Effective Products Categorization with Importance Scores and Morphological Analysis of the Titles

Abstract

During the past few years, the e-commerce platforms and marketplaces have enriched their services with new features to improve their user experience and increase their profitability. Such features include relevant products suggestion, personalized recommendations, query understanding algorithms and numerous others. To effectively implement all these features, a robust products categorization method is required. Due to its importance, the problem of the automatic products classification into a given taxonomy has attracted the attention of multiple researchers. In the current literature, we encounter a broad variety of solutions, ranging from supervised and deep learning algorithms, as well as convolutional and recurrent neural networks. In this paper we introduce a supervised learning method which performs morphological analysis of the product titles by extracting and processing a combination of words and n-grams. In the sequel, each of these tokens receives an importance score according to several criteria which reflect the strength of the correlation of the token with a category. Based on these importance scores, we also propose a dimensionality reduction technique to reduce the size of the feature space without sacrificing much of the performance of the algorithm. The experimental evaluation of our method was conducted by using a real-world dataset, comprised of approximately 320 thousand product titles, which we acquired by crawling a product comparison Web platform. The results of this evaluation indicate that our approach is highly accurate, since it achieves a remarkable classification accuracy of over 95%

Full text

Effective Products Categorization with Importance
Scores and Morphological Analysis of the Titles
Leonidas Akritidis, Athanasios Fevgas, Panayiotis Bozanis
Data Structuring & Engineering Lab
Department of Electrical and Computer Engineering
University of Thessaly
Volos, Greece
Email: {leoakr,fevgas,pbozanis}@e-ce.uth.gr
Abstract—During the past few years, the e-commerce plat-
forms and marketplaces have enriched their services with new
features to improve their user experience and increase their
profitability. Such features include relevant products suggestion,
personalized recommendations, query understanding algorithms
and numerous others. To effectively implement all these features,
a robust products categorization method is required. Due to its
importance, the problem of the automatic products classification
into a given taxonomy has attracted the attention of multiple
researchers. In the current literature, we encounter a broad
variety of solutions, ranging from supervised and deep learning
algorithms, as well as convolutional and recurrent neural net-
works. In this paper we introduce a supervised learning method
which performs morphological analysis of the product titles by
extracting and processing a combination of words and n-grams.
In the sequel, each of these tokens receives an importance score
according to several criteria which reflect the strength of the cor-
relation of the token with a category. Based on these importance
scores, we also propose a dimensionality reduction technique to
reduce the size of the feature space without sacrificing much of
the performance of the algorithm. The experimental evaluation
of our method was conducted by using a real-world dataset,
comprised of approximately 320 thousand product titles, which
we acquired by crawling a product comparison Web platform.
The results of this evaluation indicate that our approach is highly
accurate, since it achieves a remarkable classification accuracy
of over 95%.
I. INTRODUCTION
It is common knowledge that e-commerce is one of the
fastest growing Web-based enterprises. In the last few years
there has been a significant increase in the e-commerce share
of total global retail sales. From 7.4% in 2015, the percentage
of the online sales has increased to 11.9% in 2018, and it is
predicted to reach 17.5% at the end of 20211. Consequently,
the research topics related to e-commerce have been rendered
particularly important. More specifically, the effective and
efficient management, mining, and processing of the products
data are presently of top priority for the leading e-commerce
platforms.
One of the most fundamental problems in this area is the
automatic classification of products into an existing hierarchy
of categories. The successful solution of this problem can lead
to numerous novel applications, including query expansion and
1https://www.statista.com/statistics/534123/e-commerce-share-of-retail-
sales-worldwide/
rewriting, retrieval of relevant products, personalized recom-
mendations, etc. On the other hand, the current large-scale
commercial systems are now offering tens or even hundreds of
millions of products and their warehouses are getting updated
constantly at high rates. Therefore, the approach of manual
categorization is not scalable and automatic classification
emerges even more compelling.
Due to its importance, the problem in question has gained
a lot of research attention recently. The methods which we
encounter in the relevant literature take into consideration the
product titles and/or their textual descriptions to train their
classification models. A number of approaches treat the issue
of product classification as a standard short-text classification
problem and they apply slight modifications of the proposed
algorithms. Other methods perform morphological analysis of
the titles and the textual descriptions by extracting words and
n-grams. Finally, a third family of solutions employ deep
learning approaches, which are based on convolutional or
recurrent neural networks.
