Information Technology Reference
In-Depth Information
binary scale, which makes classification feasible. The good results obtained with
the listwise approach, implemented as a Genetic Algorithm that optimizes MRR,
are probably due to the fact that this approach allows for optimizing the evaluation
criterion directly.
With respect to (2), the experimental results indicate that the classification and
regression techniques are equally capable of learning to classify the data in a point-
wise setting but only if the data are balanced (by oversampling or applying a cost
factor) before presenting them to a classifier. For regression techniques, balancing
is not necessary and even has a negative effect on the results. The results also show
that it is a good option to transform the problem to a pairwise classification task for
curing the class imbalance.
With respect to (3), the experimental results indicate that for the imbalanced
dataset, techniques with hyper parameters heavily depend on tuning in order to find
sensible hyper parameter values. However, if the class imbalance is solved by bal-
ancing the data or presenting the problem as a pairwise classification task, then the
default hyper parameter values are well applicable to the data and tuning is less
important.
14.3 Multimedia Retrieval
Multimedia retrieval is an important application, which has been supported by ma-
jor commercial search engines. Due to the semantic gap in multimedia data, in order
to provide satisfactory retrieval results, textual information such as anchor text and
surrounding text is still important in determining the ranking of the multimedia ob-
jects. To further improve the ranking performance, researchers have started to pay
attention to using visual information to re-rank the multimedia objects. The goal is
to maintain the text-based search paradigm while improving the search results. Ba-
sically the initial text-based search results are regarded as the pseudo-ground truth
of the target semantic, and the re-ranking methods are used to mine the contextual
patterns directly from the initial search result and further refine it. Several learning
based re-ranking methods have been proposed in the literature of multimedia re-
trieval, which are basically based on pseudo-relevance feedback and classification.
Most recently, learning-to-rank algorithms such as Ranking SVM [ 7 , 8 ] and ListNet
[ 4 ] have also been applied to the task [ 17 , 18 ]. When evaluated on the TRECVID
2005 video search benchmark, these learning-to-rank methods perform much better
than previous methods in the literature of multimedia retrieval.
While it is promising to apply learning to rank to multimedia retrieval, there are
also some differences between standard learning to rank and multimedia re-ranking
that need our attention. First, while standard learning to rank requires a great amount
of supervision, re-ranking takes an unsupervised fashion and approximates the ini-
tial results as the pseudo ground truth. Second, for standard learning to rank, the
ranking function is trained in advance to predict the relevance scores for arbitrary
queries, while for re-ranking the function is particularly trained at runtime to com-
pute the re-ranked relevance scores for each query itself. By considering these as-
pects, a re-ranking framework as described below is proposed in [ 17 , 18 ].
Search WWH ::




Custom Search