Information Technology Reference
In-Depth Information
hand, the use of both metadata and content-based
features confronts a number of real life scenarios
where a user may query based on a sample-file or
humming as well as by the singer's name or song
title. On the other hand, the disadvantages of the
use of metadata, such as the fact that there are not
always available, the inconsistencies introduced
by their manual appointment, the fact that their
sole use in MIR requires knowledge for the query
not provided by listening and the fact that their
description lacks customization since it relies on
predefined descriptors are compromised by the
co-existence of the feature extraction process. On
the networking side of their proposal, the use of
an decentralized system avoids scalability issues
as well as potential coordinator problems. Addi-
tionally, given the assumption that the proposed
system is structured, the use of an underlying
DHT-like scheme alleviates problems concern-
ing networking searching in order to identify the
host that may have a match to a query. Though,
such an advantage comes at the cost of requiring
the peers to host foreign data. Finally, the use of
the load-balancing scheme ensures that no peer
is over exploited due to the use of the underlying
DHT-based system, at the cost of redundant data
replication and increased processing require-
ments for the distributed dynamic load-balancing
mechanism to work efficiently.
network identifiers of its neighbors by broadcast
or other algorithms. Upon a query by any peer,
the feature of the query is extracted and sent to
m
peers, randomly selected from the neighboring
set. After receiving the feature of a query, each
peer sends it to
m
random neighbor peers, as long
as a maximum allowed hop number is less than
n
,
examines the local content and returns the results
to the querying peer.
This approach is identical to the random
breadth first search (RBFS) as also presented by
Kalogeraki et al. (2002), aside from the domain of
MIR, where the querying peer
Q
propagates the
query
q
not to all but at a fraction of its neighbor
peers.
In comparison to the centralized P2P CBMIR
systems, the stability of their distributed version,
such as the PsPsM, is far better since no coordi-
nator exists that can easily overload and become
the bottleneck of the whole system. However
the number of messages that are sent by PCs of
distributed P2P CBMIR systems increases when
the scale of the system expands. Accordingly,
the load produced for a given query by PsPsM
is smaller than that of either PsCM or PsC+M,
but by no means smaller than that of the hybrid
approach PsPsCM, as it will be seen in section
“Hybrid”.
MACSIS-PP
C. Yang (2003) proposed a peer-to-peer model for
music retrieval, where nodes are interconnected
as an undirected graph and each node stores a
collection of music files. Following the MACSIS
system, analysis of raw audio and conversion to
characteristic sequences are done locally at each
node, for both the database and queries. While
building the database, characteristic sequences for
each music file are stored in multiple LSH hashing
instances (each with its own set of randomised
parameters). The same music file may appear in
many different nodes and indexed under different
sets of hashing instances. At any time, every node
decentralized unstructured
PsPs Model
Wang et al. (2002) presented a distributed CBMIR
P2P model, where the system is formed by a nu-
merous of peers without a coordinator, while the
width of the system is
m
and the depth is
n
. Peers
consist of a music set, a same feature extraction
method, a common feature matching method and
a data structure that stores the network identifiers
of neighboring peers. The main steps of a query
process in this system, PsPsM, are the follow-
ing: Each peer shares some music and gets the
Search WWH ::
Custom Search