Roffo, G., Melzi, S., Castellani, U. and Vinciarelli, A. (2017) Infinite Latent
Feature Selection: A Probabilistic Latent Graph-Based Ranking Approach.
In: IEEE International Conference on Computer Vision (ICCV 2017),
Venice, Italy, 22-29 Oct 2017, pp. 1407-
1415.(doi:10.1109/ICCV.2017.156)
This is the author’s final accepted version.
There may be differences between this version and the published version.
You are advised to consult the publisher’s version if you wish to cite from
it.
http://eprints.gla.ac.uk/149366/
Deposited on: 06 October 2017
Enlighten – Research publications by members of the University of Glasgow
http://eprints.gla.ac.uk
Infinite Latent Feature Selection:
A Probabilistic Latent Graph-Based Ranking Approach
Giorgio Roffo
University of Glasgow
Giorgio.Roffo@Glasgow.ac.uk
Simone Melzi
University of Verona
Simone.Melzi@univr.it
Umberto Castellani
University of Verona
Umberto.Castellani@univr.it
Alessandro Vinciarelli
University of Glasgow
Alessandro.Vinciarelli@Glasgow.ac.uk
Abstract
Feature selection is playing an increasingly significant
role with respect to many computer vision applications
spanning from object recognition to visual object tracking.
However, most of the recent solutions in feature selection
are not robust across different and heterogeneous set of
data. In this paper, we address this issue proposing a ro-
bust probabilistic latent graph-based feature selection al-
gorithm that performs the ranking step while considering
all the possible subsets of features, as paths on a graph,
bypassing the combinatorial problem analytically. An ap-
pealing characteristic of the approach is that it aims to dis-
cover an abstraction behind low-level sensory data, that is,
relevancy. Relevancy is modelled as a latent variable in a
PLSA-inspired generative process that allows the investiga-
tion of the importance of a feature when injected into an
arbitrary set of cues. The proposed method has been tested
on ten diverse benchmarks, and compared against eleven
state of the art feature selection methods. Results show that
the proposed approach attains the highest performance lev-
els across many different scenarios and difficulties, thereby
confirming its strong robustness while setting a new state of
the art in feature selection domain.
1. Introduction
Performance of machine learning methods is heavily de-
pendent on the choice of features on which they are ap-
plied. Different features can entangle and hide the differ-
ent explanatory factors of variation behind the data. Fea-
ture Selection (FS) aims at improving the performance of a
prediction system, allowing faster and more cost-effective
models, while providing a better understanding of the in-
herent regularities in data. In the recent computer vision
literature there are many scenarios where FS is a crucial op-
eration [5, 30, 10, 13, 24, 28]. From multiview face recog-
nition [13] where FS is used to speed up the multiview face
recognition process and to maintain the generalization per-
formance, to object recognition [30], until real-time visual
object tracking [28, 25] where FS dynamically identifies
discriminative features that help in handling the appearance
variability of the target by improving tracking performance.
In this paper, we propose a probabilistic latent graph-
based feature selection algorithm that performs the ranking
step by considering all the possible subsets of features ex-
ploiting the convergence properties of power series of ma-
trices. We map the feature selection problem to an affinity
graph (e.g., feature ≈ node), and then we consider a subset
of features as a path connecting set of nodes. An appeal-
ing characteristic of the approach is that the importance of
a given feature is modelled as a conditional probability of a
latent variable and features, namely P (z|f ). Our approach
aims to model an important hidden variable behind data,
that is, relevancy in features. Raw values are observable
while relevancy to a particular task is not (e.g., in classifica-
tion), therefore, relevancy is modelled as an abstract latent
variable. In particular, our approach consists of three main
parts:
• Pre-processing: a quantization process is applied on
raw feature distributions ~x
i
, mapping their values to
a countable nominal smaller set of tokens. The pre-
processing step assigns a descriptor f
i
to each raw fea-
ture ~x
i
.
• Graph-Weighting: we build an undirected fully-
connected graph, where nodes correspond, one by one,
to each feature f
i
, and each weighted edge among
f
i
f
j
models the probability that features x
i
and x
j
are relevant. Weights are learnt automati-
cally by a learning framework based on a variation of
the probabilistic latent semantic analysis (PLSA) tech-
nique [21], which models the probability of each co-
1
occurrence in f
i
, f
j
as a mixture of conditionally in-
dependent multinomial distributions. Parameters are
estimated using the Expectation Maximization (EM)
algorithm.
• Ranking: the ranking step is done following the idea
of the Infinite Feature Selection (Inf-FS) [30], that con-
siders all the possible paths among nodes investigating
the redundancy of any features when injected into ar-
bitrary sets of cues.
