³ÉÈË¿ìÊÖ

One area MyMedia explored was the integration of 'online update' methods into systems that require offline training. This allows sophisticated collaborative filtering systems to accommodate new users and items without having to wait for another time-consuming training cycle to be completed. We recently ran an internal field trial to see whether these online methods were effective and how well they could scale. The results were very promising, beating our best metadata-based method in terms of the perceived quality, variety and novelty of the recommendations.

A screenshot of the MyMedia recommendations interface

We are also exploring novel user interfaces and the possibility of providing recommender systems as a standalone application, rather than an integrated component of existing services. These interfaces aim to give the user more control over the recommendations and provide greater transparency. Although at an early stage it is becoming clear that providing this kind of control definitely reduces user frustration. Moreover, the additional feedback that is collected can make a significant improvement to the recommendations. Hopefully, it will be possible to make some of these prototypes publically available in the future, to determine whether a standalone recommender application would be effective and valued.

This week on our other projects:

Yves has spent some time on ABC-IP getting our automated tagging tool working on subtitles. Given a bbc.co.uk/programmes PID, get the corresponding subtitles, extract the text from them and extract tags from the text. It seems to give OK results, but hasn't been formally evaluated yet. In addition, he's also started to work on speaker segmentation and clustering for archive content, looking mainly at and got some very promising initial results from it.

The audio working group had a very short meeting this week, mostly for lack of participants (lots of people off at SXSW etc.) but it turned out to be a good way to get work done. We prepared a , and Olivier finished some work on the . Meanwhile, following a visit to the , Chris and Matt started working on small demos for our internal project: tape loops, phase vocoder, and even a wobulator.

On Chris N published the we created as part of the LIMO for live streams work and added a comment to the . The Firefox patch implements the startOffsetTime attribute for HTMLMediaElement objects for WebM streams. The team also noted some interesting discussion of how startOffsetTime should work following Sean's blog post on the .

Chris N has been working on our contribution to the next project deliverable document whilst Joanne has been analysing the data from last week's user evaluation sessions and writing up the summary report and findings. The team has also been starting to plan the next phase of the project in a bit more detail.

On Vicky B went to the Vrije Universiteit in Amsterdam to co-facilitate a workshop to gather feedback about the user experience of the project's user profiling service, including attitudes towards privacy, sharing, user control, and personalised recommendations. There's a great write up of the findings .

In other news, Michael joined the team this week and has spent his first few days quizzing everyone about what they're working on and generally getting up to speed on all our projects. Finally, here's a couple of interesting links from the team.

• is a zooming HTML5 timeline for "big history".
• Use your friends recommendations to .