Computer Science and Information Technologies

Vol. 3, No. 3, November 2022, pp. 137~147

ISSN: 2722-3221, DOI: 10.11591/csit.v3i3.pp137-147  137

Text classification to predict skin concerns over skincare using

bidirectional mechanism in long short-term memory

Devi Fitrianah1, Andre Hangga Wangsa2

1
Department of Computer Science, School of Computer Science, Bina Nusantara University, Jakarta, Indonesia
2
Department of Informatics, Faculty of Computer Science, Universitas Mercu Buana, Jakarta, Indonesia

Article Info ABSTRACT

Article history: There are numerous types of skincare, each with its own set of benefits
based on key ingredients. This may be difficult for beginners who are
Received Mar 10, 2022 purchasing skincare for the first time due to a lack of knowledge about
Revised Jul 29, 2022 skincare and their own skin concerns. Hence, based on this problem, it is
Accepted Sep 12, 2022 possible to find out the right skin concern that can be handled in each
skincare product automatically by multi-class text classification. The
purpose of this research is to build a deep learning model capable of
Keywords: predicting skin concerns that each skincare product can treat. By comparing
the performance and results of predicting the correct skin condition for each
Dermatology skincare product description using both long short-term memory (LSTM)
Multi-class text classification and bidirectional long short-term memory (Bi-LSTM), The best results are
deep learning given by Bi-LSTM, which has an accuracy score of 98.04% and a loss score
Natural language processing of 19.19%. Meanwhile, LSTM results have an accuracy score of 94.12% and
Skincare a loss score of 19.91%.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Andre Hangga Wangsa
Department of Informatics, Faculty of Computer Science, Universitas Mercu Buana
Jalan Raya Meruya Selatan no. 1, Kembangan Jakarta Barat-16550, Indonesia
Email: [email protected]

1. INTRODUCTION
Skincare products are the most mainstream cosmetics that maintain skin integrity, appearance, and
condition. The high market demand has made skincare products one of the most popular ways to deal with
skin concerns [1]. What's more, skincare trends began to rise drastically in 2020, when the COVID-19
pandemic began [2].
Skincare has various types and benefits according to the active ingredients contained in it [3]. Active
ingredients here play an important role in the performance of every skincare product because these
ingredients are chemicals that actively work on a specific target skin concern [4]. For example, salicylic acid
can reduce sebum secretion so that it can control oily skin and acne, but on the other hand, it can also cause
inflammation in sensitive and dry skin [5]. This is what makes skincare products difficult to use and
unsuitable for beginners; the user must thoroughly understand what is contained in them in order to meet
their skin care concerns and expectations [6].
Most beauty stores already manually sort all of their skincare products based on brands, skin types,
and skin concerns. But it will take a long time and require someone who knows about skincare products.
Instead, by collecting all information that relates to skincare products, such as the function of the product in
dealing with certain skin concerns, we might be able to build a model that can automatically classify and
predict the benefits of those skincare products quickly.

Journal homepage: http://iaesprime.com/index.php/csit

138  ISSN: 2722-3221

The information given to classify and predict is in the form of text data, which is a description of
skincare products, so it is called a text classification. Text classification is one of the tasks in natural language
processing (NLP), which aims to assign labels or targets to textual features or classes such as sentences,
queries, paragraphs, and documents [7]. There are two problems in text classification: binary and multi-class
classification. Binary classification is made up of only two labels, one of which is assigned a value in an
arbitrary feature space X [8]. Whereas multi-class classification has more than two labels [9].
There are various kinds of research on multiclass classification problems, despite the different uses
of domains or topics, data types, and algorithms. Although there is currently no research related to skin care
products, there are several studies that discuss the dermatology domain. The goal of Indriyani and Sudarma's
[3] study was to categorize facial skin types, which were divided into four categories: normal, dry, oily, and
combination skin. They used an image-type dataset of sixty facial images captured manually with a digital
camera. Although this makes it into computer vision instead of NLP, at least with a case that aims to classify
multiclass facial skin types and also by using a supervised learning algorithm, support vector machine
(SVM), The result is an average accuracy score of 91.66% and an average running time of 31.571 seconds,
which is higher than in previous studies [10]. The following study made extensive use of deep neural network
algorithms such as convolutional neural network (CNN), recurrent neural network (RNN), and long short-
term memory (LSTM) [11]. Although there are a few cases of binary classification because some datasets
only have two classes, the majority of the datasets have between five and ten classes. The research combines
several of those algorithms into a hybrid framework. Not only that, some algorithms are also modified into a
bidirectional mechanism. The proposed model achieved excellent performance on all tasks. A bidirectional
recurrent convolutional neural network attenuation-based (BRCAN) gave accuracy scores on the four multi-
class classification tasks of 73.46%, 75.05%, 77.75%, and 97.86%; those results are higher than all
comparison algorithms.
In relation to the aforementioned studies, we proposed a comparison of unidirectional/regular LSTM
and bidirectional long short-term memory (Bi-LSTM) in our own dataset collected from several skincare
online stores to classify skin concerns of each skincare product. The main purpose of this research is to find
out the difference between the performance results of the two proposed algorithms. In other research,
bidirectional mechanisms, which have layers that work forward and backward in sequence, are able to
outperform unidirectional LSTM [12].

