8th - 14th July

Semantic Web Summer School, much heat, much fun, much learning... Here's an index of my posts.

15th - 21st July

Friends visited.  Progress included writing notes to myself to figure out just what my PhD outcomes really are, and why.  Came up with:

1. Recommending how to usefully describe diverse amateur creative digital content (ACDC) using an ontology.
    a) What are the parts of ACDC that need to be represented?  Identify and categorise properties. How do these differentiate it from other similar content?
    b) What existing ontologies can be used to do this, and how do they need to be extended?
   
2. Building an initial set of linked data about ACDC, and providing means for its growth and use.
    a) Manual annotation of ACDC, and refinement (to test ontology).
    b) Tools for automatic annotation of the parts of ACDC that it is possible to automatically annotate.
    c) Tools for manual annotation by the community of content creators and consumers for the parts of ACDC that cannot be automatically annotated.
    d) Tools to expose the linked data for use by third-party applications.

3. Create and test an example service which uses the linked data to benefit content creators and/or consumers.
     eg. Unobtrusive recommendations for collaborative partners (most likely); content recommendation; content consumption analysis (like tracking viral content); community building / knowledge sharing in this domain; ... .

22nd - 28th July

Brainstormed with Ewan about stage 3 (above), and came up with the idea of an interface that allows content creators to allocate varying degrees of credit for roles played by different people when collaborating on a project.  This would serve to both gather collaborative bibliographic data, learn things about how different segments of the community allocate credit, and provide a potentially useful tool for content creators.  With the future value that, if we can learn enough to estimate role inputs from different people, it could be used for things like automatic revenue sharing.

Then spent the rest of the week in London, frolicking amongst the YouTubers (including attending a meeting at Google about secret YouTube-y stuff), and annotated some ACDC.  Write-up coming soon.

Harith Alani talked about using semantics to solve problems around evaluating the success of social media use in business.  The SIOC ontology is widely used to describe online community information.  It's not as simple as measuring someone's engagement with a brand's online presence - people are 'likeaholics' on Facebook, so you have to look at someone's whole behaviour profile to judge whether their like means anything or not.  It's no good just aggregating your data and spewing out numbers - you have to browse the data and try to understand where it came from.

He mentioned how little work has been done in classifying community types.  Most of the work that has been done seems to be with social networks internal to an organisation.  A bottom-up approach to community analysis can handle emergent behaviours and cope with role changes over time.  Looking at behaviour categories and roles can help an organisation to decide who to concentrate on supporting and how in order to sustain the community.  The results they have seen so far suggest that a stable mix of the different types of behaviours are needed to increase activities in forums - but they don't know what causes what.  They're reaching a point where they can use their behaviour analysis to guess what's going to happen to a community: how long it will last, how fast it will grow, how many replies a certain type of post is likely to get, etc.

Next they want to be able to classify community types, and be able to look at activities within a community over a period of time and automatically discover what kind of community it is; it might be something different than what it was set up for.

They created an alternative Maslow's Hierarchy of Needs to correspond with activities seen on forums, and found that most people are happy to stay at the lower levels of the hierarchy.  For example, join a community, lurk for a bit, ask one question and leave.  Not everyone wants or needs to be a power user.

Papers are being written that find patterns in individual datasets for a particular community in a particular context.  Harith and his team are getting tired of this; they want to generalise across communities.  So they took seven datasets and looked at how the analysis features differed as well as comparing the results across community types, randomness (vs. topicality) of datasets, and compared similar experiments.

Upcoming work includes the Reel Lives project, in which UoE is involved.  They're taking media fragments - photos, videos, audio clips, text recorded as audio - and creating automated compilations to tell a story.

Another is social methods to change energy consumption behaviour.  LiSC in Lincoln did something in this area back in the day.. an app that posted that you were listening to an embarrassing song on your facebook feed if you left your lights on.

Notes from Harith's talk are here.

SSSW 2013 - Feeding Recommender Systems with Linked Open Data from Tommaso Di Noia

From Tommaso Di Noia's talk, I learnt that recommender systems have a lot of maths behind them, especially for evaluating things, and reinforced something I already knew: I don't maths good enough to be taken seriously by most of the Informatics world.  I think I understand the principles behind the maths, but when something is descried in just maths, I have no idea what it relates to.  I'll work on this.