A significant portion of these methods take into consid-
eration additional information –known as meta-data– about
a product and extract features from brands, technical spec-
ifications, or textual descriptions. Nevertheless, these types
of metadata are not always present, whereas, even in case
they exist, they are occasionally incomplete or incorrect. In
addition, although many models have been effectively applied
to text classification, they are difficult to be applied to product
categorization due to their lack of scalability, and the sparsity
and skewness of the data contained in the commercial product
catalogs.
In this paper, we present a supervised algorithm to address
this interesting problem. Similarly to some of the existing
methods, our approach is based solely on the titles of the
products. However, in contrast to the existing methods, our
algorithm does not employ standard n-grams of fixed length;
instead, it sets a parameter Nand extracts all the 1,2, . . . , N
-grams from the titles. All generated n-grams are stored within
a dictionary data structure which, for each token, maintains a
series of useful statistics. We show that this strategy leads to
significant performance benefits.
As might be expected, a portion of the generated n-grams
are strongly correlated with a category, whereas others are

not. For instance, the unigram Apple is not a strong indicator
of a single category, since it can be correlated with multiple
potential classes such as Computers,Tablets,Mobile Phones,
and so on. On the other hand, the bigram Apple iPhone
is indeed indicative of the category of a product (Mobile
Phones). Based on this notification, we assign importance
scores to each token. The importance score essentially reflects
the informational value of a token when it is used to determine
the class of a product. During the evaluation of the method, we
experimented with multiple scoring functions and we propose
one which boosts the classification accuracy of our method.
Finally, the extraction of multiple types of n-grams undoubt-
edly leads to an increase of the size of the aforementioned
dictionary data structure. In this paper we propose a strategy
which prunes our dictionary based on the importance scores
of the tokens. That is, each token which receives a low
importance score is considered of no informational value and
it is removed from the dictionary. This method reduces the
dimensionality of the feature space and leads to a much more
compact model, whereas the degradation of the classification
performance is proved infinitesimal.
The rest of the paper is organized as follows: Section II,
contains an overview of the current state-of-the-art methods in
the area. Section III, describes in details the characteristics of
the model, the importance scores and the model training and
testing algorithms. The experimental evaluation of our pro-
posed method is presented in Section IV, whereas Section V
summarizes the conclusions of the paper.
II. RE LATE D WOR K
Due to its importance, the problem of automated product
classification has attracted multiple researchers and is being
studied systematically, particularly during at the last 5 years.
One of the first approaches in this area of research was
GoldenBullet, an early system which applied a combination
of IR and machine learning techniques (such as Naive Bayes
and k-nearest neighbor), to classify the products of an e-
commerce platform into a pre-defined taxonomy structure such
as UNSPSC2.
Supervised classification was also used in [1], where the
authors introduced a multi-class SVM product classifier in
combination with a cost-sensitive function. This classifier was
evaluated by employing a part of the UNSPSC taxonomy.
The first work which attempted to address the problem by
exploiting only the product titles was [2]. This research
introduced a binary classifier in combination with the One-
Against-All (OAA) and Error Correction Tournament (ECT)
strategies. The features for the model were extracted by using
lexical information and topic discovery algorithms, such as
Latent Dirichlet Allocation [3].
In addition, the work of [4] introduced an algorithm which
constructed document bag-of-word vectors from the product
descriptions by using traditional approaches based on tf-
idf and wordVec. In [5], [6] the classification problem is
2United Nations Standard Products and Services Code,
https://www.unspsc.org/
decomposed into a coarse level task and a fine level task.
In the sequel, simple classifier is applied for the coarse level,
and a more sophisticated model is used to separate classes
at the fine level. Then, a graph algorithm is used to discover
groups of highly similar classes. In [7] the authors introduced
an algorithm based on the combinations and permutations of
the titles tokens to achieve unsupervised products clustering.
During the past few years there is an ongoing line of re-
search which attempts to solve the problem by employing deep
learning approaches. More specifically, [8], [9] demonstrated
that the character-level convolutional networks are at least
competitive against the well-established models such as bag of
words, word-based ConvNets and recurrent neural networks.
Furthermore, [10] highlighted the effectiveness of the sim-
ple convolutional neural networks with little hyperparameter
tuning and static vectors. The authors also used task-specific
vectors and reported further gains in performance through
careful fine-tuning.