2. METHOD
This section of the paper presents the research methodology. When doing research, researchers must
obtain data that will be studied for later processing. After obtaining the data, the data is still in the form of
raw data, which then the researcher must prepare the data to become a data set that can be processed. After
the data has gone through the processing stage, the last stage the researcher must do is to evaluate the
research model or instrument to understand their performance, as well as its strengths and weaknesses. More
details, there are several stages can be seen in Figure 1.

Figure 1. The proposed research methodology

2.1. Data collection

In this research, data collection was implemented by using the Web Scraping technique. Web
Scraping is used to convert unstructured data into structured data that can be stored and analyzed in a central

Comput Sci Inf Technol, Vol. 3, No. 3, November 2022: 137-147

Comput Sci Inf Technol ISSN: 2722-3221  139

local database or spreadsheet [13]. The data is collected on a beauty online store website which is
lookfantastic.com, dermstore.com, allbeauty.com, sokoglam.com, and spacenk.com which market products
such as skincare, makeup, and beauty tools.

2.2. Data merging

The data collected are divided into three categories according to the seven skin concerns handled by
each skincare product, due to having similar symptom treatment. The three categories are drynes, redness;
anti-aging, wrinkles; acne, big pores, blemish. Data has 6 attributes which is skincare name, skincare price,
how to use, skin concerns, product description, ingredients, and active ingredients. The data that has been
collected is merged into one dataset with a total of 5183 rows.

2.3. Data cleaning

Due to data that has the same value (duplicate data) and data that has no value (null data), then data
cleaning is carried out by removing duplicate and null data evenly. Data cleaning greatly improves the
accuracy of machine learning models, which however requires broad domain knowledge to identify examples
that will influence the model [14]. After that, the total dataset is reduced to 5152 rows. However, in this
study, we will focus on the attributes of product descriptions that will become features and skin problems that
will become labels. So, we will delete the other columns that are not necessary to make the process easier
going forward.

2.4. Text preprocessing

Before fed the dataset to our models, it’s necessary to perform a data pre-processing stage.
According to the Figure 2, there is a several data pre-processing task including Case Folding, Punctuation
Removal, Whitespace Removal, Numbers Removal, Stopword Removal, Lemmatization. Case folding is the
process to convert all input words into the same form, for instance uppercase or lowercase [15]. So, we
transform all our text in description product as the features to lowercase. After that, our text data must be
clean from punctuation marks and symbol, so we applied punctuation removal. Next, we applied whitespace
removal to remove an unpredicted extra spaces between every word and line or paragraph spacing [16]. We
must make sure that our texts only contain meaningful words which aim to represent the essence of each text.
So, we need to apply stopword removal. Stopwords are actually the most common words in any language
that appears too much in a text does not add much information, such as articles, prepositions, pronouns and
conjunctions. Final step in text preprocessing is lemmatization. Lemmatization works to reduce a word
variant to its lemma and uses vocabulary and morphological analysis for returning words to their dictionary
form [17]. This step converts all of word in our texts to its basic form. Generally, lemmatazation and
stemming is a similar approach and often produce same results, but sometimes the basic form of the word
may be different than the stemming approach e.g. "caring" is stemmed to "car", but in lemmatization you will
get "care" which more appropriate than stemming.Also, in Boban [17] study, Lemmatization produces better
results.

2.4.1. Create data tensor

After our text successfully passing data preprocessing stage. We need to vectorize our features by
convert our text data into either a sequence of integers and mapping it into real-valued vector, so we can feed
it through input layer in our deep neural network models. Also, we limit the total number of words in our text
features to the most frequent words, and zero out the rest. We determine the maximum sentence length
(number of words) in each text features that will truncating long reviews and pad the shorter reviews with
zero values in the next process. According to Figure 2 there are some steps in converting our text data after
lemmatizing step called creat data tensor. First, we use tokenizer to split each word in the text. Second, we
create an index-based dictionary on each word based on the text we have or the description of skincare
products. Next, we transform our tokens from first step into sequence of integer based from our index-based
dictionary. Then, truncate and pad the input sequences, so they are all in the same length for modeling. Last
step is converting our categorical labels to numbers.

2.5. Model building and training

Next stages are model building and training. Before that, we split our dataset into three part for
training, testing, and validating. We build our LSTM and Bi-LSTM model with a similar layer structur as
illustrated in Figure 3.

Text classification to predict skin concerns over skincare using bidirectional mechanism … (Devi Fitrianah)
140  ISSN: 2722-3221

Figure 2. Text preprocessing Figure 3. Model architecture

2.5.1. Embedding layer

We put Embedding layer in first place as input layer ad map each word into a real-valued vector to
represent each word. Embedding layers works by mapping a raw user/items features in a high dimensional
space to dense vectors in a low dimensional embedding space [18]. Basically, embedding layer has similar
purpose as popular word embedding frameworks (e.g word2vec and gloVe) which provide a dense
representation of words and their relative meanings. However, there is a different between them, which are
their training process. Popular word embeddings framework like word2vec and gloVe is trained to predict if
word belongs to the context, given other words, e.g., to tell if "cuisine" is a likely word given the "The chef is
making a chinese ... " sentence begging. Word2vec learns that "chef" is something that is likely to appear
together with "cuisine", but also with "worker", or "restaurant", so it is somehow similar to "waitress", so
word2vec learn something about the language. The conclusion is embeddings created by word2vec, gloVe, or
other similar frameworks learn to represent words with similar meanings using similar vectors. Meanwhile,
embeddings learned from layer of neural network may be trained to predict a specific case, in this case is text
classification. So, the embeddings would learn features that are relevant for our text classification. If
word2vec has a pre-trained corpus or dictionary, otherwise, embedding layers doesn’t have it. But we already
created the index-based dictionary on each word from our features before and transform our features to
sequence of integer through it. It’s more efficient, doesn’t need high computing resources, and useful for
classification than using pre-trained word embedding like word2vec, even though embedding layer doesn’t
capture the semantic similarity of words like word2vec does [19].