Real world recommender systems use a variety of approaches, including collaborative (based on similar users' profiles); knowledge-based (domain knowledge, no user history); item-based (similarities between items); content-based (combination of item descriptions and profile of user interests).  Linked Open Data is used to mitigate a lack of information about entities, and helps with recommending across multiple domains.  You do have to filter the LD you use before feeding it to your recommender system though, to avoid noise.  Notes here.

Tommaso's talk was followed up by a hands-on session, where we got to poke about with some of the tools he mentioned, including FRED (transforms natural language to RDF/OWL); Tipalo (gets entity types from natural language text); and using DBpedia to feed a recommender system.

Then we worked on our mini-projects for the afternoon.  We made some progress towards breaking down the concept of serendipity and working out what properties we might need to represent as linked data, and how we could observer a user and work out if/when/how they were having serendipitous experiences without intruding too much.

In the evening we took a coach to 'nearby' historical town Segovia.  Apparently an extremely motion-sickness-inducing two and a half hour coach journey around twisty mountain paths is 'nearby'.  Fortunately I was distracted from this horrible journey by a conversation with Lynda Hardman, which I wish I had recorded.  Lynda challenged various aspects of my PhD until I could explain/justify them reasonably, including:

• Why digital creatives? (I'm used to that one now).
• What is the outcome?
• Why Semantic Web for this?

She also recommended a number of resources, including theses of her recent former students to help me with a structure for my own, and advice on maintaining a healthy balance between thinking and doing.

Plus she used to live in Edinburgh, more or less across the road from where I live now.  Cool.  Thanks Lynda!  You haven't heard the last of me :)

#travel

Once we got to Segovia, we had a guided tour of the ancient Roman architecture, interesting building façades and local legends.  It was a very good tour, but too hot to really focus.  Then they took us to a restaurant for a local speciality.  I was all set to write a whole individual blog post surveying the barbaric nature of human beings, but I didn't do it straight away and now the passion has faded slightly, so I'll leave it at a paragraph.  Some people watched the local 'ceremony' out of morbid curiosity I imagine, but it was the fact that so many people took so much pleasure in the idea of violently hacking up bodies of three-week-old piglets that really bothered me.  Fortunately the surging standing crowd allowed me (and only one other) to inconspicuously sit it out.  The veggie option was tasty, but it was difficult to really enjoy the rest of the evening whilst wondering vaguely about the states of minds of most of the people I was sharing a table with.

Tools and Techniques

Recommender systems

Input: Set of users + set of items + rating matrix.
Problem - given user, predict rating for an item.

In real world, recommendation matrix data is sparse.

Can use hybrid approaches.

Collaborative RS:

• Like Amazon.
• Based on other users with similar profiles.
• Experimentally better than content-based, but you don't always have many users.

Knowledge-based RS:

• No/little user history.
• Based on domain knowledge.

User-based collaborative recommendation:

• Pearson's correlation coefficient - baseline.
• Imagine millions of users - computing similarities takes a lot of time.
• So ..

Item-based collaborative recommendation:

• Focus on items not users.
• Compute similarity between each pair of items.
• Don't have to compute similarity between items that don't have overlapping ratings.
• Cosine similarity / adjusted cosine similarity (taking into account average rating related to a user to eliminate some bias).

Content-based RS:

• Based on description of item 
• and profile of user interests.

• Items are described in terms of attributes/features.
• Finite set of values associated with features.
• Item representation is a vector.
• Don't necessarily have complete descriptions of items - just have a 0 in your vector.

• Similarity between items: 
• Jaccard similarity.
• Cosine similarity and TF-IDF (term frequency - inverse document frequency).
• Batch compute similarities offline, then use similarities to compute ratings on the fly based on user profile.

• Predict rate only for N nearest neighbours of items in user profile, that are not in the user profile.
• An item is worth rating if more than x of N number of neighbours are within user profile.

Using LOD

To mitigate lack of information/descriptions about concepts/entities.

