I had a query regarding the model used in the paper here. What would be the problem in going for a hierarchical clustering approach?
What they did:
We solve problem (ii) by cross-validation, trying each value of S(q) (in our experiments, we consider up to 5 possible senses) and selecting the one that does best. To solve problem (i) we directly train our model, with a fixed number of senses S(q), so that the maximum sense score of a positive image is greater than the maximum sense score for a negative image, plus some margin.
This sounds a lot like the problem of not knowing how many clusters are present in an image segmentation problem. However, we can use a merge-and-split clustering algorithm which can grow the number of senses based on the responses to the senses.
An interesting example of this type of an algorithm is the ISOCLUS algorithm.
What would be the problems of using a Merge-and-Split clustering algorithm, removing weak senses(in which few data points lie) by merging and adding strong senses(in which more points lie?).
Paper with an introduction to ISOCLUS is available at: http://www.cs.umd.edu/~mount/Papers/ijcga07-isodata.pdf
On contacting the first author of this paper(Aurelien Lucchi), he told me that the clustering could worsen the ranking score used in the paper. I could not find any reasoning on why that should happen. could you perhaps provide some insight?
Also would scaling a merge and split based clustering algorithm like ISOCLUS be difficult?