2.5.2. Spatial dropout 1D layer

Next, we adding spatial dropout 1D layer. This layer performs the same function as dropout. In
standard dropout, the neuron on neural network drops independently as shown in Figure 4(a) [20].
Meanwhile, in spatial dropout it drops entire 1D feature maps instead of individual elements as shown in
Figure 4(b).

(a) (b)

Figure 4. Shows the difference between, (a) regular dropout that drops neuron in neural network
independently and (b) spatial dropout 1D that drops entire 1D feature maps

2.5.3. Unidirectional and bidirectional long short-term memory (Bi-LSTM)

Next, we use the LSTM layer and the Bi-LSTM layer on each of the two architectural models
created. LSTM is very popular for dealing with cases such as NLP, video, and audio where the data is in the

Comput Sci Inf Technol, Vol. 3, No. 3, November 2022: 137-147

Comput Sci Inf Technol ISSN: 2722-3221  141

form of a sequence. When compared with its predecessor vanilla RNN algorithm which is unable to use past
information, LSTM outperforms it with its long-term memory. LSTM transforms the memory shape of cells
withinside the RNN via way of means of reworking the tanh activation characteristic layer withinside the
RNN right into a shape containing memory devices and gate mechanisms, pursuits to determine how to make
use of and replace data saved in memory cells [21]. Now, there is a new concept of mechanism in those
sequence feed-forward neural network which called bidirectional. Bidirectional is a mechanism that able to
make a neural network works like two-way mirror, which trains an input data twice through past and future.
With implementing the bidirectional concept, a regular LSTM not only capable train the input data forward,
but also backward. According to Figure 5, Figures 6(a) and 6(b), those models are used the following formula
to calculate the predict values,

Figure 5. LSTM architecture

(a) (b)

Figure 6. Shows the differences of LSTM in: (a) unidirectional and (b) Bidirectional mechanism

𝑙𝑡 (Input Gate) = 𝜎𝑔 (𝑊𝑖 𝑋𝑡 + 𝑅𝑖 ℎ𝑡−1 + 𝑏𝑖 ),

𝑓𝑡 (Forget Gate) = 𝜎𝑔 (𝑊𝑓 𝑋𝑡 + 𝑅𝑓 ℎ𝑡−1 + 𝑏𝑓 ),
(1)
𝐶𝑡(Cell Candidate) = 𝜎𝑔 (𝑊𝑐 𝑋𝑡 + 𝑅𝑐 ℎ𝑡−1 + 𝑏𝑐 ),
𝑂𝑡(Output Gate) = 𝜎𝑔 (𝑊𝑜 𝑋𝑡 + 𝑅𝑜 ℎ𝑡−1 + 𝑏𝑜 ),

𝜎𝑔 = The gate activation function

𝑊𝑖 , 𝑊𝑓 , 𝑊𝑐 , and 𝑊𝑜 = Input weight matrices
𝑅𝑖 , 𝑅𝑓 , 𝑅𝑐 , and 𝑅𝑜 = Recurrent weight matrices
𝑋𝑡 = The data input.
ℎ𝑡−1 = The output at the previous time (t − 1)
𝑏𝑖 , 𝑏𝑓 , 𝑏𝑐 , and 𝑏𝑜 = The bias vector

Text classification to predict skin concerns over skincare using bidirectional mechanism … (Devi Fitrianah)
142  ISSN: 2722-3221

the forget gate counts the measure that decide to removes the previous memory values from the cell state.
Just like the forget gate, the input gate determines the new input to the cell state. Then, the LSTM’s cell state
Ct and the output Ht at time t are calculated,

𝐶𝑡 = 𝑓𝑡 ⊙ 𝐶𝑡 ⊙ 1 + 𝑙𝑡 ⊙ 𝑔𝑡
(2)
𝐻𝑡 = 𝑂𝑡 ⊙ 𝜎𝑐(𝐶𝑡 )

⊙ = denotes the Hadamard product (element-wise multiplication of vectors)

Also, we use another parameter in our hidden layer and output layer of both LSTM and Bi-LSTM
which are dropout, recurrent dropout, recurrent regularizer, L2 regularizers. Recurrent dropout is a
regularization that devoted recurrent neural network algorithms. Recurrent dropout works differently from
the usual dropout, which is applied to for-ward connections of feed-forward architectures or RNNs, drop
neurons directly in recurrent connections in away that does not cause loss of long-term memory instead [22].
There is a formula update on Ct when implementing recurrent dropout to the cell update vector gt,

𝐶𝑡 = 𝑓𝑡 ⊙ 𝐶𝑡 ⊙ 1 + 𝑙𝑡 ⊙ 𝑑(𝑔𝑡 )
(3)
Where d is dropout. Next parameter is usual dropout that we apply same with recurrent dropout
where in both LSTM and Bi-LSTM layer. Last parameter is L2 regularizers which is a layer weight
regularizers that enforce penalties on layer parameters or layer activity during optimization process. These
penalties are added up in a loss function that optimizes the network applied on a per-layer basis there are
three ways to apply these regularizer, in layer’s kernel, bias, and output. L2 regularizer summed the suared
weights to the loss function. L2 are often to set a value on logarithmic scale between 0 and 0.1, such as 0.1,
0.001 and 0.0001.

