Exploiting syntactic and neighbourhood attributes to address cold start in tag recommendation

摘要

Many state-of-the-art tag recommendation methods were designed considering that an initial set of tags is available in the target object. However, the effectiveness of these methods greatly suffer in a cold start scenario in which those initial tags are absent (although other features of the target object, such as title and description, may be present). To tackle this problem, previous work extracts candidate terms directly from the text associated with the target object or from similar/related objects, and use statistical properties of the occurrence of words, such as term frequency (TF) and inverse document frequency (IDF), to rank the candidate tags for recommendation. Yet, these properties, in isolation, may not be enough to effectively rank candidate tags, specially when they are extracted from the typically small and possibly low quality texts associated with Web 2.0 objects. In this work, we analyze various syntactic patterns (e.g., syntactic dependencies between words in a sentence) of the text associated with Web 2.0 objects that can be exploited to identify and recommend tags. We also propose new tag quality attributes based on these patterns, including them as new evidence to be exploited by state-of-the-art Learning-to-Rank (L2R) based tag recommenders. We evaluate our tag recommendation methods using real data from four Web 2.0 applications, finding that, for three out of our four datasets, the inclusion of our new proposed syntactic tag quality attributes brings improvements to two L2R-based tag recommenders with gains of up to 17% in precision. Furthermore, we find that recommendations provided by these methods can be further expanded exploiting the target object’s neighbourhood (i.e., similar objects). Our characterization and feature importance analysis results show that our syntactic attributes can indeed help discriminate relevant from non-relevant tags, being complementary to other, more traditional, tag quality attributes, particularly for datasets in which the textual features are short and / or present low quality.