An Approach to Content Adaptation for Mobile Computing

1
An Approach to Content Adaptation for Mobile
Computing

• Francis C.M. Lau ( W.Y. Lum)
• Department of Computer Science Information
  SystemsThe University of Hong Kong

2
The Next Gold Rush?

• The current Internet gold rush will be
  dwarfed
  by what is about to happen with
  Wireless Internet
  Access. The Economist

3
Challenges of Mobile Computing

• Device heterogeneity and constraints
• Content heterogeneity
• The network
• The user

4
Content Adaptation

• Why adapt contents?
• Most contents for viewing are for the larger
  screens
• Creating multiple versions a burden
• Even if you dont mind, there are just too many
  possible devices
• Different users want different things
• Having one, original version is easier to manage
• Content adaptation is about generating any
  content version from one single original version
• Author-once-present-anywhere (AOPA)

5
Preadaptation

• Keeping just the original version (any other
  version is runtime-generated) could be slow
• Preadaptation
• to create all possible versions, and do static
  selection at runtime, or
• to create just a few essential versions, and do
  dynamic adaptation hence the balance

6
A User-Centric, Context-Sensitive Approach

• Content adaptation is not just about adapting to
  the device, but also the user, the network,
• Adapting to the user a user-centric approach
• The user has preferences
• speed (how much delay can I tolerate?), color (do
  I mind black-and-white?), scaling (is smaller
  text size ok?), modality (do I care what
  format?),
• What the user most prefers however might not be
  feasible because of constraints of the context
• the device, the network, the requested object

7
A Collaborative Environment
Content adaptation happens here
INTERMEDIATE PROXY SERVER
CONTENTS PROVIDER
INTERNET
USER DEVICE
THE CONNECTIONS
Preadapted versions stored here
8
Content Adaptation in Two Steps

• Content negotiation
• N(Ppreference, Pdevice, Pnetwork, Pcontent) ? V
• V the recommended version
• Note version ID or metadata, not the real
  content
• Content realization
• R(V) ? O
• O the actual object returned to the client

9
(No Transcript)
10
Content Negotiation

• Negotiation users preference againstthe
  context
• Preferences represented by scoresand stored as
  score nodes in an efficient data structure
• To traverse the data structure from
  highest-scoring node until TRUE
• TRUE FALSE ? decision(score-node, context)
• where context (Pdevice, Pnetwork, Pcontent)
• note that estimated rendering time is considered
  in the process

11
CONTENTNEGOTIATION
recommendedversion

12
Negotiating for the Best Version
13
Content Realization

• To generate the object based on the desired
  version recommended by the negotiation module
• Involves one or more transcoding steps from some
  optimal preadapted version
• Tradeoff between
• real-time transcoding overhead (CPU cost, or
  time)
• storage overhead of preadaptation (I/O cost)

14
Transcoding Relation Graph

• V the set of all possible contentversions
• The edge (vi, vj) means vj can bederived from vi
  throughtranscoding
• vi could be (4-bit color, 75 scaling)vj could
  be (1-bit color, 50 scaling)
• Transcoding (?) is a lossy operation
• Edge labels are the time cost of the
  corresponding transcoding operations based on
  some cost model
• At least v0, the original content version, should
  be present in the content server

15
To Build the Preadapted Set, Vpre

• Constrained by total size allowed
• Each vertex (white) not in Vpre must be pointed
  to by exactly one edge from a vertex (black) in
  Vpre
• With least total edge cost (over all edges from a
  black vertex to a white vertex) among all the
  possibilities
• NP-complete

16
The Greedy Algorithm (GREEDY)
Vpre ? initial setwhile not exceeding space
allowed select v ? Vpre such that C(v ? Vpre) is
minimized add v to Vpre

• C(V) total edge (black-to-white) cost based on
  an optimal edge set for a given preadapted set V
• We can take space into account as well to
  maximize C'( ) which is the aggregated
  transcoding cost saving per unit spatial
  consumption

17
Example
Space limit 850 Kbytes
Vpre C( ) space
v0 13.6 500
v0, v1 8.6 750
v0, v2 8.5 700
v0, v3 10.2 580
v0, v4 7.7 600
v0, v4, v1 3.5 850
v0, v4, v2 4.3 800
v0, v4, v3 6.8 680
18
How Good is GREEDY

• Let A and B be the improvements (i.e., reduction
  in transcoding time over all content versions)
  due to the optimal solution (OPT) and GREEDY
  respectively then
• where k and k' are the numbers of versions
  selected by GREEDY and OPT respectively
• If k9 and k'8, then GREEDY is at least 70 of
  OPT in performance

19
How Good is GREEDY

• Proof based on that for a greedy selection
  algorithm for datacubes by Harinarayan et al.
  SIGMOD96
• Can we do better?
• The greedy algorithm does as well as we can hope
  any deterministic polynomial-time algorithm to
  do according to some recent result on set
  cover

20
Experimentation

• A prototype PDF document content adaptation
  system (simulation)
• User preference in five domains color,
  downloading time, scaling, modality,
  segmentation, each having a range of 4 values
• hence 45 1024 score nodes per user

21
Negotiating for the Best Version
22
Experimentation

• We measured performance in terms of the following
  against preadaptation capacity
• aggregated transcoding cost saving
• content coverage ratio of selected versions /
  all versions
• C'( ) performs better than C( ) in most
  situations
• Please refer to our Mobicom paper for the graphs

23
Experimentation

• Modality vs. downloading time all others kept
  constant
• WAP device
• (a) modality gt downloading time ? WBMP
• (b) downloading time gt modality ? WML
• PDA
• (c) PDF, (d) BMP, (e) HTML

24
Experimentation
a
c
b

• Setting maximum download time
• (a) WBMP, (b) larger WML, (c) smaller WML
• note how (c) is segmented/cropped and the use of
  the next anchor

25
Experimentation
a
c
b

• Awareness of network delays
• (a) 144 kbps ? PDF with 256 colors
• (b) 19.6 kbps ? BMP with 16 colors
• (c) 9.6 kbps ? HTML

26
Further Research

• Versions weighted according to popularity
• To exploit mutual dependencies between objects
• Dynamic (re-)preadaptation
• similar to caching
• separate caching at the proxy?
• Better algorithms than GREEDY
• Automatic content augmentation pervasive
  authoring
• Easily-transcodable contents
• Do we really need that many versions?
• User preferences how specified?
• Collaborative design device-proxy-server
• Adaptation of code

27
Related Publications

• W.Y. Lum and F.C.M. Lau, User-centric Content
  Negotiation for Effective Adaptation Service in
  Mobile Computing, IEEE Transactions on Software
  Engineering, to appear.
• W.Y. Lum and F.C.M. Lau, A Context-Aware
  Decision Engine for Content Adaptation, IEEE
  Pervasive Computing, Vol. 1, No. 3,
  July-September 2002, 41-49.
• W.Y. Lum and F.C.M. Lau, On Balancing Between
  Transcoding Overhead and Spatial Consumption in
  Content Adaptation, Proc. Mobicom 2002, Atlanta,
  USA, September 2002, 239-250.