Moreover, [11] modeled short texts by using semantic
clustering and convolutional neural networks. Particularly, the
meaningful multi-scale semantic units are initially detected
by a clustering algorithm. In the sequel, they are fed into
a convolutional layer, followed by a max-pooling operation.
Two other articles, [12] and [13] focused on large-scale deep
learning text classification. The former introduced a two stage
algorithm which consisted of a search stage and classification
stage. On the other hand, the model which was proposed in
the latter included multiple recurrent neural networks.
Finally, a comparative study which examines the current ma-
chine learning techniques for automatic product categorization
is provided in [14].
III. SUP ERVISED PRO DU CT S CATEGORIZATION
In this section, we present the characteristics of the pro-
posed method for products categorization. Initially, we provide
some preliminary elements and we establish the theoretical
background of the algorithm. In the sequel, we describe the
properties of the model and we discuss the training and testing
procedures.
A. Theoretical Background
Our analysis begins with the introduction of the taxon-
omy Y, that is, the hierarchical structure which stores all
the categories (also called classes, or labels) in the dataset.
This hierarchy is usually represented by a tree data structure
which organizes the elements of Yinto various levels of
subcategories. More specifically, the categories tree contains
two types of nodes: i) the parent nodes, which may include
any number of subcategories; and ii) the leaf nodes, which do
not have any children.
Furthermore, we consider a set X={x1, x2, . . .}, which
includes all the products in the dataset. There exist various
methods to categorize the products of Xinto the classes of
Y. The most common scenario in the current e-commerce
platforms dictates that each product xiis mapped to only one
leaf category yj∈Y. In other words, a product cannot be

assigned multiple categories, and, additionally, it cannot be
mapped to a parent category. The full categorization of a prod-
uct constitutes a path from the root of the aforementioned tree
to the leaf category where the product belongs to. For example,
Computers/Hardware/CPUs may represent a categorization of
three levels for a computer processor.
Each product in the dataset is described by its title τ, which
consists of a sequence of several words (w1, w2, . . . , wlx),
where lxdenotes the length of the title of xin words. In
the real-world commercial environments, a product may also
be accompanied by extended descriptions, or other types of
explicitly stated meta-data such as brand names, manufacturers
etc. However, such information is occasionally not consistent
(i.e., it is not provided for all products), incomplete, inaccurate,
or noisy. Consequently, in this paper we introduce a supervised
algorithm which builds a classification model based on a set
of previously classified products, by taking into consideration
the provided titles only.
B. Model Properties
In this subsection, we describe in details the characteris-
tics of the product categorization model and we present the
feature extraction method, as well as the training and testing
algorithms. Most works in the area construct correlations of
the title words with the categories of the given taxonomy by
applying tf-idf or similar methods. However, a single word
may be correlated with multiple categories (in other words,
it may be ambiguous), or it may incorrectly indicate a strong
relationship with a particular category. As stated earlier, we
anticipate that the word Apple will be proved ambiguous,
since it is strongly correlated with multiple categories such
as Computers,Tablets,Mobile Phones, and others. The same
phenomenon is expected to occur multiple times, especially
for brand names, color and size descriptors, capacities, etc.
For this reason, in our work we also employ n-grams; that
is, sequences of nsuccessive words in the title of a product.
Compared to the single title words, the n-grams have a lower
possibility of being ambiguous. In the previous example, we
explained that the word Apple may be proved problematic
due to its correlation with multiple categories. Nonetheless,
the bigrams Apple iPhone and Apple iPad should be much
more specific. We anticipate that the former will be frequently
correlated with the category Mobile Phones, whereas for the
later, the category which we primarily expect to encounter is
Tablets. In the discussion which follows, we will collectively
refer to both n-grams and words by using the term tokens.
The previous discussion reveals that the generated tokens
may be correlated with one or more categories with a different
manner. That is to say, each token texhibits its own level of
ambiguity with respect to a given category y. Therefore, it
is required that we take into consideration how important a
token is during the classification of a product. We quantify
this importance by assigning an importance score Itto each
token tof a product title according to the following equation:
It=ltlog |Y|
ft
(1)
where ltis the length of the token tin words, and ftis a
frequency value which represents the number of the categories
which have been correlated with t. This definition embodies
the following notations:
•A token which has been correlated with multiple cat-
egories is considered of low informational value and
should receive a low importance score. The opposite
holds for the tokens which frequently appear in combi-
nation with only one or two categories.
