ContentBased Music Information Retrieval in Wireless Adhoc Networks

1
Content-Based Music Information Retrieval in
Wireless Ad-hoc Networks

• Ioannis Karydis
• Alexandros Nanopoulos
• Apostolos Papadopoulos
• Dimitrios Katsaros
• Yannis Manolopoulos
• Aristotle University of Thessaloniki, Greece

2
A walk in the park
song excerpt
propagate
reply
song excerpt
reply
3
An emerging paradigm in music distribution

• The new trend is here wireless devices that can
  do much (lots of MHz!)
• The music industry found a blooming application
  music has turned into commodity over WWW
• How can we extend this success to the new trend
  of wireless networks?
• Is this another way to help piracy?
• No! Licensed distribution of digital music
  offers
• minimisation of distribution costs
• custom orders (track selection)
• instant delivery (temporal spatial)

4
What we need to make this true

• CBMIR for wireless P2P networks
• Consider the frequent alteration of the network
  topology
• Optimise the traffic for the constrained
  bandwidths of wireless networks (find effective
  representations of music data)
• Design the routing of music data over the
  wireless ad-hoc network

5
Why not existing (wired) solutions?

• In wireless ad-hoc networks two nodes can
  communicate only if in close proximity
  (in-range).
• Network peers
• participate randomly
• participate for short term
• change frequently their location.
• These factors cause existing approaches, e.g.,
  indexing, to become inapplicable.

6
Layout

• Background
• Problem definition
• Proposed method
• Experimental results
• Summary

7
Mobile ad-hoc networks

• Wireless mobile ad-hoc network (MANET)
• Collection of wireless mobile hosts
• Temporary network
• NO centralised administration
• NO standard support services
• The ad-hoc nature requires path discovery
• Need for routing policies in MANETs

8
Routing in MANETs

• Rely on some form of broadcasting, e.g.
• source-initiated on-demand routing protocols
• hybrid routing protocols
• Flooding is the simplest broadcasting approach
• each node in the network forwards a packet
  exactly once
• generates too many redundant transmissions gt
  broadcast storm problem
• To address flooding
• probabilistic approaches
• deterministic approaches

9
Layout

• Background
• Problem definition
• Proposed method
• Experimental results
• Summary

10
Problem definition

• Given a mobile client that wants to find music
  documents that are similar to a query, search all
  approachable peers in an MANET and return
  possible answers to the querier.

11
Layout

• Background
• Problem definition
• Proposed method
• Experimental results
• Summary

12
Template for CBMIR in MANETs

• User poses a query
• Query transformed to a representation form R
• R is broadcasted to all peers in range
• Qualifying sequences (true- and false- positives)
  comprise an answer-set
• Answer-sets are broadcast back to the querier
• Resolution of false-positives at
• peers that provide answers
• intermediate peers
• the querier
• Return of actual matches to the user/application

FWD traffic
BWD traffic
13
Options to represent the query

• The whole query sequence itself (time domain)
• Large size
• The first few coefficients of a frequency-domain
  transformation
• DFT, DCT,
• We choose DWT (Haar) transformation
• Small size
• A sample of the query sequence and the first few
  DWT coefficients
• Medium size

14
DWT (Haar) transformation provides

• simple but yet efficient representation of audio
  considering
• non-uniform frequency resolution,
• impulsive characteristics (C. Roads and Poli,
  1997)
• The Haar wavelet transformation
• is easy to compute incrementally,
• capable in capturing
• time dependant properties of data
• overall multi-resolution representation of
  signals (Kin-Pong Chan and Yu, 2003)

15
Options for false-alarm resolution

• At the qualifying peers
• Possible when using the whole query sequence
• No false-alarms
• At the querier
• When choosing representation only with DWT
  coefficients
• False-alarms (many!)
• At the querier, but intermediate peers help
• Significantly reduced number of false-alarms
• Intermediate peers prune many of them

16
Resulting approaches
17
ST example
18
Layout

• Background
• Problem definition
• Proposed method
• Experimental results
• Summary

19
Experiments

• Simulation test-bed
• 100 network nodes
• 300 songs (various music genres, e.g. pop, greek,
  rock, classical) average length 5 min
• Each song was randomly repeated 4 times
• Mobility simulator (GSTD)
• Area 4 km
• Peer radius 500m
• Peer velocity 5km/h
• Metrics
• average traffic
• time 1st and last result were discovered

2
20
Time of 1st last results vs. Max-hop
Increase in available Max-Hop gt more peers
examined gt longer times
21
Traffic vs. Max-hop
BWD phase is more demanding for all algorithms
22
Time of 1st last results vs. Range of query
23
Traffic vs. Range of query
24
Time of 1st last results vs. query size
increase in query size gt increased processing
required for the determination of matching
excerpts
25
Traffic vs. query size
increase in query size gt propagation of larger
representations
26
Traffic vs. NF parameter
High NF, limits the effectiveness of the policy
for the BWD phase, since most peers are selected
at random by this policy
27
Traffic vs. initial sample factor
Forward traffic increases with increasing sample
size
28
Layout

• Background
• Problem definition
• Proposed method
• Experimental results
• Summary

29
Summary

• Introduced CBMIR application in wireless ad-hoc
  networks
• Recognised new challenges posed by wireless
  ad-hoc networks.
• Proposed a novel algorithm, with twofold
  optimisation
• use of query representation with reducing length,
  
• selective policy for routing answers, which
  performs additional pruning of traffic.
• Result
• significant reduction in response times and
  traffic
• The examined context does not depend on specific
  features and distance measure

30
Content-Based Music Information Retrieval in
Wireless Ad-hoc Networks

• Thank you!