The proposed method is compared against 11 state of
the art feature selection methods selected from recent lit-
erature in the machine learning and pattern recognition do-
mains, reporting results for a total of 576 unique tests (note,
the source code is available at Matlab-Central). We se-
lected 10 publicly available benchmarks of cancer classifi-
cation and prediction on DNA microarray data (Colon [32],
Lymphoma [14], Leukemia [14], Lung [15], Prostate [1]),
handwritten character recognition (GINA [2]), text classi-
fication from the NIPS feature selection challenge (DEX-
TER [18]), and a movie reviews corpus for sentiment
analysis (POLARITY [26]). More extensively, two object
recognition datasets have been taken into account (PAS-
CAL VOC 2007-2012 [11, 12]). Results show that the pro-
posed approach represents the most robust algorithm, which
achieves the highest level of performance across many dif-
ferent domains and challenging scenarios.
The rest of the paper is organized as follows: Sec. 2 illus-
trates the related literature, mostly focusing on the compara-
tive approaches we consider in this study. Sec. 3 details the
proposed approach, also giving a formal justification and
interpretation based on absorbing Markov chain (Sec. 3.4).
Extensive experiments are reported in Sec. 4, and, finally,
in Sec. 5, conclusions are given, and future perspectives are
envisaged.
2. Related Work
Since the mid-1990s, few domains used more than 20
features. The situation has changed considerably in the past
few years and most papers explore domains with hundreds
to tens of thousands of features. New approaches were pro-
posed to address these challenging tasks involving many ir-
relevant and redundant variables and often comparably few
training examples. Typically, FS techniques are partitioned
into three classes [19]: Filters, Wrappers and Embedded
methods. The proposed approach is a filter method, which
analyzes intrinsic properties of data, ignoring the type of
classifier. Conversely, wrappers use classifiers to score a
given subset of features, and embedded methods inject the
selection process directly into the learning process of the
classification framework.
Among the most used filter-based strategies, Relief-
F [23] is an iterative, randomized, and supervised approach
that estimates the quality of the features according to how
well their values differentiate data samples that are near to
each other. Another effective yet fast filter method is the
Fisher method [17], which computes a score for a feature
as the ratio of inter-class separation and intra-class vari-
ance, where features are evaluated independently. A Mu-
tual Information based approach (MI) is proposed in [35].
MI considers as a selection criterion the mutual informa-
tion between the distribution of the values of a given fea-
ture and the membership to a particular class. Even in the
last case, features are evaluated independently, and the final
feature selection occurs by aggregating the m top ranked
ones. In unsupervised learning scenarios, a widely used
method is the Laplacian Score (LS) [20], where the im-
portance of a feature is evaluated by its power of locality
preserving. In order to model the local geometric struc-
ture, this method constructs a nearest neighbor graph. LS
algorithm seeks those features that respect this graph struc-
ture. The unsupervised feature selection for multi-cluster
data is denoted MCFS in [8], which selects those features
such that the multi-cluster structure of the data can be best
preserved. [34] proposed a L2,1-norm regularized discrim-
inative feature selection for unsupervised learning (UDFS)
which selects the most discriminative feature subset from
the whole feature set in batch mode. Feature selection
and kernel learning for local learning-based clustering (LL-
CFS) [36] associates a weight to each feature and incorpo-
rates it into the built-in regularization of the LLC algorithm
to take into account the relevance of each feature for the
clustering. In the experiments, we also compare our ap-
proach against the unsupervised graph-based filter method
dubbed Inf-FS [30]. In the Inf-FS formulation, each feature
is a node in the graph, a path is a selection of features, and
the higher the centrality score, the most important (or most
different) the feature. Another widely used FS method is
SVM-RFE (RFE) [19], which is a wrapper method that se-
lects features in a sequential, backward elimination manner,
ranking high a feature if it strongly separates the samples by
means of a linear SVM. Finally, for the embedded methods,
the feature selection via concave minimization (FSV) [7]
is a popular FS strategy, where the selection process is in-
jected into the training of an SVM by a linear programming
technique. For further information, please see Tab. 2.
3. Our Approach
Given a training set X represented as a set of feature dis-
tributions X = {~x
1
, ..., ~x
n
}, where each m × 1 vector ~x
i
is the distribution of the values assumed by the i
th
feature
with regards to the m samples, we build an undirected graph
G, where nodes correspond to features and edges model re-
lationships among pairs of nodes. Let the adjacency matrix
A associated to G defining the nature of the weighted edges:
each element a
ij
of A, 1 ≤ i, j ≤ n, models pairwise re-
lationships between the features. Each weight represents
the likelihood that features ~x
i
and ~x
j
are good candidates.
Weights can be associated to a binary function of the graph
nodes:
a
ij
= ϕ(~x
i
, ~x
j
), (1)
where ϕ(·, ·) is a real-valued potential function learned
by the proposed approach in a PLSA-inspired framework.
