/// <summary>
/// Finds duplicates in a list of <c>NewsItem</c>s and the database.
/// </summary>
/// <param name="archivist">
/// An <c>Archivist</c> instance with database access.
/// </param>
/// <param name="news">
/// The list of news to compare with the matrix.
/// </param>
/// <returns>
/// Returns the filtered list.
/// </returns>
private List<NewsItem> FilterDatabaseRedundancy(Archivist archivist,
List<NewsItem> news)
{
// Create result list containing all news items. Redundant news
// will later be removed.
List<NewsItem> result = new List<NewsItem>(news);
DateTime dateLimit = DateTime.Now.Subtract(
new TimeSpan(24 * 7 * 2, 0, 0));
SparseMatrix matrix = GetTermDocumentMatrix(archivist, news, dateLimit);
// If matrix == null, there are no news items in the db, and thus
// no redundant news.
if (matrix == null)
{
return news;
}
// Transpose matrix and thus prepare for multiplication.
SparseMatrix mTransposed = matrix.Transpose();
// Normalize all rows to make cosine similarity calculation faster.
mTransposed = mTransposed.NormalizeRows();
// Find redundancy value for each news item.
foreach (NewsItem item in news)
{
// Get tf-idf vector representing this news item.
SparseVector newsVector = archivist.GetTfIdfVector(item);
newsVector = newsVector.Normalize();
// Calculate product.
SparseVector product = mTransposed.VectorProduct(newsVector);
// Remove item from result list if product contains element
// with value >= threshold.
foreach (int entry in product.NonZeroIndices)
{
if (product[entry] >= THRESHOLD)
{
result.Remove(item);
break;
}
}
}
return result;
}
/// <summary>
/// Write the cosine similarity of the given <c>NewsItem</c>s to
/// the file cosine_similarity_log.txt.
/// </summary>
/// <param name="archivist">
/// The <c>Archivist</c> for getting database information.
/// </param>
/// <param name="news">
/// The <c>NewsItem</c>s to print the cosine similarity of.
/// </param>
private static void WriteCosineSimilarity(Archivist archivist, List<NewsItem> news)
{
// Sort the news according to title.
news.Sort((n1, n2) => int.Parse(n1.Title).CompareTo(int.Parse(n2.Title)));
// Get the vector of all news and add them to a matrix.
List<SparseVector> vectors = new List<SparseVector>();
foreach (NewsItem n in news)
{
vectors.Add(archivist.GetTfIdfVector(n));
}
// Create a matrix.
SparseMatrix matrix =
new SparseMatrix(vectors.First().Dimension, vectors.Count);
// Add entries to the matrix.
for (int i = 0; i < vectors.Count; i++)
{
for (int j = 0; j < vectors.First().Dimension; j++)
{
matrix[j, i] = vectors[i][j];
}
}
// Transpose matrix and thus prepare for multiplication.
SparseMatrix mTransposed = matrix.Transpose().NormalizeRows();
// Calculate the result.
SparseMatrix result = mTransposed.Product(matrix.Transpose().NormalizeRows().Transpose());
// Write to file.
using (StreamWriter file = new StreamWriter("cosine_similarity_log.txt"))
{
int rowLength = result.Rows;
int columnLength = result.Columns;
// Print header.
file.Write(" ");
for (int i = 0; i < result.Columns; i++)
{
file.Write("dok" + (i + 1).ToString("0#") + " ");
}
file.WriteLine();
// Print the matrix.
for (int i = 0; i < rowLength; i++)
{
// Print doc title.
file.Write("dok" + (i + 1).ToString("0#") + " ");
for (int j = 0; j < columnLength; j++)
{
file.Write(result[i, j].ToString("0.##0") + " ");
}
if (i != rowLength - 1)
{
file.WriteLine();
}
}
}
}
/// <summary>
/// Finds redundant <c>NewsItem</c>s in the given list,
/// and returns a new list without the <c>NewsItem</c>s.
/// </summary>
/// <param name="archivist">
/// An <c>Archivist</c> instance with database access.
/// </param>
/// <param name="news">
/// The list of <c>NewsItem</c>s to filter.
/// </param>
/// <returns>
/// A filtered list of <c>NewsItem</c>s.
/// </returns>
private List<NewsItem> FilterListRedundancy(Archivist archivist,
List<NewsItem> news)
{
List<NewsItem> result = new List<NewsItem>(news);
// Find redundancy for each news item.
for (int i = 0; i < news.Count - 1; i++)
{
// Get Tf-idf vector for current item.
SparseVector itemVector = archivist.GetTfIdfVector(news[i]);
itemVector = itemVector.Normalize();
// For all items with index j > i, calculate cosine similarity.
for (int j = i + 1; j < news.Count; j++)
{
// Get sparsevector from news item to compare.
SparseVector compareVector = archivist.GetTfIdfVector(news[j]);
compareVector = compareVector.Normalize();
// Remove from list if cosine similarity exeeds or equals
// threshold.
if (compareVector.DotProduct(itemVector) >= THRESHOLD)
{
result.Remove(news[i]);
break;
}
}
}
return result;
}
