A Recommendation System is a sub-type of an information filtering system that is used to predict/select the most relevant items for a user given a goa

A Recommendation System is a sub-type of an information filtering system that is used to predict/select the most relevant items for a user given a goal, and potentially a context. For example, Andy would like to watch a comedy on Friday night. In this example, the goal is to find a comedy that Andy would enjoy, and the context is Friday night. A pertinent question then is: how well did our system do? Related work has long focused on measuring the effectiveness of RSs, with a big body of research devoted to devising novel metrics to measure the accuracy of produced recommendations and novel algorithms that produce more accurate recommendations based on these metrics. However, RSs can also be evaluated along another dimension: efficiency; that is, based on the resources required to train and to use a RS in practice.

Recommendation systems are expected to exhibit inherent efficiency-effectiveness trade-offs. In this project we thus make three main hypotheses: (a) The efficiency of a recommendation system can be modelled through a series of equations that encapsulate characteristics of the input, properties of the algorithms, and what quality of service constraints/goals are imposed/expected; (b) A predictive model can be built so as to estimate the efficiency-effectiveness curve for families of recommendation algorithms subject to predefined resource constraints; and (c) A methodology can be devised to allow arbitrary RS algorithms to be cast in this efficiency-effectiveness space so as to facilitate comparisons to other related algorithms.

Consequently, in this project we will examine the efficiency of recommendation systems, by mapping out their theoretical computational (space/time) complexity, actual training time and resource consumption, in tandem with the attained effectiveness quantified through various quality/accuracy metrics. Then, we will propose a methodology that maps out the design space of recommendation systems, by taking into account the recommendation task, the choice of algorithms, the available resources, the quality of service constraints/goals, and the effectiveness of the recommendations. The proposed methodology will encapsulate the efficiency-effectiveness of recommendation systems, highlighting the advantages and limitations in each case. In doing so, we also expect to uncover uncharted territories in the design space, possibly leading to new and improved RS algorithms.

The work in this project aligns with the "Content Creation and Consumption" and "Digital Business Models" priority areas of the Digital Economy EPSRC theme, and falls squarely in the "Information Systems" research area of the Information and Communication Technologies theme. This work stems from and augments research carried out as part of the EU-funded project PRIMES -- a collaboration among the University of Glasgow, HT2 Labs, and two secondary schools from France and the Netherlands.