MARKOV RANDOMFIELDS AND MAXIMUMENTROPY MODELING

FOR MUSIC INFORMATION RETRIEVAL

Jeremy Pickens and Costas Iliopoulos
Department of Computer Science
Kings College London

London WC2R 2LS, England

jeremy,[email protected]
ABSTRACT
Music information retrieval is characterized by a number
of various user information needs. Systems are being de-
veloped that allow searchers to nd melodies, rhythms,
genres, and singers or artists, to name but a few. At the
heart of all these systems is the need to nd models or
measures that answer the question how similar are two
given pieces of music. However, similarity has a variety
of meanings depending on the nature of the system be-
ing developed. More importantly, the features extracted
from a music source are often either single-dimensional
(i.e.: only pitch, or only rhythm, or only timbre) or else as-
sumed to be orthogonal. In this paper we present a frame-
work for developing systems which combine a wide vari-
ety of non-independent features without having to make
the independence assumption. As evidence of effective-
ness, we evaluate the system on the polyphonic theme
similarity task over symbolic data. Nevertheless, we em-
phasize that the framework is general, and can handle a
range of music information retrieval tasks.
Keywords: Random Fields, Music Modeling
1 INTRODUCTION
Recent interest in the area of music information retrieval
and related technologies is exploding. Digital music col-
lections are becoming available locally (computer hard
disks, mp3 players) and remotely (online music stores, di-
gitial libraries). However, few of the existing techniques
take advantage of recent developments in statistical mod-
eling. In this paper we discuss an application of Markov
Random Fields to the problem of creating accurate yet
exible statistical models of music. With such models
in hand the challenges of developing effective search-

This work is supported by EPSRC grant GR/N63123/01

Permission to make digital or hard copies of all or part of
this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for prot or
commercial advantage and that copies bear this notice and the
full citation on the rst page.
c 2005 Queen Mary, University of London
ing, browsing and organization techniques for the growing
bodies of music collections may be successfully met.
Random Fields are a generalization of Markov chains
to multi-dimensional spatial processes. They are incred-
ibly exible, allowing us to model arbitrary interactions
between variables. They have been very popular in mod-
eling of physical systems, and recently have demonstrated
superior performance in a number of language-related ap-
plications such as Berger et al. (1996), Della Pietra et al.
(1997) and Lafferty et al. (2001). Finally, and perhaps
most importantly, random elds are extremely attractive
from a theoretical standpoint. Probability distributions
over the elds have exponential form, consistent with the
maximum-entropy framework for inference. The object-
ive function is globally -convex with respect to paramet-
ers of the model, and so parameters can be learned effect-
ively through iterative methods. Furthermore, there exists
a principled way of inducing the structure of the eld that
guarantees improvement in the objective function, and in
some cases allows closed-form solutions.
2 MUSIC REPRESENTATION
Music has several possible representations. In its most un-
structured form, music can be represented as a sequence of
audio signal samples, as for example in a wav or mp3 le.
On the other end of the spectrum, music may be repres-
ented as instructions to a performer, as in sheet music. In
the middle are representations such as MIDI, in which the
onset, duration, and pitch of every note in a piece of mu-
sic is known but no other information is necessarily avail-
able. Not to be overlooked, music may also be represented
using metadata, such as song title, artist, genre, year, in-
strumentation and so on. For the evaluation of our current
model, we will focus on symbolic pitch information. But
we emphasize that the model in general can accomodate
features from any representation, including both content-
and metadata-related features, simultaneously, as long as
the feature is the answer to a binary question. This will be
covered in more detail at the end of the paper.
In polyphonic music, we may think of the notes as
a two-dimensional graph, with ticks (time) along the x-
axis, and MIDI note number (1 to 128) along the y-axis.
At any point along the y-axis, notes turn on, remain on
for a particular duration, and then turn back off again.
As a quick example, see the gures below. Black circles
207
represent notes being on. White circles represent notes
being off.
Figure 1: Example music sequence, preprocessed
For our current purposes (though not necesssary in
general) we make the following simplications. We begin
by selecting only the onset times of each new pitch in the
sequence, ignoring the duration of the note. Next, we re-
move all onset times which contain no pitches. (Thus, we
are throwing away not only the duration of the notes them-
selves, but the duration between notes, including rests) Fi-
nally, we reduce the 128-note y-axis to a 12-note octave-
invariant pitch set by taking the mod-12 value of every
symbolic pitch number. Our example above becomes:
Figure 2: Example music sequence, postprocessed
3 MARKOV CHAIN RETRIEVAL
MODEL
In the previous section we showed how polyphonic mu-
sic can be represented as a temporal progression of 12-
dimensional binary vectors. In this section we introduce
as a baseline approach a simple modeling and retrieval
system based on extracting multiple, sometimes overlap-
ping Markov chains from this sequence.
A Markov chain model proceeds in three stages. In
the rst stage, we extract the chains (or features). In the
second stage, we use the relative frequency counts of these
extracted chains to estimate a state transition distribution.
In the nal stage, we compare models estimated from both
a music query and a collection document and rank collec-
tion documents by this similarity score.
For feature extraction, we simply count the number of
length m sequences through a piece of music. This ap-
proach is broadly similar to that of Doraisamy and R uger
(2001). For example, if we extract chains of length m=2
from the music in Figure 2, we end up with the following
set of sequences: {0;2, 2;1, 1;0, 1;2, 0;1, 2;1}.
In the second stage, we divide these chains into two
parts, the previous state, or history, and the current
state. We dene the history H as the rst m-1 notes in
the sequence, and the current state n as the last note in the
sequence. For example, with an m=2 chain 1;2, the
history is the state 1 and the current state is 2. With
an m=3 chain 1;2;1, the history is the state 1;2
and the current state is 1
We next obtain parameters for the conditional probab-
ility distribution