2.6. Model evaluation and prediction

Final stages are model evaluation and prediction with a validation dataset. The evaluation contains a
several score to measure the performance of model training and testing. We use an accuracy score by
obtaining precision, recall, and f-measure.
𝑇𝑃+𝑇𝑁
Accuracy = (4)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

TP = True Positive is a skin concern that is in the actual label and appears in the prediction.
FP =False Poisitive is a skin concern that is in actual label but doesn’t appears in the prediction.
FN = False Negative is skin concern that is not in the actual label but appears in the prediction.
TN = True Negative is a skin concern that is neither in the actual label nor the prediction
Precision is the percentage of positive cases that were actually predicted to be truly positive [23].
Precision is calculated,
𝑇𝑃
Precision = (5)
𝑇𝑃+𝐹𝑃

recall is the Percentage of actual positive cases that were correctly predicted. It actually measures the
coverage of positive cases and accurately reflects the predicted cases [23]. Recall is calculated,
𝑇𝑃
Recall = (6)
𝑇𝑃+𝐹𝑁

F1- Measure is a composite measure that captures the trade-offs related to precision and recall and calculated,

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥𝑅𝑒𝑐𝑎𝑙𝑙
F1-Measure = (7)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

loss function that being used is categorical cross-entropy. Categorical cross-entropy is specifically used for
the case of multi-class classification which increasing or decreasing the relative penalty of a probabilistic
false negative for an individual class [24]. The categorical cross-entropy loss function is used the following
formula,
output
size
Loss = − ∑𝑖=1   𝑦𝑖 ⋅ log 𝑦ˆ𝑖 (8)

Comput Sci Inf Technol, Vol. 3, No. 3, November 2022: 137-147

Comput Sci Inf Technol ISSN: 2722-3221  143

𝑦ˆ𝑖 = 𝑖-th scalar value in the model output

𝑦𝑖 = Corresponding target value
Output Size = The number of scalar values in the model output
this loss function measures the distance of dissimilarity between the true label distribution and the predicted
label distribution. The y_i defines the probability that event i occurs. The sum of all y_i is 1 that means one
event may occur. The minus sign guarantees that the closer the distributions are to each other, the smaller the
loss. Also, we use a confusion matrix to calculate the total of true or false a prediction generated by the
classification model. Confusion matrix is machine learning concept that contains information about the actual
and predicted classifications performed by the classification system which has two dimention divided for
indexing the actual class of an object, and the other is indexing the class that the classifier predicts [25].

3. RESULTS AND DISCUSSION

All stages of this research were carried out with the python programming language. The results of
this research are measured using several scores. The scores measure the performance of the proposed model
classification prediction, by looking at the accuracy and loss scores in each experiment carried out.

3.1. Train-test-validation split evaluation

The first experiment was carried out by splitting the dataset into three parts. The three parts of
dataset used for train, test, and validation process. Table 1 shows the result of dataset spliting with the best
result of 80% train dataset, 1% test dataset, and 19% validation dataset with an acuracy score of 98.04% and
loss 19.19% from Bi-LSTM.

Table 1. The result based on the distribution of dataset splitting

Train/Test/Validation Split LSTM Accuracy LSTM Loss Bi-LSTM Accuracy Bi-LSTM Loss
80/1/19 0.9412 0.1991 0.9804 0.1919
80/ 2/18 0.9401 0.2069 0.9800 0.1991
80/3/17 0.9405 0.1919 0.9801 0.2007
80/4/16 0.9400 0.2020 0.9611 0.1910
80/5/15 0.9258 0.2276 0.9690 0.2205
80/6/14 0.9635 0.2481 0.9800 0.2287
80/7/13 0.9643 0.2311 0.9750 0.2210
80/8/12 0.9681 0.2490 0.9800 0.2294
80/9/11 0.9292 0.2410 0.9790 0.2910
80/10/10 0.9261 0.2911 0.9601 0.2411
80/11/9 0.9600 0.2450 0.9780 0.2910
80/12/8 0.9278 0.2101 0.9501 0.2934
80/13/7 0.9210 0.2105 0.9309 0.2451
80/14/6 0.9181 0.2980 0.9187 0.2410
80/15/5 0.9082 0.2949 0.9182 0.2410
80/16/4 0.9009 0.2910 0.8890 0.3800
80/17/3 0.8829 0.3991 0.8898 0.3809
80/18/2 0.8821 0.3929 0.8810 0.3876
80/19/1 0.8832 0.3901 0.8824 0.3792
90/1/9 0.9290 0.3519 0.9790 0.2509
90/2/8 0.9283 0.3210 0.9174 0.2410
90/3/7 0.9043 0.3210 0.9111 0.2901
90/4/6 0.9021 0.3410 0.9019. 0.3100
90/5/5 0.9080 0.3410 0.8978 0.3240
90/6/5 0.8880 0.3450 0.8901 0.3210
90/7/3 0.8821 0.3421 0.8981 0.3209
90/8/2 0.8901 0.3450 0.8999 0.3210
90/9/1 0.9059 0.3592 0.9079 0.3465
Best Score 0.9804 0.1919

