+ SPE documentation

shogun-toolbox · Apr 3, 2012 · a985a56 · a985a56
1 parent 39731a4
commit a985a56
Showing 1 changed file with 26 additions and 1 deletion.
diff --git a/src/shogun/converter/StochasticProximityEmbedding.h b/src/shogun/converter/StochasticProximityEmbedding.h
@@ -30,7 +30,32 @@ enum ESPEStrategy
  * Dimensionality Reduction Toolkit) used to construct embeddings of data using 
  * the Stochastic Proximity algorithm.
  *
- * TODO more documentation
+ * Agrafiotis, D. K. (2002)
+ * Stochastic Proximity Embedding
+ * Retrieved from:
+ * http://www.dimitris-agrafiotis.com/Papers/jcc20078.pdf
+ *
+ * This class provides two different strategies for the computation of the embedding. 
+ * In each iteration, both strategies choose two sets of feature vectors whose 
+ * representation in the embedded space is updated. The first set is randomly chosen 
+ * in both strategies. However, the second set is obtained differently depending on 
+ * the strategy used. In the SPE_GLOBAL strategy, the second set is still 
+ * chosen at random. On the other hand, if SPE_LOCAL is used, first of all, the 
+ * K-Nearest Neighbors  of each of the feature vectors in the first set is obtained 
+ * and secondly, a number of feature vectors among these K-Nearest Neighbors is chosen
+ * to form the second set.
+ *
+ * The parameter K for K-Nearest Neighbors in SPE_LOCAL corresponds to the class member
+ * "m_k". Each of the two sets used on every iteration is formed by "m_nupdates" feature
+ * vectors. Therefore, the number of feature vectors given must be always at least two
+ * times the value of "m_nupdates".
+ *
+ * In order to avoid problems with memory in case a large number of features vectors is
+ * to be embedded, the distance matrix is never computed explicitily. This has the 
+ * drawback that it is likely that the same distances are computed several times during
+ * the process.
+ *
+ * Only CEuclidianDistance distance is supported for the moment.
  *
  */