P(n|H) by doing maximum likelihood
estimation on the complete set of extracted chains, where
|H, n| is the number of times the chain with history H
followed by note n is observed.

P(n|H) =
|H, n|

H
i
|H
i
, n|
(1)
Basic Dirichlet smoothing is used to overcome the
zero frequency problem (Zhai and Lafferty). Figure 3
shows an (unsmoothed) example for a sample 3-state
model trained with the data from Figure 2.
Counts
0 1 2
0 0 1 1
1 1 0 1
2 0 2 0

P(n|H)
0 1 2
0 0 0.5 0.5
1 0.5 0 0.5
2 0 1.0 0
Figure 3: Example 1
st
-order (m = 2) state transition
counts (left) and state transition distribution (right)
Prior to retrieval, at indexing time, we estimate

P(n|H) for every piece of music in the collection. At

retrieval time, when presented with a query, we estimate
a model for the query in the exact same manner. Similar-
ity is calculated between the query and every document in
the collection using the Kullback-Leibler (KL) divergence
measure. Documents are ranked by increasing dissimilar-
ity to the query.
4 MARKOV RANDOMFIELD
RETRIEVAL MODEL
In the previous section, we modeled polyphonic music as
a sequence of Markov chains across neighboring 12-bit
binary note vectors. However, there is one main prob-
lem with this method: Since the music is polyphonic, the
extracted Markov chain features overlap. This overlap
causes the model to count certain notes more than once,
which has the effect of overestimating certain transitions.
Using Markov chains makes the assumption that notes oc-
curing at the same point in time are independent of each
other; this is clearly not true. Therefore, we turn to a
framework that allows us to selectively model the inter-
actions (dependencies) between the individual notes.
4.1 Nature of the Field
Suppose we are given a musical piece represented by T
simultaneities (binary vectors of length 12). With this
piece we will associate a lattice of binary variables {n
i,t
},
indexed by the time t = 1. . .T, and by note index
i = 0. . .11. Each variable {n
i,t
} can be either 0 or 1,
indicating whether a note i is on or off at time t. Fig-
ure 4 illustrates the lattice. Our goal is to develop a model
that will allow us to predict the value n
i,t
from the values
of the surrounding variables. In other words, we would
like to develop an estimate for the probability distribution
P(n
i,t
|{n
j,s
: j=i or s=t}). It is important to stress that
we do not want to assume independence among the condi-
tioning variables, or restrict the conditioning to the imme-
diate neighbors of n
i,t
. On the contrary, we believe that
208
n
0,1
n
0,2
n
0,3
n
0,4
. . . n
0,t
. . .
n
1,1
n
1,2
n
1,3
n
1,4
. . . n
1,t
. . .
n
2,1
n
2,2
n
2,3
n
2,4
. . . n
2,t
. . .
. . .
n
11,1
n
11,2
n
11,3
n
11,4
. . . n
11,t
. . .
Figure 4: Variables of a Musical Markov Random Field
the value of n
i,t
is strongly inuenced by both its short-
range and long-range neighbors in the lattice. However,
for the scope of this paper we will impose several limita-
tions on what kind of dependencies may exist in our eld.
The rst limitation we impose concerns the temporal
nature of music. In the most general formulation of a ran-
dom eld, the value of the note n
i,t
may be inuenced by
both the notes that precede it {n
j,st
}, and notes that fol-
low it {n
j,s>t
}. For comparison with the chain model, in
our initial random eld model we will restrict the depend-
encies to only those notes that precede the target note in
the sequence. For every note n
i,t
we dene the concept
of history or neighborhood H
i,t
to include the notes that
either occur before time t, or notes that occur at time t, but
have an index lower than i:
H
i,t
= {n
j,s
: s < t} {n
j,s
: s = t, j < i} (2)
Notes in H
i,t
are the ones that can be examined when
we are making the prediction regarding n
i,t
. In other
words, we assume that the probability of note i playing at
time t is completely determined by H
i,t
in our model. We
reiterate that we still allow arbitrary dependencies within
the subset dened by the neighborhood H
i,t
.
In general, we need not constrain our neighborhood
in this manner. We believe that better models could be
created by considering notes that will be played in the