The learning framework models the probability of each co-
occurrence in ~x
i
, ~x
j
as a mixture of conditionally indepen-
dent multinomial distributions, where parameters are learnt
using the EM algorithm. Given the weighted graph G, the
proposed approach analyses subsets of features as paths
connecting them. The cost of each path is given by the joint
probability of all the nodes belonging to it. The method ex-
ploits the convergence property of the power series of ma-
trices as in [30], and evaluates in an elegant fashion the rele-
vance of each feature with respect to all the other ones taken
together. For this reason, we dub our approach infinite la-
tent feature selection (ILFS).
3.1. Discriminative Quantization process
Since the amount of possible distinct values in ~x
i
is huge,
we map this large set of values to a countable smaller set,
hereinafter referred to as set of tokens. Tokens are the words
of our dictionary of features. Thus, each feature will be
represented by a new low-dimensional vocabulary of mean-
ingful tokens. The way used to assign each value to a spe-
cific token is based on a quantization process, we called dis-
criminative quantization (DQ). The rationale behind the DQ
process is to take into account how well a given feature is
representative of a class before performing the many-to-few
mapping.
Firstly, the Fisher criterion is used to compute a scoring
vector Φ = [·, ..., ·] which takes into account both means
and standard deviations of the classes, for each sample and
feature. In binary classification scenarios, this is given by
Φ =
1
Z
h
(s − µ
1
)
2
σ
2
1
+ σ
2
2
,
(s − µ
2
)
2
σ
2
1
+ σ
2
2
i
, (2)
where s is a sample from the i
th
feature ~x
i
, µ
k
and σ
k
denote the mean and standard deviation of class k, respec-
tively. A normalization factor Z is introduced to ensure that
the scores are a valid distribution over both classes. A nat-
ural generalization of these scores into a multi-class frame-
work is given by
Φ =
1
Z
h
(s − µ
1
)
2
P
K
k=1
σ
2
k
, ...,
(s − µ
K
)
2
P
K
k=1
σ
2
k
i
, ∀
k∈K
(3)
where K is the number of classes, s is a single sample from
the i
th
feature. Therefore, considering all the samples, Φ
results to be a m × K matrix.
Now, let us assume that the sample s belongs to class k.
If ~x
i
is a strong discriminant feature, s will score high at
f
z t
n
Features Latent Variables
Tokens
t
1
t
2
t
6
z
1
𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑐𝑦
z
2
𝐼𝑟𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑐𝑦
f
1
f
2
f
n
(a)
(b)
Figure 1. Illustration of the general structure of the model. (a) The
intermediate layer of latent topics that links the features and the
tokens. (b) The graphical model using plate representation.
Φ
k
. Then, we derive our priors π by extracting Φ scores for
each feature according to the ground truth as follows:
π = diag(ΦY )
where Y is the 1-of-K representation of the ground truth. It
is a particularly convenient representation where the class
labels are represented by K-dimensional vectors in which
one of the elements equals 1, and all remaining elements
equal 0. As a result, π ∈ [0, 1] is a 1 × m vector containing
a score for each element of a particular feature i. It takes
into account how well each element is represented by the
feature i according to Eq.3.
Finally, quantization is performed. The first step is to di-
vide the entire range of values [0, 1] into a series of T inter-
vals (i.e., we use T = 6 in this work: interval 1 corresponds
to not-well-represented samples, and interval 6 is associated
to well-represented samples). Secondly, we assign a token
to values falling into each interval. Given the outcomes of
the DQ process, we obtain a meaningful new representa-
tion of our training data X in the form of F = {f
1
, ..., f
n
},
where each feature is described by a vocabulary of few to-
kens. In other words, the derived feature representation f
i
comes from x
i
where each value is assigned to a token T .
According to this formulation, a strong discriminative fea-
ture will be intuitively associated to a descriptor f
i
contain-
ing many relatively large tokens (e.g., 5, 6) rather than small
ones (e.g., 1, 2).
3.2. From co-occurrences to graph weighting
Weighting the graph according to the nodes discrimina-
tory power has a great influence on the quality of the rank-
ing process. We designed a framework to automatically per-
form the graph weighting from training data, such that the
learnt parameters can be used to sort features according to
their degrees of relevance or importance.
Our solution is based on a variation of the PLSA [21]
technique, that considers co-occurrences of tokens and fea-
tures, ht, fi, to model the probability of each co-occurrence
as a mixture of conditionally independent multinomial dis-
tributions.
In order to better understand the intuition behind the pro-
posed model, we need to make some assumptions. We as-
sume that a feature consists of only two topics represent-
ing the two main latent variables of any feature selection
algorithms: Relevancy and Irrelevancy. Therefore, we in-
troduce an unobserved class variable Z = {z
1
, z
2
} obtain-
ing a latent variable model for co-occurrence tokens. As
a result, there is a distribution P (z|f) over the fixed num-
ber of topics for each feature f. Similarly, original PLSA
model does not have the explicit specification of this dis-
tribution but it is indeed a multinomial distribution where
P (z|f) represents the probability that topic z appears in fea-
ture f . Fig. 1.(a) shows the general structure of the model,
each feature can be represented as a mixture of concepts
(Relevant/Irrelevant) weighted by the probability P (z|f )
and each token expresses a topic with probability P (t|z).