•The length of a n-gram (in words) is another parameter
which contributes to its importance. The previous discus-
sion demonstrated that bigrams are more important than
unigrams. More generally, the longer the token, the higher
the importance score should be.
Another point, which distinguishes our work from the cur-
rent approaches, is that here we do not merely employ n-grams
of constant length or of constant number. More specifically,
we define a parameter Nand we construct all the 1,2, . . . , N
-grams. For example, for N= 3, the algorithm will generate
all the unigrams, bigrams and trigrams of the words of a prod-
uct title. As demonstrated later in the experimental evaluation,
this approach leads to significant performance improvements
of the algorithm.
Our proposed model consists of a lexicon data structure
which stores all the distinct tokens of the product titles in
the training set. Moreover, for each stored entry t, the lexicon
maintains:
•its frequency value ft;
•its importance score It; and
•arelevance description vector (RDV), which, for each
token-category relationship, includes a pair in the form
(y, ft,y ), where yis the ID of the category and ft,y is
another frequency value that reflects how many times t
has been correlated with the category y.
This lexicon will assist us in classifying the products of the
test set, based on the tokens of the observed products of the
training set. Since it includes n-grams of multiple types, we
expect that its size will grow quickly, as more samples are
examined. For this reason, in Subsection III-E, we propose a
dimensionality reduction technique for limiting its size.
C. Model Training
In Algorithm 1 we present a pseudocode of the training
algorithm. For each product in the training set, we initially
retrieve its respective category ID y, and its title τ. Before
the extraction of the tokens, each single word in the title is
processed in three ways (step 5): i) casefolding (all capital
letters are converted to lower case); ii) punctuation removal
(that is, removal of several unnecessary symbols such as
commas, exclamation and question marks, ampersands, etc.);
and iii) stemming (we compute and store the ‘root’ of a word
instead of the word itself). Regarding stemming, we were
particularly interested in studying the effects of eliminating
the singular/plural differences in the words of the titles.
The n-grams are typically extracted by employing a window
of length n, which captures sequences of nsuccessive words.

Algorithm 1: Model training algorithm
1initialize the lexicon L;
2for each product xin the training set do
3retrieve the category y;
4extract the title τ;
5perform linguistic processing of τ;
6for each n∈[1, N]do
7compute all tokens Tnof length n;
8for each token t∈Tndo
9if L.search(t) == false then
10 L.insert(t);
11 set ft←1;
12 L.insertRDV(t,y);
13 set ft,y ←1;
14 else
15 if L.searchRDV(t,y) == false then
16 set ft←ft+ 1;
17 L.insertRDV(t,y);
18 set ft,y ←1;
19 else
20 set ft,y ←ft,y + 1;
21 end
22 end
23 end
24 end
25 end
26 for each token t∈Ldo
27 set It←ltlog(|Y|/ft);
28 sort RDV of tin decreasing ft,y order;
29 end
To acquire all the tokens of a product title, this window
progressively slides from the left to the right by taking steps
of one word at a time. In addition, since we are interested in
obtaining all 1,2, . . . , N -grams, it is required that we employ
Nsuch windows, with lengths ranging from 1 to N.
For each token tcaptured by the aforementioned window,
we perform a search in the lexicon L(step 9). In case it is not
present, we insert it and we set its frequency ftequal to 1.
We also initialize its RDV and we insert the pair (y, ft,y )with
ft,y = 1 (steps 9–12). In the opposite case, where talready
resides in the lexicon, we traverse its RDV by looking up for
y. In other words, we initially search if thas been previously
correlated with y. If it has not (i.e. ydoes not exist in the RDV
of t), we initially increase ftby one (step 16). In the sequel,
we insert the the pair (y, ft,y )with ft,y = 1 in the RDV of
t. Finally, in case yhas been previously correlated with t, we
merely increase the corresponding ftyvalue by one.
At this point, all tokens in the titles of the products have
been inserted into the lexicon L. In the last three steps (28–
31), we traverse Land, for each token t∈L, we perform two
operations: i) we compute its importance score by applying
equation 1; and ii) we sort its respective RDV in decreasing
ft,y order. In this way, the strongest (t, y)correlations are
Algorithm 2: Categorization algorithm
1for each product xin the test set do
2initialize candidates list Y0;
3extract the title τ;
4perform linguistic processing of τ;
5for each n∈[1, N]do
6compute all tokens Tnof length n;
7for each token t∈Tndo
8if L.search(t) == true then
9for each pair (y, ft,y)in the RDV of t
do
10 if Y0.search(y) == false then
11 Y0.insert(y);
12 set Sy←f(It,y, Qt,y )
13 else
14 set Sy←Sy+It,y
15 end
16 end
17 end
18 end
19 end
20 sort Y0in decreasing Syorder;
21 set y←Y0[0];
22 end
moved towards the beginning of the list. This arrangement
will assist us in the next phase (Subsection III-E) where we
introduce a pruning strategy for reducing the size of the model.
