Since its invention in 1989 the world wide web has served to place vast libraries of resources at our fingertips [1]. Over time many of these resources have become too large to navigate manually. Now in the information age we seem to be constantly bombarded with choice in nearly every aspect of our lives, from which breakfast cereal to buy, to which songs to play on the drive into work. This is especially true in the case of online music streaming services like Spotify which boast catalogues of up to 20 million songs [2].
Faced with such choice we need some sort of automated system to narrow our focus to the content that we are actually interested in. Our current solution to this problem is to rely upon recommendation systems which attempt to match our preferences to a subset of the material available.
The main problem with this existing solution (particularly in artistic endeavours) is it's lack of personalisation. It is thought that too often users of these systems are simply being recommended a small set of obvious mainstream music based on broad and sometimes outdated metadata [3].
It is this personalisation problem which I am investigating throughout my PhD. By the completion of my PhD I hope to be able to demonstrate that it is a dual natured problem resulting at times from a lack of accuracy and at other times paradoxically a lack of spontaneity.
My current approach is to enable greater personalisation through interface design by allowing users to observe and manipulate the input to the recommender algorithms.
In order to research this I have designed a series of crowd-sourcing studies which build upon one another to create a narrative, leading to a fuller understanding of how user-centric controls can affect the personalisation of music recommenders.
I have decided to move away from highly qualitative interview based research towards a more quantitative style of research. This has several advantages in that it produces more falsifiable data which can be swiftly analysed and repeated for validity purposes. Furthermore plays to my strengths and background data analytics and statistics.
The first step in producing a system which allows users to filter input is to produce a taxonomy of attributes to enable them to categorise this input. To this end I have designed a crowd-sourcing studying in which participants are asked to identify a single user from a set of nine Last.fm profiles to match given search requirements e.g. ‘looking for a mellow folk song from the 1970’s’ or ‘looking for a song to listen with my mates’. Participants can use any of the the available information in the user profiles to help them make their decision.
The output of the study will be to produce the taxonomy of attribute that the participants most commonly used to identify their chosen user. I will also keep a log of which users where matched to which scenario most often. This will allow me to look for further similar traits between these users which participants might have not explicitly acknowledged.
Once a ground truth or base taxonomy has been produced from this study I intend to use it to design a follow up study in which participants can use the taxonomy to filter the entire user-base to reveal those users which they deem most appropriate to a given recommendation scenario. The idea is to see whether participants actually use the taxonomy in the way that they believe they do. Is the method by which they think they select users i.e. the characteristics they look for and filter by, the actual characteristics they use. Additionally is there any significant difference in participants behaviour depending on the scenario they are attempting to identify users for.
Leading on from the previous studies I hope to investigate the dynamic nature of music recommendations. One of the most common frustrations with conventional music recommendation systems is that they are too generic and only have one monolithic perspective of a given user’s tastes. In reality most people have a dynamic range of taste preferences depending on the context they are in e.g. jogging, going to work or relaxing with friends. In the system I am proposing this problem should be alleviated by the fact users can pre-filter the input to produce drastically different recommendations at different times.
Of course it remains to be decided how users should be presented with the controls to pre-filter this information. Are different interfaces more or less useful in different contexts? For instance is a user interface that displays users as friendship clusters more useful when recommending for social contexts? Is a genre based mapping of the user base more useful when looking for similar artist style recommendations?
The outcome from this aspect of my research will hopefully be to identify several good practises for designing interfaces to facilitate human in-the-loop dynamic music recommendations.
A deeper understanding of the personalisation problem as a paradoxical set of conflicting requirements exposes several novel areas for future research. One direction of investigation might be to identify a framework of set of best practises for designing interfaces to faciliate human-in-the-loop media input filtering. Another interesting area would be identifying specific moments or milestones in the development of a recommendation system when the personalisation problem becomes too complex to address in a purely algorithmic fashion.
This author is supported by the Horizon Centre for Doctoral Training at the University of Nottingham (RCUK Grant No. EP/G037574/1), by the RCUK’s Horizon Digital Economy Research Institute (RCUK Grant No. EP/G065802/1), and by FAST Grant Committee.