Context-based Global Expertise in Recommendation Systems

The concept of expertise can be simply defined as the skill or knowledge that a person has in a particular field; more formally, expertise is "the ability to discriminate meaningful classes of domain features and patterns, and to take decisions or actions that are appropriate to the class at hand" [9]. Apart from definitions, less trivial matters are how the expertise can be evaluated, and which effective applications it can have. The expertise assessment can be hard, expecially in nowadays virtual social networks (e.g. Facebook, or even e-commerce oriented, as eBay), due to the the lack of real person to person interactions used in real world to judge someone’s expertise level. Several approaches to this issue have been proposed [6, 12, 31, 32]. In particular, in [6] the authors aim at ranking the expert candidates in a given topic based on a data collection, hence they locate three components, i.e. a supporting document collection, a list of expert candidates, and a set of expertise topics. This work (as others) shows that an expertise rank is strictly related to a topic, so the question of which set of topics as well as their relationship (for instance, an arrangement into an ontology) should be addressed. That paper also highlights that expertise is commonly inferred from a set of documents (personal profiles, web pages, forum messages etc.) that represent the use case where expertise is applied, and they are the virtual counterpart of real world interactions between persons usually used to assess the each other’s expertise. The evaluation of an expertise rank by exploiting some data is not a new idea[13], and it has been successfully applied in other scenarios, e.g. Usenet news messages [26] or computer supported cooperative work (CSCW) [16]. In [12], topic expertise is ascertained by exploiting collaborative tagging mechanisms that enable the formation of social networks around tags or topics. Authors state that inferring expertise from data as personal profiles is problematic since users should keep them updated, and they also debate about the granularity in skill levels that should not be either too coarse or too fine (in the former case, automated systems have a difficult time selecting the right people, whereas in the latter users can hardly determine their levels in relation to others). The work presented in [31] is a propagation-based approach for expertise assessment that takes into account both person local information and relationships between persons; this raises the question of local vs global approaches that is frequent whenever complex networks are considered, hence not only in expert finding scenarios but also others (e.g. trust, recommendation systems etc.) In [32] the question of expertise within online communities is addressed, and network structure as well as algorithms are tailored to the case of Java Forum; this suggests that a fundamental role is played by the specific (possibily complex) network being considered and by its properties [18, 2]. In this work we present a method to rank people according to their expertise in a set of topics. We perform this assessment in an expertise network, i.e. where the relationship among nodes is the expertise rank assigned in a given context. In particular, we aim at evaluating the global expertise a node v has in a network within a specific context based on local expertise ranks assigned by v’s neighbor nodes. The placement of our work in comparison with issues highlighted in works cited so far can be schematized as follows: