Back to index
Semantic
Multicast: Intelligently Sharing Collaborative Sessions
Son Dao, Lixia Zhang, et al.
One-line summary:
Overview/Main Points
- Build channel-to-channel aggregation proxies, and allow
subscription on the outputs based on user profiles.
Sure, I knew that, but the hard part is how
to extract metadata describing the channel content. THis paper
assumes the metadata infrastructure and standard formats are
already in place, as in: "Various fields are associated with
weighted similarity ontologies."
- Basic construct: semantic multicast graph, where each node
represents a "range and quality" of semantic topics present in
the set of collaborative streams and user profiles. We add a
node dynamically if a "large set" of user profiles is likely to
be interested in it, and if the node would have access to input streams
close to the topic of interest.
- To compute this, intersect user profiles by comparing
nodes in the two profiles that have a "common parent" in
the similarity ontology that describes the topic
space, and let the intersection be the common parent
weighted by an error factor proportional to the weights
between the topics and their parent in the ontology.
The basic idea is that given some similarity weightings
between a subtopic and its parents, you can estimate how
similar two subtopics are to each other.
- Geographic criteria: try to instantiate nodes "close to" clusters
of users that would use them, to avoid a lot of overlapping
network traffic.
- Each aggregation proxy announces itself and independently
computes semantic multicast graph optimized for its user
population and the stream data of other proxies it can see.
- Similarly, each user announced himself to any proxy, and his
profile automatically gets forwarded to other appropriate proxies
that this one knows about.
Relevance
- Nice generalization of multicast with semantics subscription
groups.
- Nice application of multicast announcement and distributed soft
state (seems like each user can store his profile locally and
just announce it to any proxy at signup time) to "groups of
cooperating proxies" problem. But the paper doesn't seem to
exploit this, arguing instead for statefulness in proxies
remembering disconnection state of their users.
Flaws
- Metadata infrastructure and "similarity ontologies" are the hard
part. Standards and deployment aside, everyone else who's tried
to construct a static universal namespace for just abuot anything
has foundered. Would be nice if there were a way for namespaces
to describe themselves, as an extension of the ontology stuff;
then perhaps some minimal user-interaction with particular stream
servers could be used as the basis of how to derive the ontology
info for a particular set of sources.
- Each proxy is supposed to remember connected/disconnected state
of each of its users, so when a disconnected user reconnects,
proxy can "fast forward" thru session or allow user to catchup on
missed stuff. How this is accomplished is not clearly spelled out.
- Implementation as described is stateful, but could be statless or
SRM! (See "relevance" above)
Back to index