3.2. Hyper-parameters tuning

There are a several hyper-parameters used in model training. Memory units (Mu), Optimizers (O),
Activity function (Af) tuning as shown in Table 2. The Bi-LSTM model still outperformed the LSTM with a
memory unit setting of 100, RMSprop optimizers, and Activity function softmax.
Next, early stopping callback is a parameter that stop the training process when metric has stopped
improving by stores the model’s weights at the optimal epoch. These parameter attain the highest accuracy
intraining regardless of the epoch setting [26]. These parameter has two hyper-parameter which is patience
(p) and minimal delta (-Δ). The result of tuning these two hyper-parameter as shown in Table 3. The
Bi-LSTM model still outperformed the LSTM with a patience of 5 and min delta of 0.0001.
Text classification to predict skin concerns over skincare using bidirectional mechanism … (Devi Fitrianah)
144  ISSN: 2722-3221

Table 2. Memory units, optimizers, activity function tuning

Model Mu/O Mu = 100 Mu = 200 Af
Adam Acc=0.9054 Loss=0.3811 Acc=0.8899 Loss=0.3723
RMSprop Acc=0.9412 Loss=0.1991 Acc=0.9009 Loss=0.3792
SGD Acc=0.7821 Loss=0.5978 Acc=0.7811 Loss=0.5985 Softmax
Adadelta Acc= 0.7890 Loss=0.5821 Acc= 0.7799 Loss=0.4951
Adagrad Acc=0.7829 Loss=0.5435 Acc=0.7826 Loss=0.5433
Adam Acc=0.8054 Loss=0.4811 Acc=0.8099 Loss=0.4723
RMSprop Acc=0.8059 Loss=0.4592 Acc=0.7963 Loss=0.4352
SGD Acc=0.6821 Loss=0.6978 Acc=0.6816 Loss=0.6985 Sigmoid
Adadelta Acc= 0.6890 Loss=0.6821 Acc= 0.6799 Loss=0.5951
Adagrad Acc=0.7829 Loss=0.6435 Acc=0.6826 Loss=0.6433
Adam Acc=0.5063 Loss=0.5841 Acc=0.5099 Loss=0.4323
RMSprop Acc=0.5059 Loss=0.5592 Acc=0.5003 Loss=0.5365
SGD Acc=0.6821 Loss=0.6978 Acc=0.4821 Loss=0.5953
LSTM Adadelta Acc= 0.5890 Loss=0.5821 Acc= 0.5939 Loss=0.5951 ReLu
Adagrad Acc=0.6829 Loss=0.5435 Acc=0.5826 Loss=0.6433
Adam Acc=0.5054 Loss=0.5811 Acc=0.5099 Loss=0.4723
RMSprop Acc=0.5059 Loss=0.5592 Acc=0.7963 Loss=0.4152
SGD Acc=0.5821 Loss=0.5978 Acc=0.4816 Loss=0.5985
Adadelta Acc= 0.6890 Loss=0.6821 Acc= 0.6799 Loss=0.4951 Tanh
Adagrad Acc=0.5829 Loss=0.6435 Acc=0.5826 Loss=0.6433
Adam Acc=0.6054 Loss=0.4811 Acc=0.5099 Loss=0.4723
RMSprop Acc=0.6059 Loss=0.4592 Acc=0.7963 Loss=0.4152 Hard
SGD Acc=0.6821 Loss=0.4978 Acc=0.6821 Loss=0.4953
Adadelta Acc= 0.6890 Loss=0.4821 Acc= 0.4939 Loss=0.5951 Sigmoid
Adagrad Acc=0.6829 Loss=0.4435 Acc=0.6826 Loss=0.4433
Adam Acc=0.9034 Loss=0.3851 Acc=0.8889 Loss=0.3783
RMSprop Acc=0.9804 Loss=0.1919 Acc=0.9069 Loss=0.3592
SGD Acc=0.7861 Loss=0.5988 Acc=0.7881 Loss=0.5985
Adadelta Acc= 0.7890 Loss=0.5871 Acc= 0.7799 Loss=0.4961 Softmax
Adagrad Acc=0.7869 Loss=0.5495 Acc=0.7876 Loss=0.5483
Adam Acc=0.8054 Loss=0.4811 Acc=0.8089 Loss=0.4763
RMSprop Acc=0.8059 Loss=0.4592 Acc=0.7983 Loss=0.4392
SGD Acc=0.6821 Loss=0.6978 Acc=0.6886 Loss=0.6975
Adadelta Acc= 0.6890 Loss=0.6821 Acc= 0.6799 Loss=0.5671 Sigmoid
Adagrad Acc=0.7829 Loss=0.6485 Acc=0.6876 Loss=0.6483
Adam Acc=0.5073 Loss=0.5881 Acc=0.5079 Loss=0.4383
RMSprop Acc=0.5079 Loss=0.5582 Acc=0.5093 Loss=0.5375
SGD Acc=0.6861 Loss=0.6678 Acc=0.4881 Loss=0.5993
Bi-LSTM Adadelta Acc= 0.5870 Loss=0.5851 Acc= 0.5989 Loss=0.5971 ReLu
Adagrad Acc=0.6869 Loss=0.5475 Acc=0.5826 Loss=0.6493
Adam Acc=0.5064 Loss=0.5871 Acc=0.5079 Loss=0.4783
RMSprop Acc=0.5089 Loss=0.5572 Acc=0.7973 Loss=0.4182
SGD Acc=0.5861 Loss=0.5978 Acc=0.4896 Loss=0.5995
Adadelta Acc= 0.6860 Loss=0.6881 Acc= 0.6779 Loss=0.4971 Tanh
Adagrad Acc=0.5879 Loss=0.6465 Acc=0.5886 Loss=0.6493
Adam Acc=0.6074 Loss=0.4861 Acc=0.5069 Loss=0.4783
RMSprop Acc=0.6089 Loss=0.4572 Acc=0.7993 Loss=0.4172 Hard
SGD Acc=0.6861 Loss=0.4978 Acc=0.6881 Loss=0.4963 Sigmoid
Adadelta Acc= 0.6860 Loss=0.4881 Acc= 0.4979 Loss=0.5981
Adagrad Acc=0.6889 Loss=0.4475 Acc=0.6896 Loss=0.4463

