = 0; i < matrixHeight; i++)
{
funcMatrix[i] = new double[vectorCount - (i + 1)];
}
if (this.Verbose)
{
StatusMessage.Write(string.Format("Comparing top {0} vs. {1} other items...", matrixHeight, vectorCount - 1));
}
var timer = new Stopwatch();
timer.Start();
for (int y = 0; y < matrixHeight; y++)
{
for (var x = 0; x < funcMatrix[y].Length; x++)
{
if (func == VectorFunction.Distance)
{
funcMatrix[y][x] = Distance(this[y], this[x + (y + 1)]);
}
else
{
funcMatrix[y][x] = VectorBase.CosineSimilarity(this[y], this[x + (y + 1)]);
}
}
}
timer.Stop();
return(funcMatrix);
}