Fig. 1.(b) describes the generative process for each of the
n features in the set by using plate representation. We can
write the probability a token t appearing in feature f as fol-
lows:
P (t|f ) = P (t|z
1
)P (z
1
|f) + P (t|z
2
)P (z
2
|f).
By replacing this for any feature in the set F we obtain,
P (f ) =
Y
t
n
P (t|z
1
)P (z
1
|f) + P (t|z
2
)P (z
2
|f)
o
.
The unknown parameters of this model are P (t|z) and
P (z|f). As for PLSA, we derived the equation for com-
puting these parameters by maximum likelihood. The log-
likelihood function is given by
L =
X
f
X
t
Q(f, t) log[P (t|f )]
where Q(f, t) is the number of times token t appearing in
feature f . The EM algorithm is used to compute optimal
parameters. The E-step is given by
P (z|f, t) =
P (z)P (f |z)P (t|z)
P (z
1
)P (f |z
1
)P (t|z
1
) + P (z
2
)P (f |z
2
)P (t|z
2
)
,
and the M-step is given by
P (t|z) =
P
f
Q(f, t)P (z|f, t)
P
f,t
0
Q(f, t
0
)P (z|f, t
0
)
,
P (f |z) =
P
t
Q(f, t)P (z|f, t)
P
f
0
,t
Q(f
0
, t)P (z|f
0
, t)
,
P (z) =
P
f,t
Q(f, t)P (z|f, t)
P
f,t
Q(f, t)
.
The responsibility for assigning the “condition of be-
ing relevant” to features lies to a great extent with the un-
observed class variable Z. In particular, we initialize the
model priors P (t|z) in order to link z
1
to the abstract topic
of Relevancy, and hence z
2
to Irrelevancy. By construc-
tion we limited the range of the tokens to values between
1 and 6 (see Sec.3.1), with 1 that behaves the same way
as being the lowest rating for a sample of a particular fea-
ture, and 6 being the highest quality. As a result, a natural
way to initialize these priors is to generate a pair of linearly
spaced vectors assigning a higher probability P (t
0
|Z = z
1
)
for those tokens t
0
which score higher, and consequently the
opposite for P (t
0
|Z = z
2
).
Finally, the graph can be weighted by the estimated prob-
ability distribution P (Z = z
1
|f). According to Eq.1, each
element a
i
j of the adjacency matrix is the joint probability
that the abstract topic of relevancy appears in feature f
i
and
f
j
, namely:
a
ij
= ϕ(~x
i
, ~x
j
) = P (Z = z
1
|f
i
)P (Z = z
1
|f
j
), (4)
where mixing weights P (Z = z
1
|f
i
) and P (Z = z
1
|f
j
) are
conditionally independent. Indeed, knowledge of whether
P (Z = z
1
|f
i
) occurs provides no information on the like-
lihood of P (Z = z
1
|f
j
) occurring, and knowledge of
whether P (Z = z
1
|f
j
) occurs provides no information on
the likelihood of P (Z = z
1
|f
i
) occurring.
3.3. Probabilistic Infinite Feature Selection
Let γ = {v
0
= i, v
1
, ..., v
l−1
, v
l
= j} denote a path
of length l between nodes i and j, that is, features ~x
i
and
~x
j
, through other nodes v
1
, ..., v
l−1
. For simplicity, sup-
pose that the length l of the path is lower than the total
number of nodes n in the graph. In this setting, a path is
simply a subset of the available features/nodes that come
into play. Moreover, the network is characterized by walk
structure [6], where nodes and edges can be visited multiple
times.
We can then estimate the joint probability that γ is a good
subset of features as
P
γ
=
l−1
Y
k=0
a
v
k
,v
k+1
. (5)
Let us define the set P
l
i,j
as containing all the paths of
length l between i and j; to account for the energy of all the
paths of length l, we sum them as follows:
C
l
(i, j) =
X
γ∈P
l
i,j
P
γ
, (6)
which, following standard matrix algebra, gives:
C
l
(i, j) = A
l
(i, j),
![Table 2. Feature selection approaches considered in the experiments [29, 27]. The table reports their Type, class (Cl.), complexity (Compl.), and execution times in seconds (Exec.Time). As for the complexity, T is the number of samples, n is the number of initial features, i is the number of iterations in the case of iterative algorithms, and C is the number of classes.](/figures/table-2-feature-selection-approaches-considered-in-the-2in0z9sm.webp)