future, not just in the past. However, we wish to do
two things with this paper: (1) introduce random elds
as a modeling framework suitable for music information
retrieval, and (2) evaluate this framework by comparing
it with a similar model (Markov chains). Markov chains
only consider the past. If we were to introduce a version of
the random eld that considered the future as well as the
past, then it would be unclear as to whether performance
gains of elds over chains were due to inherent modeling
superiority, or to the fact that the elds had use of more
data. Thus, we constrain the eld neighborhood in such
a manner as to make it fair and analogous as possible to
the chain models. Future work (no pun intended) need not
adhere to this constraint.
4.2 Conjunctive Features
The second limitation we impose on the conditional prob-
ability P(n
i,t
|H
i,t
) concerns the nature of dependencies
that will be modeled by a eld. In general, a random
eld framework allows arbitrary dependencies (or fea-
tures) between the target n
i,t
and its neighborhood H
i,t
.
For example, n
i,t
may depend on the answer to the fol-
lowing question: what is the total number of times note
i was played in the history H
i,t
?. For the sake of sim-
plicity and elegance we will deliberately restrict allowed
Figure 5: Examples of musical features that may be in-
duced to predict the probability of note 2 being played
at time t. Black circles represent notes that are part of
the feature function. Boxed black circle denotes the note
n
2,t
. Boxed area represents the history H
2,t
. From left to
right, the features are: {n
2,t
n
1,t
}, {n
2,t
n
2,t1
n
2,t2
},
{n
2,t
n
1,t
n
3,t1
n
3,t2
}, {n
2,t
n
0,t
n
2,t2
n
0,t2
}
dependencies to binary questions of the form: was note j
played at time s before t?. We will also allow generaliza-
tions where a question is asked about some subset S of the
notes in the allowed history H
i,t
. The answer to a ques-
tion of this form will be called the feature function f
S
,
and S will be referred to as the support of f. For a given
support S H
i,t
, the feature function f
S
is dened as
the conjunction of answers about the individual notes in
n
j,s
S:
f
S
(n
i,t
, H
i,t
) = n
i,t

n
j,s
S
n
j,s
(3)
Dened in this manner, our feature functions are al-
ways boolean, and equal to 1 if all the notes dened by S
were played before the target note n
i,t
. A feature function
always includes the target note n
i,t
. This is not a fallacy
in the model, since n
i,t
will never actually be considered
a part of its own history. Presence of n
i,t
in the feature
serves only to tie the occurrences of notes in S to the oc-
currence of n
i,t
. If the feature is considered likely, that
is evidence in favor of predicting n
i,t
= 1. If the feature
does not occur, it suggests that n
i,t
is likely to be zero.
One nal comment: we choose to make features time-
invariant, but not index invariant. This means that a fea-
ture is expected to characterize the same kind of depend-
ency, regardless of the time index t of the target n
i,t
, but
that the feature is not (pitch) transposition invariant. Con-
sequently, we will index the time component of the notes
in S not in absolute values but relative to the time t. We
do not do the same for the note index i, so these indices
will remain absolute. As an illustration, Figure 5 contains
some examples of features that could have an impact on
note 2 at time t.
4.3 Exponential Form
At this point we are ready to select the parametric form
that we will be using for computing the probabilities
P(n
i,t
|H
i,t
). There are a number of different forms we
could choose, but it turns out that for random elds there
is a natural formulation of the distribution that is given by
the maximum-entropy framework. Suppose we are given
a set F of feature functions that dene the structure of
the eld. The maximum-entropy principle states that we
should select the parametric form that is: (i) consistent
with the structure imposed by F and (ii) makes the least
209
amount of unwarranted assumptions that is the most
uniform of all distributions consistent with F. The family
of functions that satises these two criteria is the exponen-
tial (or log-linear) family, expressed as:

P(n
i,t
|H
i,t
) =
1
Z
i,t
exp
_
_
_

fF

f
f(n
i,t
, H
i,t
)
_
_
_
(4)
In equation (4), the set of scalars = {
f
: f F}
is the set of Lagrange multipliers for the set of structural
constraints F. Intuitively, the parameter
f
ensures that
our model predicts feature f as often as it should occur
in reality. Z
i,t
is the normalization constant that ensures
that our distribution sums to unity over all possible values
of n
i,t
. In statistical physics, it is known as a partition
function and is dened as follows:
Z
i,t
=

n
exp
_
_
_

fF

f
f(n, H
i,t
)
_
_
_
(5)
For a general random eld, the partition function Z
i,t
is exceptionally hard to compute, since it involves sum-
mation over all possible states of the system. In a typical
system the number of states is exponential in the number
of eld variables, and direct computation of equation (5)
is not feasible.
In our case, our assumption of no dependencies
between the current state notes n
i,t
makes computation
of the partition function extremely simple. Recall that
all underlying eld variables are binary, so equation (5)
only needs to be computed for two cases: n
i,t
= 0 and
n
i,t
= 1. We can further simplify the problem if we re-
call that every feature function f is a binary conjunction,
and every f includes n
i,t
in its support. As a direct con-
sequence, f(n
i,t
, H
i,t
) is non-zero only if n
i,t
= 1. The
assertion holds for all feature functions fF, which im-
plies that the summation inside the exponent in equations
(4) and (5) is zero whenever n
i,t
= 0. These observa-
tions allow us to re-write equation (4) in a form that allows
very rapid calculations:

P(n
i,t
= 1|H
i,t
) =
_
_
_

fF

f
f(1, H
i,t
)
_
_
_

P(n
i,t
= 0|H
i,t
) = 1 P(n
i,t
= 1|H
i,t
) (6)
Here is the sigmoid function, dened as:
(x) =
e
x
1+e
x
. We have stated equation (6) as a special
case applicable to our particular setting. In the remaining
arguments we will use the form given by equations (4) and
(5) to ensure generality.
4.4 Objective Function
The ultimate goal of this project is to develop a probab-
ility distribution

P(n
i,t
|H
i,t
) that will accurately predict
the notes n
i,t
in polyphonic music. There exist a number
of different measures that could indicate the quality of pre-
diction. We choose one of the simplest log-likelihood
of the training data. Given a training set T consisting of
T simultaneities with 12 notes each, the log-likelihood is
simply the average logarithm of the probability of produ-
cing every note in T :
L

P
=
1
12T
log
T

t=1
11

i=0

P(n
i,t
|H
i,t
) (7)
=

P(H)

P(n|H) log

P(n|H)
In the second step in equation (7) we re-expressed
the log-likelihood in terms of the expected cross-entropy
between the target distribution

P(n|H) and the estimate

P(n|H) produced by our eld. The target empirical dis-

tribution

P(n|H) can be computed directly fromthe train-
ing set T , it is just the relative frequency of observing a
note n together with the history H across all the positions
(i, t) in the eld:

P(n|H) =
1
12T
T

t=1
11

i=0
(n, n
i,t
)(H, H
i,t
) (8)
Here refers to the Kronecker delta function, dened
as:
(x, y) =
_
1 if x = y
0 otherwise
(9)
Returning our attention to equation (7), we stress that
the expectation

H
[. . .] is performed over all possible
values that a history H of a note might take. This set
is exponentially large, and a direct computation would be
infeasible. However, for computation we always use the
rst part (top) of equation (7), whereas the second part
(bottom) comes very handy in the algebraic derivations of
the eld induction algorithm.
4.5 Maximum Entropy
To summarize, in the previous two subsections we re-
stricted ourselves to the exponential (Gibbs) form of the
probability distribution

P(n|H), and declared that our
objective is to maximize the likelihood of the training
data within that parametric form. It is important to
note that there is a different statement of objectives that
provably leads to exactly the same exponential solution

P(n|H). Rather than focus on maximum-likelihood, we

could search for the most uniform distribution

P(n|H)
that is consistent with the structure imposed by F. To cla-
rify what we mean by the structure consistency, suppose
f F is a feature of the eld. Let

E[f] denote the em-
pirical or target expected value of f, which is simply how
often the feature actually occurs in the training data T :

E[f] =

P(H)

P(n|H)f(n, H) (10)
=
1
12T
T

t=1
11

i=0
f(n
i,t
, H
i,t
)
Similarly, our estimate

P(n|H) gives rise to the pre-
dicted expectation

E[f] for the function f. Predicted ex-
pected value is simply how often our model thinks that
f should occur in the training set:
210