Recommender systems are usually vertical, but LD lets you easily build a multi-domain recommender system.

To avoid noisy data, you have to filter it before feeding your RS.

Freebase.


Tiapolo

• Automating typing of DBPedia entities.

Vector space model for LOD

• MATHS.

Social Media Analytics with a Pinch of Semantics

Using semantics to solve problems (not solving problems of semantics).

SM for businesses:

• Analytics.
• How to measure success?

SM silos impeding progress.
In-house social platforms increasing, so even more so.

SIOC to integrate online community information.
SIOC + FOAF + SKOS.

FB Graph.
People are likeaholics.  Their 'likes' become meaningless, so you need to take this into account when making recommendations.
Browse your data and understand user actions.

Behaviour analysis.

Bottom-up analysis.
Can handle unexpected or emerging behaviours.

• Community members classified into roles.
• Identify unknown roles.
• Cope with role changes over time.
• Clustering to identify emerging roles.

eg. focussed novice; mixed novice; distributed expert; ...
Spectrum across users you can or can't do without.

Extending an ontology built on SIOC.

Encoding rules in ontologies with SPIN.

Three categories of features:

• Social features (people you follow, people follow you, ...)
• Content features (what you're posting, keywords, ...)
• Topical/semantic features

Which behaviour categories you need to cater for more than others?  How roles impact activity in online community.

Consistently see that you need some sort of stable mixture of behaviours for activities in forums to increase.

==> Don't know what's causing which.

What is a healthy community?

Use behaviour analysis to guess what's going to happen to community. Eg.

• Churn rate.
• User count.
• Seeds/non-seeds prop (how many / if people reply to you).
• Clustering.

Unexpected: the fewer focused experts in the community, the more posts received a reply.
(But quality of answers?)

Community types (Little work in this space)

Muller, M. (CHI 2012) community types in IBM Connections:

• Communities of Practice
• Teams
• Technical support
• ..
• .. (see slides..)

Need an ontology and inference engine of community types.
Wants an automated process to tell you what type of community it is - it might be something it wasn't set up for.
Then you could determine what sort of patterns you would expect to find.
Noone has done this yet.

Measurements of value and satisfaction

Answers different across communities.  They ran it on IBM Connections - corporate community.

Most of this work is for managers of communities - see what's happening and help to predict what might be coming next.

Can classify users based on Maslow's Hierarchy of Needs?
Mapping the hierarchy to social media communities.
~90% users happily staying at the lower levels of the 'needs hierarchy'.

Behaviour evolution patterns

What paths they follow over time.
eg. people who become moderators eventually.

Engagement analysis

What's the best way to write a tweet so that people care about it?
Which posts are likely to generate more attention?

Getting bored of people finding patterns in individual datasets.  What can be generalised to other communities?

So experimented with 7 datasets and looked at how results differed across:

• community types.
• randomness (vs. topicality) of datasets.
• related experiments.

And people use different features.

Semantic sentiment analysis in social media

Too much research going on, especially on twitter.

Extract semantic concepts from tweets; likely sentiment for a concept.
Tweetnator.
Semantics increases accuracy by 6.5% for negative sentiment; 4.8% for positive sentiment.

OUSocial.
Students don't use in-house networks because they already use facebook groups etc. Want to analyse what's happening on them.

Upcoming

Reel Lives (inc. Ed.)
Fragmented digital selves.
Want to automate compilations of media (photos, messages) posted online.

Changing energy consumption behaviour.
Providing information is not enough.

Social Eco feedback technology.

The 10th Summer School for Ontology Engineering and the Semantic Web

Sunday

Arriving by train into Cercedilla, north of Madrid, we immediately encountered other confused looking folk with poster tubes.  So we shared taxis (EUR 10) from Cercedilla station to the summer school residence further north, in the forest.

After getting keys for our pleasant, single, en-suite rooms, arrivals congregated in the shade by the building  to introduce ourselves.. Again, and again, and again, as new people continuously arrived over the space of a few hours.