D. Model Testing
Now we are ready to apply our trained model to categorize
the unlabeled products of the test set. Algorithm 2 shows
the basic steps of the categorization phase. At first, for each
product xin the test set, a list of candidate categories Y0is
initialized. In the sequel, the title τis processed by following
the same three steps of the training phase; i.e., casefolding,
punctuation removal, and stemming.
We continue in a similar spirit to the training algorithm,
and we extract all the tokens from the title of the product
by applying the same sliding windows of length [1, N]. For
each extracted token t, a search in the lexicon Lis performed.
In case it is successful, we retrieve the frequency ft, the
importance score It, and the entries of the respective RDV.
In the sequel, we traverse the RDV of tand for each
category ywe look-up the candidates list Y0. In case the
look-up fails, we insert yin Y0, and we compute a score St,y
based on the importance score Itof each involved token and
another quantity Qt,y (we will shortly present the details of
the calculation of St,y). In the opposite case, we merely update
Sy. After all the tokens have been examined, we sort the
candidates list Y0in decreasing score order and we categorize
the product xin the highest scoring category Y0[0].
Now let us elaborate on the computation of the St,y scores.
Here we adopt a strategy similar to the ranking in Web

search engines and we define St,y as a linear combination
of the ‘query independent’ importance scores Itand a ‘query
dependent’ term Qt,y as follows:
St,y =k1It+k2Qt,y (2)
where k1and k2are two constant quantities. Regarding the
Qt,y term, we experimented with dozens of functions and
we found that the best accuracy is achieved by employing a
variation of the well-established td-idf scoring scheme which
has been proved quite effective in many classification tasks.
The variation we used is given by the following equation:
Qt,y = log ft,y log |Y|
ft
(3)
where ft,y is the aforementioned frequency value which stores
the number of correlations of a token twith a category y.
E. Dimensionality Reduction & Model Pruning
The existence of multiple types of tokens in the aforemen-
tioned lexicon structure is translated to an increase in the
overall number of features extracted by the model. This in
turn leads to a high-dimensional feature space, which must
be reduced to ensure the scalability of our solution in case
large datasets are utilized. In this subsection, we present a
method which effectively decreases the dimensionality of the
involved feature space, and, also, it has a minimal impact on
the classification effectiveness of the algorithm.
More specifically, our proposed dimensionality reduction
approach takes into consideration the importance scores It,y
which we discussed earlier and preserves only the tokens
whose importance score is greater than a pre-defined constant
cut-off threshold C. The rest of the tokens are considered of
low informational value (i.e. they are considered ambiguous)
and are evicted from the model. The cut-off threshold is
defined as a percentage Tof the greatest importance score
in the lexicon according to the following equation:
C=Tmax
t,y It,y (4)
We experimented with various values of Tand in our
evaluation we present detailed diagrams of how Taffects the
overall size of the model, and the fluctuation of the accuracy of
our algorithm against the adversary classification approaches.
IV. EXP ER IM EN TS
Now let us we demonstrate the effectiveness of the proposed
method in terms of both classification accuracy and trained
model size.
All the experiments were performed on a workstation
equipped with a single Intel CoreI7 7700@3.6GHz CPU
(single-core mode) and 32GB of RAM, running Ubuntu Linux
16.04 LTS.
The dataset we used was acquired by deploying a crawler on
a popular product comparison platform, shopmania.com. This
service enlists tens of millions of products organized in a three-
level hierarchy comprised of 230 categories. The two higher
levels of this hierarchy include 39 categories, whereas the third
lower level accommodates the rest 191 leaf categories. Each
product is categorized into this tree structure by being mapped
to only one leaf category. Some of these 191 leaf categories
contain millions of products. However, shopmania.com allows
only the first 10,000 products to be retrieved per category.
Under this restriction, our crawler managed to collect 313,706
products.