E[f] =

P(H)

P(n|H)f(n, H) (11)
=
1
12T
T

t=1
11

i=0

P(n|H
i,t
)f(n, H
i,t
)
The key difference between

E[f] and

E[f] is that we
do not look at the actual value n
i,t
when we compute

E[f], instead we predict it from our model

P(n|H).
Given the two expectations in equations (10) and (11) it
is natural to strive that they be equal, that is wed like to
arrange our model in such a way that predicted frequency

E[f] of any feature f matches its actual frequency of oc-

currence

E[f]. Furthermore, if there are multiple distribu-
tions

P(n|H) that honor the constraint that

E[f] =

E[f],
the maximum-entropy principle would guide us to pick
the distribution that makes the least amount of assump-
tions about the data, or equivalently, maximizes its own
expected entropy:
Ent

P
=

P(H)

P(n|H) log

P(n|H) (12)
Curiously, maximizing the entropy subject to the con-
straint that

E[f] =

E[f] for every feature f turns out
to be equivalent to assuming an exponential form for our
probability distribution

P(n|H) and maximizing the like-
lihood given by equation (7).
4.6 Feature Induction
In the previous sections we outlined the general structure
of a random eld over polyphonic music and stated our
objective: to learn the probability distribution

P(n|H)
that maximizes the likelihood of the training data (equa-
tion (7)). Recall that we selected the exponential form for

P(n|H). If we examine equation (4) we note that there

are two things the model depends on. The rst and the
most important in our opinion is the structure of the eld
F, represented as a set of constraints or feature functions
fF. These constraints represent most signicant de-
pendencies between the variables of the eld. The second
thing we learn is the set of weights = {
f
}, one for
each feature fF. We know that and F are intimately
intertwined and we need to learn them simultaneously,
but for the sake of clarity we split the discussion in two
sections. This section will describe how we can incre-
mentally induce the structure F of the eld, starting with
a very at, meaningless structure (primitive, atomic fea-
tures) generalizing to more interesting complex relation-
ships.
The eld induction procedure closely follows the al-
gorithmdescribed in Della Pietra et al. (1997), the primary
difference being that we are dealing with a conditional
eld, whereas Della Pietra uses a joint model. We start
with a eld that contains only individual notes, without
any dependencies: F
0
= {n
i,t
: i = 0. . .11}. We will
incrementally update the structure F by adding the fea-
tures g that result in the greatest improvement in the ob-
jective function. Suppose F
k
= {f
S
} is the current eld
structure. Also assume that the corresponding weights
k
are optimized with respect to F
k
. We would like to add to
F
k
a new feature g that will allow us to further increase
the likelihood of the training data. In order to do that we
rst need to form a set of candidate features G that could
be added. We dene G to be the set of all one-note exten-
sions of the current structure F:
G =
_
f
S
n
j,s
: f
S
F and there exists
n
j

,s
S such that |s s

| 2
_
(13)
In other words, we form new candidate features g tak-
ing an existing feature f and attaching a single note n
j,s
that is not too far from f in time (in our case, not more
than by two simultaneities). Naturally, we do not include
as candidates any features that are already members of F.
Now, following the reasoning of Della Pietra, we would
like to pick a candidate gG that will result in the max-
imum improvement in the objective function. Suppose
that previous log-likelihood based only on F
k
was L

P
.
Now, if we add a feature g weighted by the multiplier ,
the new likelihood of the training data would be:
L

P+{g}
= L

P
+

E[g] log

E[e
g
] (14)
When we add a new feature g to the eld, we would
like to add it with a reasonable weight , preferably
the weight that maximizes the contribution of . We
can achieve that by differentiating the new log-likelihood
L

P+{g}
with respect to and nding the root of the de-
rivative:
0 =
L

P+{g}

= log
_

E[g](1

E[g])

E[g](1

E[g])
_
(15)
An important observation to make is that we arrived
at a closed-form solution for the optimal weight to be
assigned to the new feature g. The closed-form solution is
a special property of binary feature functions, and greatly
simplies the process of inducing eld structure. Know-
ing the optimal value of in closed formallows us to com-
pute the resulting improvement, or gain, in log-likelihood,
also in closed form:
Gain =

E[g] log

E[g]

E[g]
+ (1

E[g]) log
1

E[g]
1

E[g]
(16)
The nal form is particularly interesting, since it
represents the Kullback-Leibler divergence between two
Bernoulli distributions with expected values