Table 3. Patience and min delta tuning

Model P/-Δ -Δ = 0.01 -Δ = 0.001 -Δ = 0.0001
P=1 Acc=0.8999 Loss=0.3492 Acc=0.8854 Loss=0.3111 Acc=0.8814 Loss=0.3121
P=2 Acc=0.8829 Loss=0.3461 Acc=0.8839 Loss=0.3413 Acc=0.8889 Loss=0.3433
LSTM P=3 Acc=0.8899 Loss=0.3401 Acc=0.8808 Loss=0.3323 Acc=0.8878 Loss=0.3333
P=4 Acc=0.8959 Loss=0.3392 Acc=0.8854 Loss=0.3111 Acc=0.8821 Loss=0.3111
P=5 Acc=0.8829 Loss=0.3400 Acc=0.8808 Loss=0.3323 Acc=0.9412 Loss=0.1991
P=1 Acc=0.9009 Loss=0.3292 Acc=0.8999 Loss=0.3811 Acc=0.8954 Loss=0.3221
BI-LSTM P=2 Acc=0.8829 Loss=0.3461 Acc=0.8839 Loss=0.3453 Acc=0.8839 Loss=0.3443
P=3 Acc=0.8979 Loss=0.3398 Acc=0.8854 Loss=0.3117 Acc=0.8821 Loss=0.3111
P=4 Acc=0.8829 Loss=0.3400 Acc=0.8808 Loss=0.3323 Acc=0.9079 Loss=0.3562
P=5 Acc=0.8909 Loss=0.3409 Acc=0.8858 Loss=0.3111 Acc=0.9804 Loss=0.1919

3.3. Model evaluation and prediction

After the hyper-parameter tuning, we get the best settings are as shown in Table 4. We evaluate our
proposed models with a validation dataset as much as 980 skincare products. To measure theperformance of
model training and testing, we used an accuracy score by obtaining precision, recall, and f-measure as shown
in Table 5. Testing and validation confusion matrix in Bi-LSTM models are shown in Figures 7(a) and 7(b).

Comput Sci Inf Technol, Vol. 3, No. 3, November 2022: 137-147

Comput Sci Inf Technol ISSN: 2722-3221  145

Table 4. The best models’ settings

Hyper-Parameter LSTM Bi-LSTM
Train/Test/Validation data split 80/1/9 80/1/9
Max.Number of Words 50000 50000
Max.Sequence Length 512 512
Embedding Dimension 500 500
Memory Units 100 100
Optimizers RMSprop RMSprop
Activation Function Softmax Softmax
Spatial Dropout 1D 0.3 0.3
Dropout 0.3 0.3
Recurrent Dropout 0.3 0.3
Recurrent Regularizer 0.01 0.01
Kernel Regularizer 0.01 0.01
Bias Regularizer 0.01 0.01
Patience 5 5
Min.Delta 0.0001 0.0001
Accuracy Score = 0.9412 0.9804
Loss Score = 0.1991 0.1919

Table 5. Classification report on the validation data in the proposed models

Model Precision Recall F1-Score
0.9012 0.8939 0.8975 Micro Avg
LSTM 0.9004 0.8797 0.8894 Macro Avg
0.9015 0.8939 0.8972 Weighted Avg
0.8939 0.9939 0.8939 Samples Avg
0.8981 0.8908 0.8945 Micro Avg
BI-LSTM 0.8905 0.8839 0.8866 Macro Avg
0.8994 0.8908 0.8946 Weighted Avg
0.8908 0.8908 0.8908 Samples Avg

(a)

(b)

Figure 7. BI-LSTM Testing, (a) validation and (b) confusion matrix

3.3.1. Models inference

After getting the fine-tuned in each model, we tested the models to predicting what’s skin concern
that every skincare product overcomes by manually input the skincare product description to the models. The
actual labels over skincare description that we manually input before are taken from official website of each
skincare products. The results can be seen in Table 6.

Text classification to predict skin concerns over skincare using bidirectional mechanism … (Devi Fitrianah)
146  ISSN: 2722-3221

Table 6. Models inference

Skincare Description Actual LSTM Bi-LSTM
Label Prediction Prediction
Niacinamide 10% + Zinc 1% from The Ordinary is a water-based vitamin and mineral acne, acne, acne,
formula with 10% niacinamide and 1% zinc PCA. This water-based serum is great for big pores, big pores, big pores,
those looking for solutions for visible shine/enlarged pores/textural irregularities blemish blemish blemish
Benefits
Address signs of ageing with the Retinol Serum 0.2% in Squalane from The Ordinary; anti-aging, anti-aging, anti-aging,
a water-free, multipurpose, potent solution formulated to refine pores, reduce the wrinkles wrinkles wrinkles
appearance of dark spots and wrinkles and improve skin texture. Enriched with a
0.2% concentration of the anti-ageing powerhouse Retinol, which is a derivative of
Vitamin A, the lightweight serum has a plumping and firming effect on the
complexion, as well as protecting the skin from harmful environmental aggressors.
Another key antioxidant ingredient Squalane prevents UV damage and the formation
of age spots whilst counteracting harmful bacteria, leaving you with flawless skin.
Quench your skin in a wave of pure hydration with The INKEY List Hyaluronic Acid drynes, drynes, drynes,
Serum. This powerful ingredient attracts up to 1000x its weight in water, binding redness redness redness
moisture to restore the skin’s natural barrier. The gentle serum is suitable for all skin
types to restore balance.