A really broad mix of people are here in terms of nationalities and places and levels of study, but I still haven't quite got used to the fact that answering 'Semantic Web stuff' is not specific enough in this crowd, when someone asks you what your research is about.  Nobody needs convincing that these technologies are useful!

Later we received schedules, maps, ill-fitting t-shirts* and very helpful name badges, and headed for dinner at the bar down the road.

As is traditional when I write about my experiences in new places, I will describe the food every day.  It has become apparent, at this residence at least, that variety of ingredients is not ordinary, so in this respect meals are simple.  Dinner that first night started with a salad (lettuce, olives, tomato, onion, shredded beetroot and a single slice of hard boiled egg; no dressing), followed by - for the majority - slices of meat (beef? Pork? I dunno..) and fries.  Mine was a plate of mushy green vegetables with a little seasoning, that was pretty tasty.  Dessert was a single pear, delivered with ceremony, but otherwise unadorned.  Healthy, at least.

Yet we were all (those I sat with at least) were left feeling a little unsatisfied.

I shared a table with a French, Spanish, Italian and Irish guy.  Conforming appropriately to stereotypes, and setting up reputations for the rest of the week, the French and the Italian shared the bottle of wine on the table; the rest of us went without.

I returned to bed after a couple of hours of socialising and enjoying the cool air in and around the bar.

* For next year, they could ask for t-shirt sizes when they ask for dietary preferences?

Monday

The day started early, and with no hot water or wifi for anyone.  Breakfast was combinations of sweet pastries, coffee, tea, juice and bread.

Punctuated variously by coffee breaks, the learning began in earnest.

During the introduction by Mathieu D'Aquin, I found out that I am one of 53 students selected out of 96 applicants to attend this year's Summer School of the Semantic Web!  I had no idea it was that selective, or that there had been that much competition.

The first keynote was by Frank van Harmelen, about all the Semantic Web questions we couldn't ask ten years ago.

Slides:

Semantic Web questions we couldn't ask 10 years ago from Frank Van Harmelen

Frank started by saying that the early Semantic Web vision has morphed into the more manageable vision of a Web of Data, or a Giant Global Graph, and outlined the principles of the Semantic Web as they appear to stand at present:

1. Give everything a name (entities).
2. Relations form graph between things.
3. Names are addresses on the Web (so we inherit properties of Web like AAA).
4. Add semantics.

Frank pointed out the advantages of the fact the Linked Data crowd, grown naturally and not designed, is now so big we don't know how many triples it contains, nor how fast it is growing.  Companies and organisations (like Google, NXP, BBC, DataGov) are using Semantic Web technologies to achieve their own ends, for a variety of different use cases, without caring much about the Semantic Web, and this is contributing to the growth.

This growth has given rise to a number of research areas that were impossible to realisitically ask questions about ten years ago, including self-organisation, distribution of data, provenance, dynamics and change, errors and noise (how to deal with disagreements).

Frank asserted that rules and structures, algorithms and patterns in data, exist whether we are looking at them or not.  He used the analogy that OWL is our microscope, and it may be the tool that distorts our vision of the information universe rather than properties of what we are looking at (for example, structures in data presenting themselves well in some domains but not others).

He went on to promote the roll of the Informatician to be to test theories, hypothesis and falsify, as scientists rather than engineers.  To discover, rather than build.

I struggle with this view of the world, and feel instinctively that theory and practice are intrinsically linked; one can't exist without the other, not just in the grand scheme of things, but in day to day work and research.  This is one of the main points of contention with my own PhD, and I've no doubt there will be many more blog posts about this issue in the near future as I reconcile my need to create something immediately useful with the necessity of producing a contribution to knowledge at large.

See my raw notes here.

We had an Introduction to Linked Data by Mathieu D'Aquin (raw notes here), followed by a workshop.  We wrote SPARQL queries to populate a pre-written web page with information about Open University courses, sub-courses and locations thereof.

Lunch, similar to the previous night's dinner, was a starter salad, an entire half chicken (or something) plus fries for the carnivores and the most unappealing risotto of my life for (not that I'm ungrateful, but I have never been unable to finish a meal due to boredom before).  I went for a walk with some others to grab some fresh air before the afternoon's work, and missed out on watermelon.