E[g] and

E[g] respectively.
4.7 Parameter Estimation
In the previous section we described how we can auto-
matically induce the structure of a random eld by in-
crementally adding the most promising candidate feature
g G. We also presented the closed form equations that
allow us to determine the improvement in log-likelihood
that would result from adding g to the eld, and the
optimal weight that would lead to that improvement.
What we did not discuss is the effect of adding g on the
211
weights of other features already in the eld. Since the
features f F are not independent of each other, adding
a new feature will affect the balance of existing features.
From equation (16) we know that the new log-likelihood
L

P+{g}
is always going to be better than the old one L

P
(unless the eld is saturated and cannot be improved any-
more). However, this does not guarantee that the current
set of weights is optimal for the new structure. We may
be able to further improve the objective by re-optimizing
the weights for all functions that are now in the eld.
Assume now that the structure F contains all the de-
sired features. We would like to adjust the set of weights
, so that the objective function L

P
is maximized. This
is accomplished by computing the partial derivatives of
L

P
with respect to each weight
f
, with the intention of
driving these derivatives to zero:
L

=

E[f

]

E[f

] (17)
Unfortunately, there is no closed-form solution that
would allow us to set the weights to their optimal values.
Instead, we utilize an iterative procedure that will drive
the weights towards the optimum. There are a number
of algorithms for adjusting the weights in an exponential
model, the most widely known being the Generalized It-
erative Scaling (GIS) algorithm proposed by Darroch and
Ratcliff (1972). However, iterative scaling is extremely
slow; much faster convergence can be achieved by using
variations of gradient descent. Given the current value of
the weight vector , we will update it by a small step in the
direction of the steepest increase of the likelihood, given
by the vector of partial derivatives:

k+1
f

k
f
+
L

f
=
k
f
+
_

E[f

]

E[f

]
_
(18)
Equation (18) will be applied iteratively, for all f F,
until the change in likelihood is smaller than some pre-
selected threshold. Note that while

E[f] is computed only
once for each feature f, we will have to re-compute the
value

E[f] after every update. This makes the learning
procedure quite expensive. However, the learning proced-
ure is guaranteed to converge to the global optimum. Con-
vergence is ensured by the fact that the objective function
L

P
is -convex with respect to the weights
f
. One may
verify this by computing the second-order derivative of
L

P
and observing that it is everywhere negative.
4.8 Field Induction Algorithm
We are nally ready to bring together the results of the
previous subsections into one algorithm for automatic in-
duction of random elds models for polyphonic music:
1. Initialization
(a) Let the feature set F
0
be the set of single-note
features: F
0
= {n
i,t
: i = 0 . . . 11}
(b) Set the initial features weights
f
= 1 for all
f F
0
2. Weight Update
(a) Set
k+1
f

k
f
+
_

E[f]

E[f]
_
for each fea-
ture f F
(b) If there is noticeable change in likelihood, repeat
step (2a)
3. Feature Induction
(a) Enumerate the set of candidate features
(b) For every candidate feature g G compute the
optimal weight
g
= log
_

E[g](1

E[g])

E[g](1

E[g])
_
(c) For every g G compute expected improvement
(gain) from adding g to the structure F
(d) Pick the candidate g that promises the highest im-
provement, add it to the structure F, and set
g
=
g
(e) If there is noticeable change in likelihood, go to
step (2), otherwise return F and as the induced
eld
4.9 Discussion on Markov Chains
At this point we should note that our random eld ap-
proach in some sense encompasses the Markov chain ap-
proach. Instead of inducing the features of the eld,
one could easily preselect as features all possible one-
dimensional chains under a certain xed length and then
learn the weights of those features directly. One advant-
age to our current approach, however, is that by selectively
adding only the best features to the model, not only is the
nal number of parameters much smaller, but the features
themselves grow in the direction of the data. In our ex-
periments, for some songs we learned features covering
6 simultaneities (onsets) within the rst thousand features
induced. A Markov chain would require 12
6
2.99 mil-
lion parameters to cover this same onset range.
4.10 Music Retrieval using Random Fields
Now that we have a framework for creating a Markov
Random Field model of a piece of music we wish to
use this model for retrieval. We do this by estimating
a model from a query and then observing how well that
query model predicts the notes in each document in the
collection. In other words, our measure of similarity is
the expected cross-entropy between the empirical distribu-
tion

P
D
(n|H) of the document and the estimate

P
Q
(n|H)
produced by the query model. This measure is essentially
equation (7) from above, with the document rather than
the query as the target distribution.
If the model estimated fromthe query does well at pre-
dicting the notes in a piece of music, then the query and
that piece could have been drawn from the same underly-
ing distribution. Therefore we regard them as similar.
This process is repeated for all pieces in the collection,
and pieces are then ranked in order of increasing cross-
entropy (dissimilarity).
212
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
P
r
e
c
i
s
i
o
n
Recall
11-pt Recall/Precision graph
MRF
MC=2
MC=3
MC=4