4. CONCLUSION
The findings have produced a satisfactory performance, with an adequate score obtained both for
accuracy and loss. The performance of the bidirectional LSTM model, which makes use of this bidirectional
mechanism, outperforms that of the LSTM model, which makes use of the same mechanism and produces an
accuracy score of 94.12% and a loss value of 19.91%. The bidirectional LSTM model produces a score of
98.04% for its accuracy and a loss value of 19.19% for its loss. The usage of an embedding layer where the
data was previously transformed into a tensor form may be adjusted by employing a popular word embedding
such as word2vec or gloVe, which requires a large amount of computer resources but can extract the
semantic meaning of the features. Both of the models that have been provided have been able to effectively
train on the dataset that we obtain from well-known websites that specialize in the sale of skincare items.
Because of this, the prediction successfully maps the skin's worries over the description of each skincare
product, both with unseen data or validation data and the description that we manually enter into the models.
Additionally, given the dataset that we have, this research has the potential to be further expanded into a
recommendation system for online retailers that offer skincare goods as well as a mobile application.

ACKNOWLEDGEMENTS
We would like to thank all colleagues at the Faculty of Computer Science, Universitas Mercu Buana
who were involved in this research, either in terms of knowledge assistance and for their other support.

REFERENCES
[1] J. E. Lee, M. L. Goh, and M. N. Bin Mohd Noor, “Understanding purchase intention of university students towards skin care
products,” PSU Research Review, vol. 3, no. 3, pp. 161–178, 2019, doi: 10.1108/prr-11-2018-0031.
[2] H. Symum, F. Islam, H. K. Hiya, and K. M. A. Sagor, “Assessment of the Impact of COVID-19 pandemic on population level
interest in Skincare: Evidence from a google trends-based Infodemiology study,” medRxiv, 2020, doi:
10.1101/2020.11.16.20232868.
[3] Indriyani and I. Made Sudarma, “Classification of facial skin type using discrete wavelet transform, contrast, local binary pattern
and support vector machine,” Journal of Theoretical and Applied Information Technology, vol. 98, no. 5, pp. 768–779, 2020.
[4] A. Borrego-Sánchez, C. I. Sainz-Díaz, L. Perioli, and C. Viseras, “Theoretical study of retinol, niacinamide and glycolic acid with
halloysite clay mineral as active ingredients for topical skin care formulations,” Molecules, vol. 26, no. 15, 2021, doi:
10.3390/molecules26154392.
[5] S. Khezri and K. Khezri, “The side effects of cosmetic consumption and personal care products,” Journal of Advanced Chemical
and Pharmaceutical Materials (JACPM), vol. 2, no. 3, pp. 152–156, 2019, [Online]. Available:
http://advchempharm.ir/journal/index.php/JACPM/article/view/121%0Ahttp://advchempharm.ir/journal/index.php/JACPM/article
/download/121/258.
[6] S. Cho et al., “Knowledge and behavior regarding cosmetics in Koreans visiting dermatology clinics,” Annals of Dermatology,
vol. 29, no. 2, pp. 180–186, 2017, doi: 10.5021/ad.2017.29.2.180.
[7] S. Minaee, N. Kalchbrenner, E. Cambria, N. Nikzad, M. Chenaghlu, and J. Gao, “Deep Learning Based Text Classification: A
Comprehensive Review,” 2020, [Online]. Available: http://arxiv.org/abs/2004.03705.
[8] A. Arami, A. Poulakakis-Daktylidis, Y. F. Tai, and E. Burdet, “Prediction of Gait Freezing in Parkinsonian Patients: A Binary
Classification Augmented With Time Series Prediction,” IEEE transactions on neural systems and rehabilitation engineering : a
publication of the IEEE Engineering in Medicine and Biology Society, vol. 27, no. 9, pp. 1909–1919, 2019, doi:
10.1109/TNSRE.2019.2933626.
[9] L. Tang, Y. Tian, and P. M. Pardalos, “A novel perspective on multiclass classification: Regular simplex support vector
machine,” Information Sciences, vol. 480, pp. 324–338, 2019, doi: 10.1016/j.ins.2018.12.026.
[10] S. A. Wulandari, W. A. Prasetyanto, and M. D. Kurniatie, “Classification of Normal , Oily and Dry Skin Types Using a 4-

