Ikumi Suzuki

Ridge regression, hubness, and zero-shot learning

This paper discusses the effect of hubness in zero-shot learning, when ridge regression is used to find a mapping between the example space to the label space. Contrary to the existing approach, which attempts to find a mapping from the example space to the label space, we show that mapping labels into the example space is desirable to suppress the emergence of hubs in the subsequent nearest neighbor search step. Assuming a simple data model, we prove that the proposed approach indeed reduces hubness. This was verified empirically on the tasks of bilingual lexicon extraction and image labeling: hubness was reduced with both of these tasks and the accuracy was improved accordingly.

Centered kNN Graph for Semi-Supervised Learning

Graph construction is an important process in graph-based semi-supervised learning. Presently, the mutual kNN graph is the most preferred as it reduces hub nodes which can be a cause of failure during the process of label propagation. However, the mutual kNN graph, which is usually very sparse, suffers from over sparsification problem. That is, although the number of edges connecting nodes that have different labels decreases in the mutual kNN graph, the number of edges connecting nodes that have the same labels also reduces. In addition, over sparsification can produce a disconnected graph, which is not desirable for label propagation. So we present a new graph construction method, the centered kNN graph, which not only reduces hub nodes but also avoids the over sparsification problem.

Reducing Hubness: A Cause of Vulnerability in Recommender Systems

It is known that memory-based collaborative filtering systems are vulnerable to shilling attacks. In this paper, we demonstrate that hubness, which occurs in high dimensional data, is exploited by the attacks. Hence we explore methods for reducing hubness in user-response data to make these systems robust against attacks. Using the MovieLens dataset, we empirically show that the two methods for reducing hubness by transforming a similarity matrix(i) centering and (ii) conversion to a commute time kernel-can thwart attacks without degrading the recommendation performance.

Reducing Hubness for Kernel Regression

In this paper, we point out that hubness--some samples in a high-dimensional dataset emerge as hubs that are similar to many other samples--influences the performance of kernel regression. Because the dimension of feature spaces induced by kernels is usually very high, hubness occurs, giving rise to the problem of multicollinearity, which is known as a cause of instability of regression results. We propose hubness-reduced kernels for kernel regression as an extension of a previous approach for kNN classification that reduces spatial centrality to eliminate hubness.

Annotating Cohesive Statements of Anatomical Knowledge Toward Semi-automated Information Extraction

Anatomical knowledge written in a textbook is almost completely unreusable computationally, because it is embedded in a cohesive discourse. In discourse contexts, the frequent use of cohesive ties such as reference expressions and coordinated phrases not only troubles the function of automated systems (i.e., natural language parsers) to extract knowledge from the resulting complicated sentences, but also affects the identification of mentions of anatomical named entities (NEs). We propose to revamp the prose style of anatomical textbooks by transforming cohesive discourse into itemized text, which can be accomplished by annotating reference expressions and coordinating conjunctions. Then, automatically, each anaphor will be replaced by its antecedent in each reference expression, and the conjoined elements are distributed to sentences duplicated for each coordinating conjunction connecting phrases. We demonstrate that, compared to the original text, the transformed one is easy for machines to process and hence convenient as a way of identifying mentions of anatomical NEs and their relations. Since the transformed text is human readable as well, we believe our approach provides a promising new model for language resources accessible by both human and machine, improving the computational reusability of textbooks.