MRF

MC=2

MC=3

MC=4
%Change %Change %Change
Retrieved: 151000 151000 151000 151000
Relevant: 8801 8801 8801 8801
Rel|ret: 7291 4508 -38.17* 6805 -6.67* 6037 -17.20*
Interpolated Recall - Precision
at 0.00 1.0000 1.0000 0.0 1.0000 0.0 1.0000 0.0
at 0.10 0.5316 0.3171 -40.3* 0.2176 -59.1* 0.1319 -75.2*
at 0.20 0.3802 0.2388 -37.2* 0.1522 -60.0* 0.0970 -74.5*
at 0.30 0.2787 0.1404 -49.6* 0.1316 -52.8* 0.0931 -66.6*
at 0.40 0.2063 0.0848 -58.9* 0.0873 -57.7* 0.0691 -66.5*
at 0.50 0.1598 0.0492 -69.2* 0.0754 -52.8* 0.0550 -65.6*
at 0.60 0.1209 0.0307 -74.6* 0.0656 -45.7* 0.0498 -58.8*
at 0.70 0.0880 0.0134 -84.8* 0.0530 -39.8* 0.0406 -53.9*
at 0.80 0.0571 0.0021 -96.2* 0.0407 -28.7* 0.0404 -29.2*
at 0.90 0.0218 0.0000 -100.0* 0.0116 -46.8* 0.0372 +70.5*
at 1.00 0.0028 0.0000 -100.0* 0.0002 -91.2* 0.0197 +603.4*
Average precision (non-interpolated) over all rel docs
0.2175 0.1245 -42.76* 0.1021 -53.05* 0.0790 -63.68*
Figure 6: Recall-Precision results for Markov Random Field versus Markov Chain models
5 EXPERIMENTS AND ANALYSIS
For evaluation, we have assembled four collections. The
rst is a set of approximately 3000 polyphonic music
pieces from the CCARH at Stanford. These are mostly
baroque and classical pieces from Bach, Beethoven, Co-
relli, Handel, Haydn, Mozart and Vivaldi. Longer scores
have sometimes been broken up into their various move-
ments, but otherwise each piece is unique. Our remaining
three sets of music are pieces which were intentionally
composed as variations on some theme:
Twinkle 26 individual variations on the tune known to
English speakers as Twinkle, twinkle, little star (in
fact a mixture of mostly polyphonic and a few mono-
phonic versions);
Lachrimae 75 versions of John Dowlands Lachrimae
Pavan from different 16th and 17th-century sources,
sometimes varying in quality (numbers of wrong
notes, omissions and other inaccuracies), and in scor-
ing (for solo lute, keyboard or ve-part instrumental
ensemble);
Folia 50 variations by four different composers on the
well-known baroque tune Les Folies dEspagne.
For retrieval, we select a piece from the three sets of
variations and use that as the query. All other pieces from
that same variation set are judged relevant to the query,
and the rest of the collection is judged non-relevant. This
process is repeated for all pieces in all three sets of vari-
ations, for a total of 151 queries.
Let us dene
MC
as the retrieval system based on
Markov Chain models, and
MRF
as the retrieval system
based on Markov Random Field models. Figure 6 shows
the recall-precision results for the ranked lists produced by
each of the various modeling approaches,
MRF
as well
as
MC
with the length of the chain set to 2, 3, and 4 (1
st
-
, 2
nd
-, and 3
rd
-order models, respectively). In the table,