Comput Sci Inf Technol, Vol. 3, No. 3, November 2022: 137-147

Comput Sci Inf Technol ISSN: 2722-3221  147

Connectivity and 8-Connectivity Region Properties Based on Average Characteristics of Bound,” Jurnal Transformatika, vol. 17,
no. 01, pp. 78–87, 2019, [Online]. Available: journals.usm.ac.id/index.php/transformatika.
[11] J. Zheng and L. Zheng, “A Hybrid Bidirectional Recurrent Convolutional Neural Network Attention-Based Model for Text
Classification,” IEEE Access, vol. 7, pp. 106673–106685, 2019, doi: 10.1109/ACCESS.2019.2932619.
[12] R. L. Abduljabbar, H. Dia, and P. W. Tsai, “Unidirectional and bidirectional LSTM models for short-term traffic prediction,”
Journal of Advanced Transportation, vol. 2021, 2021, doi: 10.1155/2021/5589075.
[13] R. Diouf, E. N. Sarr, O. Sall, B. Birregah, M. Bousso, and S. N. Mbaye, “Web Scraping: State-of-the-Art and Areas of
Application,” Proceedings - 2019 IEEE International Conference on Big Data, Big Data 2019, pp. 6040–6042, 2019, doi:
10.1109/BigData47090.2019.9005594.
[14] S. Hara, A. Nitanda, and T. Maehara, “Data cleansing for models trained with SGD,” Advances in Neural Information Processing
Systems, vol. 32, 2019.
[15] M. A. Rosid, A. S. Fitrani, I. R. I. Astutik, N. I. Mulloh, and H. A. Gozali, “Improving Text Preprocessing for Student Complaint
Document Classification Using Sastrawi,” IOP Conference Series: Materials Science and Engineering, vol. 874, no. 1, 2020, doi:
10.1088/1757-899X/874/1/012017.
[16] F. Mohammad, “Is preprocessing of text really worth your time for toxic comment classification?,” 2018 World Congress in
Computer Science, Computer Engineering and Applied Computing, CSCE 2018 - Proceedings of the 2018 International
Conference on Artificial Intelligence, ICAI 2018, pp. 447–453, 2018.
[17] I. Boban, A. Doko, and S. Gotovac, “Sentence retrieval using Stemming and Lemmatization with different length of the queries,”
Advances in Science, Technology and Engineering Systems, vol. 5, no. 3, pp. 349–354, 2020, doi: 10.25046/aj050345.
[18] X. Zhaok et al., “AutoEmb: Automated Embedding Dimensionality Search in Streaming Recommendations,” Proceedings - IEEE
International Conference on Data Mining, ICDM, vol. 2021-December, pp. 896–905, 2021, doi:
10.1109/ICDM51629.2021.00101.
[19] D. López-Sánchez, J. R. Herrero, A. G. Arrieta, and J. M. Corchado, “Hybridizing metric learning and case-based reasoning for
adaptable clickbait detection,” Applied Intelligence, vol. 48, no. 9, pp. 2967–2982, 2018, doi: 10.1007/s10489-017-1109-7.
[20] G. Cheng, V. Peddinti, D. Povey, V. Manohar, S. Khudanpur, and Y. Yan, “An exploration of dropout with LSTMs,”
Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, vol. 2017-
August, pp. 1586–1590, 2017, doi: 10.21437/Interspeech.2017-129.
[21] D. Fitrianah and R. N. Jauhari, “Extractive text summarization for scientific journal articles using long short-term memory and
gated recurrent units,” Bulletin of Electrical Engineering and Informatics, vol. 11, no. 1, pp. 150–157, 2022, doi:
10.11591/eei.v11i1.3278.
[22] A. Dutta, S. Kumar, and M. Basu, “A Gated Recurrent Unit Approach to Bitcoin Price Prediction,” Journal of Risk and Financial
Management, vol. 13, no. 2, p. 23, 2020, doi: 10.3390/jrfm13020023.
[23] F. Ramzan et al., “A Deep Learning Approach for Automated Diagnosis and Multi-Class Classification of Alzheimer’s Disease
Stages Using Resting-State fMRI and Residual Neural Networks,” Journal of Medical Systems, vol. 44, no. 2, 2020, doi:
10.1007/s10916-019-1475-2.
[24] Y. Ho and S. Wookey, “The Real-World-Weight Cross-Entropy Loss Function: Modeling the Costs of Mislabeling,” IEEE
Access, vol. 8, pp. 4806–4813, 2020, doi: 10.1109/ACCESS.2019.2962617.
[25] J. Xu, Y. Zhang, and D. Miao, “Three-way confusion matrix for classification: A measure driven view,” Information Sciences,
vol. 507, pp. 772–794, 2020, doi: 10.1016/j.ins.2019.06.064.
[26] S. Shin, Y. Lee, M. Kim, J. Park, S. Lee, and K. Min, “Deep neural network model with Bayesian hyperparameter optimization
for prediction of NOx at transient conditions in a diesel engine,” Engineering Applications of Artificial Intelligence, vol. 94, 2020,
doi: 10.1016/j.engappai.2020.103761.

BIOGRAPHIES OF AUTHORS

Devi Fitrianah is working in Bina Nusantara University as a lecturer in

Department of Computer Science, School of Computer Science. Her academic qualification is
Dr., S.Kom., M.T.I. As Data Scientist, she does research in Artificial Intelligence, Data
Mining, Machine Learning and Information Systems (Business Informatics). Recent
researches are in the field of Natural Language Processing for sentiment analysis and
Automated Text Summarization. She can be contacted at email: [email protected].

Andre Hangga Wangsa is Student in Universitas Mercu Buana who are pursuing
a bachelor's degree in Computer Science His interest are in Data Science fields such as Natural
Language Processing and Computer Vision. He can be contacted at email:
[email protected].

Text classification to predict skin concerns over skincare using bidirectional mechanism … (Devi Fitrianah)