In the following analysis of the experimental results, we
will refer to the proposed algorithm by using the abbreviation
SPC (Supervised Products Classifier). The accuracy of SPC is
compared against two well-established supervised classifica-
tion algorithms: Logistic Regression (LogReg), and Random
Forests (RanFor).
A. Classification Accuracy
In this subsection we examine the effectiveness of our
approach in terms of products classification accuracy. Figure 1
illustrates the performance of SPC and its two adversary
algorithms for the four following feature spaces:
•N= 1: this feature space includes only the words of the
product titles, that is, no n-grams are constructed (top-left
diagram).
•N= 2: we take into consideration both the words and the
bigrams extracted from product titles (top-right diagram).
•N= 3: the feature space consists of all the unigrams,
bigrams and trigrams (bottom-left diagram).
•N= 4: the feature space consists of all the 1,2,3,4-
grams (bottom-right diagram).
For each feature space, we measure the classification accu-
racy of all methods for various pruning levels. In particular,
we modify the value of the parameter Tof equation 4 in
the range [0.0,1.0] by increasing its value by 0.1 each time.
For instance, the pair N= 2, T = 0.5is translated into a
feature space which includes the words and the bigrams of the
product titles whose importance score is greater than half of
the highest computed importance score. Consequently, for the
requirements of this experiment we constructed 44 different
datasets, that is, 11 datasets for each value of N.
The performance evaluation was conduced by using 60% of
the records in the dataset for training, whereas the rest 40% are
left for testing. To ensure the correctness of our experiments
and to present unbiased results, all the adversary approaches
were given exactly the same set of features with the same
importance scores as weights.
For the first feature space (N= 1, top-left diagram of
Figure 1) and with all words present (T= 0), our method
achieved a remarkable product classification accuracy of 88%.
This accuracy is essentially left intact even if we remove all
the words whose importance score is lower than the 20% of the
maximum. At T= 0.3and T= 0.4the accuracy is slightly
decreased at 86.1% and 84.8% respectively. After these values,
the performance degrades quickly. SPC outperformed both
the opponent classification methods in all cases. Regarding
Logistic Regression, the accuracy started from roughly 82% (at
T= 0) and it decreased at similar rates as SPC. On the other

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold T
Accuracy
SPC LogReg RanFor
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold T
Accuracy
SDC LogReg RanFor
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold T
Accuracy
SDC LogReg RanFor
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold T
Accuracy
SDC LogReg RanFor
Fig. 1. Comparison of the classification performance of SPC against Logistic Regression (LogReg) and Random Forests (RanFor). Accuracy fluctuation with
respect to the value of the cut-off threshold Tfor: i) N= 1 (top left), ii) N= 2 (top right), iii) N= 3 (bottom left), and iv) N= 4 (bottom right).
hand, the performance of Random Forests was significantly
lower; for T= 0 it was measured at 59.5% and it dropped in
higher rates compared to the other two algorithms.
Similar notifications can be made for the rest of the ex-
amined feature spaces. In the second feature space (N= 2,
top-right diagram of Figure 1), the accuracy of SPC increases
to 92.5% and begins to decrease for values of T > 0.7. On
the other hand, the accuracy achieved by Logistic Regression
and Random Forests for all words and bigrams (T= 0) was
85% and 65.5%, respectively. The diagram shows that the
performance of Logistic Regression degrades at a rate which
is similar to the one of SPC, whereas for Random Forests, the
degradation occurs at a remarkably higher rate.
Our algorithm achieved its highest performance for N= 3
and T∈[0.0,0.2]; its accuracy in these cases reached 95.1%.
Similarly to the previous case, the accuracy falls below 90%
for a threshold value of T > 0.7, whereas for T= 0.6it is
measured at 93%, only 2.1% lower than the maximum. As we
will see shortly in the following subsection, the model size
for T= 0.6is smaller by 47% in terms of number of tokens
in the lexicon, and 54% in terms of stored RDV entries. In
other words, by pruning nearly half of the model data, we
lose only about 2% of its effectiveness. Regarding the two
other methods, the accuracy for Logistic Regression in this
experiment was measured at 87%, significantly lower than that
of our SPC. The performance of Random Forests was even
worse, at 68%.
