- author = "Deepa Anand",
- title = "Enhancing Accuracy of Recommender Systems through various approaches to Local and Global Similarity Measures",
- year = "2011",
- school = "Computer and System Sciences, Jawaharlal Nehru University",
- address = "New Delhi, India",
- month = jul,
- keywords = "genetic algorithms, genetic programming, recommender systems",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/Anand_thesis.pdf",
- size = "172 pages",
-
abstract = "Web 2.0 represents a paradigm shift in the way that
internet is consumed. Users' role has evolved from that
of passive consumers of content to active prosumers,
implying a plethora of information sources and an ever
increasing ocean of content. Collaborative Recommender
systems have thus emerged as Web 2.0 personalisation
tools which aid users in grappling with the overload of
information by allowing the discovery of content in
contrast to plain search popularised by prior web
technologies. To this end Collaborative filtering (CF)
exploit the preferences of users who have liked similar
items in the past to help a user to identify
interesting products and services. The success of CF
algorithms, however, is hugely dependent on the
technique designed to determine the set of users whose
opinion is sought. Traditionally user closeness is
assessed by matching their preferences on a set of
common experiences that both share. The challenge with
this kind of computation is the overabundance of
available content to be experienced, at the user's
disposal, thus rendering the user-preference space very
sparse. The similarity so computed is thus unstable for
user pairs sharing a small set of experiences and is in
fact incomputable for most user pairs due to a lack of
expressed common preferences.
To remedy the sparsity problems we propose methods to enrich the set of user connections obtained using measures such as Pearson Correlation Coefficient (PCC) and Cosine Similarity (COS). We achieve this by leveraging on explicit trust elicitation and trust transitivity. When interacting with anonymous users online, in the absence of physical cues apparent in our daily life, trust provides a reliable measure of quality and guides the user decision process on whether or not to interact with an entity. These trust statements in addition to identifying malicious users also enhance user connectivity by establishing links between pairs of users whose closeness cannot be determined through preference data. In addition transitivity of trust can also be leveraged to further expand the set of neighbours to collaborate with. We first explore a bifurcated view of trust: functional and referral trust i.e. trust in an entity to recommend items and the trust in an entity to recommend recommenders and propose models to quantify referral trust. Such a referral-functional trust framework leads to more meaningful derivation of trust through transitivity resulting in better quality recommendations.
Though trust has been extensively used in literature to support the CF process, distrust information has been explored very little in this context. We thus propose a tri-component computation of trust and distrust using preference, functional trust and referral trust in order to densify the network of user interconnections. To maintain a balance between increased coverage and the quality of recommendations, however, we quantify risk measures for each trust and distrust relationship so derived and prune the network to retain high quality relationships thus ensuring good connections formed between users through transitivity of trust and distrust.
In the absence of supplemental information such as trust/distrust to provide extra knowledge about user links the local similarity connections can be harnessed to deem a pair of users similar if they are share preferences with the same set of users thus estimating the global similarity between user pairs. We investigate the effectiveness of various graph based global or indirect similarity computation schemes in enhancing the user or item neighborhood thus bettering the quality and number of recommendations obtained.",
-
abstract = "In addition to the inadequacy of similarity measures
such as PCC and VS in forming a rich user neighbourhood
they are static and may not capture user matching
satisfactorily and guarantee optimal performance under
diverse data situations. We propose to learn similarity
measures which not only adjust to the type of data at
hand but also ensure optimal performance over time.
Evolutionary techniques are employed to learn such
adaptive similarity measures.
Finally sparsity variant fusion of predictions from local and global similarity measures have been shown to offer quality recommendations. In particular the fact that local similarity measures suffice when the preference data is dense but overtaken in performance by global similarity links when preference data is scarce can be leveraged to fuse the recommendations from the two systems. We define sparsity not only for the overall system but also at the user and user-item level. We use these measures to suggest a fusion scheme tailored for each user and/or for each item to be predicted by estimating the apportionment of influence local and global similarity measures have on each prediction.
We demonstrate the effectiveness of the proposed techniques through experiments performed on real world datasets.",
- notes = "Supvervisor: K. K. Bharadwaj. JNU",