MRF
is shown rst, and each of the
MC
systems are
shown in comparison, with percentage change (whether
positive or negative) and an asterisk to indicate statistical
signicance (t-test at a 0.05 level).
No matter the chain length, the random eld approach
outperforms the Markov chain approach on just about
every level of precision-recall. On average, the
MC
sys-
tems are from 42% to 63% worse. These results show the
value of the random eld approach.
We believe what is happening is that the random elds
are less sensitive to the noise that appears with vari-
213
ations. For example, suppose there is an insertion of a
single note in one of the variations. The Markov chain
approach counts all possible paths through a polyphonic
sequence. The number of these paths is exponential in the
length of the chain. A single note insertion therefore dis-
proportionately increases the number (and character) of
paths being counted. This analysis is borne out by the
fact higher-order Markov chains actually do progressively
worse (see Figure 6). The longer the chain, the more a
single insertion affects the model. Two note insertions
create even more paths.
On the other hand, random elds take into account the
dependencies between the features. They work by calcu-
lating feature expectations over the data (see section 4.5),
and adjusting the weights so that the contribution of
each feature toward the prediction of a note label (on
or off) is balanced. A single note insertion may activ-
ate a few additional feature functions, but the contribution
of these additional features do not throw off the overall
correctness of the note probability estimate because their
weights are learned initially with non-independence in
mind. Random elds are more robust when it comes to
detecting variations.
6 CONCLUSION
In this paper we developed a retrieval system based on
automatically-induced random elds and show the superi-
ority of these models over Markov chains. Central to our
approach is the notion of a binary feature function, a con-
junction of notes positioned at xed pitches and locations.
Yet these features are not the only ones possible. Re-
call from section 4.2 that features are just functions that
return a boolean value. What happens inside the function
is as limitless as one need imagine. For example, one can
create models of rhythmic patterns by choosing onset con-
junction features of the sort: f
1
= was there an onset 100
ticks prior to the current onset, and another onset 300 ticks
prior to the current onset? (We are currently developing
such models.)
One is not limited to rhythm alone, any more than
one is limited to pitch alone. Alongside pitch-only and
rhythm-only features, our set of active feature functions F
can contain features that are a mixture of both pitch and
duration. E.g.: Let f
3
= the previous note was an E and it
lasted for 200 ms. Not only does this feature contain both
pitch and duration information, but if the model already
contains the pitch-only feature f
2
= the previous note was
an E, we may add f
3
without worry. The training that oc-
curs as part of weight updating (section 4.7) insures that
the values given to all feature are balanced, to automat-
ically take into account statistical dependencies between
features. (It should be clear that f
2
and f
3
are not in-
dependent.) Early work by Doraisamy and R uger (2001)
and Lemstr om et al. (1998) experimented with features of
this nature, but were forced by lack of framework to make
the independence assumption. With random elds, this
assumption is no longer necessary.
Finally, it goes without saying that in addition to pitch
and rhythm, as long as one can craft a binary wrapper (fea-
ture function) one can use any other type of musical in-
formation one has available, including metadata and data
obtained from semantic analysis of audio. Features could
include functions such as: f
4
= my timbre classier gives
me condence > 0.9 that there is a trombone playing at
this point in this audio recording, and my beat tracker es-
timates the current tempo for this song at between 80 and
100 bpm, and there was an onset around 300 ticks ago, and
the metadata tells me this song was recorded in the 1970s
This feature function may (when properly weighted and
combined with the rest of f F) be a strong positive
indication that the note C# is on.
Random elds are a framework for sequential mu-
sic modeling in which combination of multiple, non-
independent sources and types of data may be explored.
Markov random elds are more robust than Markov
chains. They are accurate, without overtting, as we can
see from the retrieval results above. They also offer a
method for attaching relative importance to various fea-
tures without having to make independence assumptions
between the features used. In short, they are an important
framework for developing the kinds of models needed for
music information retrieval applications.
7 ACKNOWLEDGEMENTS
Many thanks to Victor Lavrenko, whose early collabora-
tion made retrieval systems of this sort possible. We also
wish to thank the conference reviewers for their invaluable
comments.
REFERENCES
A. L. Berger, S. A. Della Pietra, and V. J. Della Pietra.
A maximum entropy approach to natural language
processing. Computational Linguistics, 22(1):3971,
1996.
J. Darroch and D. Ratcliff. Generalized iterative scal-
ing for log-linear models. In Ann. Math. Statistics, 43,
pages 14701480, 1972.
S. Della Pietra, V. Della Pietra, and J. Lafferty. Inducing
features of random elds. In IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 19, pages 380
393, 1997.
S. Doraisamy and S. M. R uger. An approach toward a
polyphonic music retrieval system. In J. S. Downie and
D. Bainbridge, editors, Proceedings of the 2nd Annual
ISMIR, pages 187193, Indiana University, Blooming-
ton, Indiana, October 2001.
J. Lafferty, A. McCallum, and F. Pereira. Conditional ran-
dom elds: Probabilistic models for segmenting and
labeling sequence data. In Proc. 18th International
Conf. on Machine Learning, pages 282289. Morgan
Kaufmann, San Francisco, CA, 2001.
K. Lemstr om, A. Haapaniemi, and E. Ukkonen. Retriev-
ing music - to index or not to index. In The Sixth ACM
International Multimedia Conference (MM 98), pages
6465, Bristol, UK, September 13-16 1998.
C. Zhai and J. Lafferty. Dual role of smoothing
in the language modeling approach. URL cite-
seer.ist.psu.edu/zhai01dual.html.
214