Notice that the same values of maximum accuracy was also
measured in the largest feature space (N= 4), for all the three
examined methods. This leads to the conclusion that there is
no practical need to compute tokens which are longer than 3
words. In this case, the decrease in the performance is slightly
faster for SPC, since the accuracy falls below 90% for T >
0.6. The performance degradation pattern is similar for both

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.5
1
1.5
2
2.5
3
·105
Threshold T
Model size
Tokens RDV entries
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
1
2
3
4
5
6
7
8
9
·105
Threshold T
Model size
Tokens RDV entries
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.25
0.5
0.75
1
1.25
1.5
1.75
2
·106
Threshold T
Model size
Tokens RDV entries
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.5
1
1.5
2
2.5
3
·106
Threshold T
Model size
Tokens RDV entries
Fig. 2. Model size (overall number of tokens and RDV entries) reduction with respect to the value of the cut-off threshold Tfor: i) N= 1 (top left), ii)
N= 2 (top right), iii) N= 3 (bottom left), and iv) N= 4 (bottom right).
Logistic Regression and Random Forests.
B. Model Size
Here we study the size of the model generated by SPC and
we evaluate the effectiveness of our dimensionality reduction
method.
In Table IV-B we report the increase in the size of our
model with respect to the value of n. In case n= 1 (i.e. only
single words are included and no n-grams are constructed),
the model consists of 127,407 words and their RDVs include
226,450 relationships with various categories. This means that
each word is correlated on average with roughly 1.8 categories.
In the next row we show the model size when only the bigrams
are inserted into the lexicon. There are about 460 thousands
of such bigrams in the product titles of our dataset. Since their
RDVs store approximately 542 thousands items, each bigram
is associated on average with only 1.17 categories. On a similar
spirit, the trigrams and the 4-grams are correlated on average
with 1.07 and 1.04 categories. These numbers verify one of
our basic assumptions which states that the longer a token is,
the less ambiguity it carries. Consequently, higher importance
scores should be assigned to longer tokens.
In the sequel, we examine how our proposed dimensionality
reduction strategy affects the size of the model. The setup
of this experiment is the same as in the evaluation of the
classification performance; that is, we have the aforementioned
four feature spaces and we modify the value of the cut-off
threshold Tin the range [0.0,1.0]. The results are depicted in
Figure 2 and illustrate the reduction in the number of tokens
and the number of RDV entries stored in the lexicon, as T
increases. Notice that for T= 0, the model sizes are identical
to the ones that we report in Table IV-B.
Here we are primarily interested in two model sizes: i) the

TABLE I
MOD EL SI ZE S FOR VAR IOU S VALU ES OF n.
nTokens RDV entries
1(words) 127,407 226,450
2(bigrams) 460,144 542,180
3(trigrams) 653,525 699,639
4(4-grams) 744,699 772,578
1,2587,551 768,630
1,2,31,241,076 1,468,269
1,2,3,41,985,775 2,240,847
initial, and ii) the size of the model at the point where accuracy
starts to decrease significantly. For this reason, it is required
that we combine the data from both Figures 1 and 2.
For N= 1, the top-left diagram of Figure 1 shows that at
T= 0.4the accuracy starts to decrease at high rates. At this
point, Figure 2 tells us that the model includes 126,528 tokens
and 194,184 RDV entries. Compared to the initial model size
(127,407 tokens and 226,450 RDV entries), we have a decrease
of less than 1% in the number of tokens and about 15% in
the number of RDV entries. This means that for N= 1,
our method prunes only a few words which have long RDVs
(i.e. they are correlated with a large number of categories,
consequently, they are ambiguous). Although the reduction
in the model size is not very large, the situation improves
significantly for the other larger feature spaces (which also
yield higher accuracy values).
Indeed, in the feature space which includes both words and
bigrams (N= 2), the accuracy starts to decrease at T=
0.7(top-right diagram of Figure 1). At that point, the model
contains 455,635 tokens and 503,292 RDV entries. Compared
to the initial 587,551 tokens and 768,630 RDV entries, the
reduction we achieve is approximately 23% in the number of
tokens and 35% in the RDV entries.
Finally, for N= 3 and N= 4, the reduction in the model
sizes is even greater: the number of tokens is decreased by 47%
and 63%, respectively, whereas the number of RDV entries is
reduced by 54% and 66%, respectively.
V. CONCLUSIONS AND FUTURE WORK
In this paper we introduced a new supervised learning
method for automatic product categorization. The problem is
particularly important for all e-commerce platforms as a robust
solution can lead to numerous novel applications such as sim-
ilar product suggestions and personalized recommendations.