Manfred Hauswirth presented some really exciting stuff about annotating and using streams of data.  Particularly challenging is how to integrate this with static data and make inferences over the lot.  Streams include sensor data, as well as ever-flowing social media streams for example; anything that changes over time.

They've built some systems to process this kind of data, and one of them is available as middleware.

My raw notes are here.

In the afternoon we had a poster session, where all participants pinned up posters about their work, and discussed at length with anyone who was interested.  Here's evidence that I participated.

And here's Paolo's:

I wrote a few notes about things from other peoples' posters that I need to look up.

The main feedback I received was about making sure I focus, narrow down my topic, and concentrate on some evaluatable deliverables that are PhD-worthy.

Questions like (paraphrasing) "why should we care about digital creatives?" threw me, because I thought the obvious answer - that they are people too, Web users, technology users, contributors to culture and an ecosystem of digital content and data - was apparently not enough from an academic standpoint.

I was simultaneously told to focus more, and to explain why the problem I'm trying to solve is applicable to all domains, not just digital creatives.  But some of the problems I'm looking at have been (or are being) solved in other domains (like e-health, biological research, education) and the reason what I'm doing is interesting is because none of these solutions quite work for digital creatives, and I want to find solutions that do, and try to figure out why.

I'm still stuck in some sort of struggle between theory and practice; thinking and doing.  And the long-standing problem of how to decide which doing actually worked.

I've started scribbling notes about the narrowing down problem.  I'll need to have this figured out before my first year review in August anyway, so stay tuned for another post all about it.

Then I sneaked off for a nap.

Dinner at the bar again; the usual salad, plus some eggy fish thing for most.  I got a plate of artichoke.  Artichoke is great, I love it, and I'm all for simple meals.  But I remain unconvinced that a plate of only artichoke constitutes an acceptable level of effort on the part of caterers.  And the sheer quantity made it start to taste a bit funny after a while.  But not to worry; we rounded off with a solitary peach apiece.

Further socialising, and appreciation of the night sky, before returning to bed write blog posts.

I'm super excited and inspired by the talks, work I've heard about so far, and the atomsphere of the place.  I'm excited to learn a helluva lot, and remind myself that I'm not facing impossible problems, and am not facing many problems alone.  I remember that I am instinctively passionate about the Web and the possibilities it holds (and indeed has already realised) for the empowerment of individuals.  I remember how lucky I am to be able to sustain myself through studying something I love so much, and to have the potential to make a change, and through my work maybe even facilitate others to be able to make a living doing what they love, as well.

4th - 10th February

When helping to organise an event, it's impressive how much time sending emails, updating websites and spreadsheets, having meetings and generally coordinating things can take up.   So that's a lot of what I've done this week.  Not PhD, but related enough to be excusable... sort of... (Smart Data Hack and Open Data Day: Joined Up Edinburgh)

I have finally booked plane tickets to Serbia for Resonate new media festival, at which I'm hoping to make useful contacts with independent digital media producers, learn some stuff about big data visualisation at the workshops (my places at the ones I've applied to haven't been confirmed yet though), and generally learn more about the digital art scene (and figure out ways SW technologies can benefit people who are part of it).  And also to enjoy Belgrade, as I'll have a spare couple of days either side of the festival.  Because the flights were cheaper then!  Honest!  I actually lost an entire afternoon hunting for cheapest flights on websites that weren't obviously scams (Cheap-o-Airlines?  Really??  Ultimately booking through JAT directly was the best option), then negotiating cheapest trains to London in order to catch flights, then sending CouchSurfing requests, because why spend money on a hostel when you can meet wonderful new people and get tourist advice for free?

I also applied for a place at the Semantic Web Summer School in Spain in July.  It looks fantastic and educational and stuff.  And I applied for funding to help cover then 900EUR entry cost (a whole week, accommodation and meals included), and SICSA are providing £500, yay!

Two thirds of Ontologies with a View met this week, and we discussed BBC use of Linked Data.  I only managed to skim the paper, but knew the general principles from articles I'd read about their work before... I will read it properly at some point, though I think a more technical discussion of what they did might be useful.