Our proposed model is based on the extraction of various
types of n-grams which, in combination with the single words
and other types of tokens, improves the effectiveness of the
categorization process. According to the proposed method,
all tokens are inserted into a lexicon and they are assigned
importance scores based on various parameters which reflect
their ambiguity. The ambiguity is a notion which represents
how significant a token is when it is used to determine the
class of an unlabeled product.
The existence of multiple types of tokens may enlarge
the model significantly and, therefore, affect its scalability
in a negative manner. For this reason, we also proposed a
dimensionality reduction method based on the importance
scores of the tokens. According to this method, the non
important tokens are removed from the lexicon of the model,
since they are considered of low informational value.
The evaluation of our method demonstrated its effectiveness,
since for a feature space which includes the words, the bigrams
and the trigrams of the product titles, the classification ac-
curacy exceeds 95%. Furthermore, we experimentally proved
that in this case, the pruning of nearly half of the model data
decreases accuracy by an insignificant margin of only 2%.
Finally, we intend to further study and develop this model
with the aim of testing it in real-world large-scale and high-
dimensional data from actual product catalogs. A detailed
investigation of the efficiency of the model is also within our
intentions.
REFERENCES
[1] J. Chen and D. Warren, “Cost-sensitive Learning for Large-scale Hier-
archical Classification,” in Proceedings of the 22nd ACM International
Conference on Information and Knowledge Management, 2013, pp.
1351–1360.
[2] Z. Kozareva, “Everyone Likes Shopping! Multi-class Product Cate-
gorization for E-commerce,” in Proceedings of the 2015 Conference
of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, 2015, pp. 1329–1333.
[3] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet Allocation,”
Journal of Machine Learning Research, vol. 3, pp. 993–1022, 2003.
[4] V. Gupta, H. Karnick, A. Bansal, and P. Jhala, “Product Classifica-
tion in E-commerce using Distributional Semantics,” arXiv preprint
arXiv:1606.06083, 2016.
[5] D. Shen, J. D. Ruvini, M. Somaiya, and N. Sundaresan, “Item Catego-
rization in the E-commerce Domain,” in Proceedings of the 20th ACM
International Conference on Information and Knowledge Management,
2011, pp. 1921–1924.
[6] D. Shen, J.-D. Ruvini, and B. Sarwar, “Large-scale Item Categorization
for E-commerce,” in Proceedings of the 21st ACM International Confer-
ence on Information and Knowledge Management, 2012, pp. 595–604.
[7] L. Akritidis and P. Bozanis, “Effective Unsupervised Matching of
Product Titles with k-Combinations and Permutations,” in Proceedings
of the 14th IEEE International Conference on Innovations in Intelligent
Systems and Applications, 2018.
[8] X. Zhang, J. Zhao, and Y. LeCun, “Character-level Convolutional
Networks for Text Classification,” in Advances in Neural Information
Processing Systems, 2015, pp. 649–657.
[9] S. Lai, L. Xu, K. Liu, and J. Zhao, “Recurrent Convolutional Neural
Networks for Text Classification,” in Proceedings of the 29th AAAI
Conference on Artificial Intelligence, vol. 333, 2015, pp. 2267–2273.
[10] Y. Kim, “Convolutional Neural Networks for Sentence Classification,”
in Proceedings of the Conference on Empirical Methods in Natural
Language Processing, 2014, pp. 1746–1751.
[11] P. Wang, J. Xu, B. Xu, C. Liu, H. Zhang, F. Wang, and H. Hao,
“Semantic Clustering and Convolutional Neural Network for Short Text
Categorization,” in Proceedings of the 53rd Annual Meeting of the
Association for Computational Linguistics and the 7th International
Joint Conference on Natural Language Processing, vol. 2, 2015, pp.
352–357.
[12] G.-R. Xue, D. Xing, Q. Yang, and Y. Yu, “Deep Classification in Large-
scale Text Hierarchies,” in Proceedings of the 31st Annual International
ACM SIGIR Conference on Research and Development in Information
Retrieval, 2008, pp. 619–626.
[13] J.-W. Ha, H. Pyo, and J. Kim, “Large-scale Item Categorization in E-
commerce using Multiple Recurrent Neural Networks,” in Proceedings
of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, 2016, pp. 107–115.
[14] C. Chavaltada, K. Pasupa, and D. R. Hardoon, “A Comparative Study of
Machine Learning Techniques for Automatic Product Categorisation,” in
International Symposium on Neural Networks, 2017, pp. 10–17.