You are on page 1of 10

Skin Care Product Recommendation System

Using Android-Based Hybrid Method (Content


Based Filtering and Collaborative Filtering)
Victor Amrizal1, Hendra Bayu Suseno2, Arini3, Handan Siswaningrum4
Teknik Informatika. Universitas Islam Negeri Syarif Hidayatullah.Jakarta, Indonesia
victor.amrizal@uinjkt.ac.id, hendra.bayu@uinjkt.ac.id, arini@uinjkt.ac.id, handan.siswaningrum15@mhs.uinjkt.ac.id

Abstract— The development of skin care products in facial The production of skin care products in Indonesia itself is
skin care in Indonesia can be proven by data on the market value increasing along with the increasing public interest in the need
of the cosmetic industry from 2010 to 2023. The largest market of for using skin care. Based on a survey conducted in 2018 by
the Indonesian cosmetic industry is the skin care segment with a (“Beauty and Personal Care in Indonesia | Euromonitor,”
market volume of USD 1,673 million. The number of skin care 2021) with the title “Mass Beauty and Personal Care in
products is a consideration for someone in choosing a product. Indonesia”, stated that the growth of the cosmetic industry in
The recommendation system can help provide product Indonesia continues to increase with the highest sales being in
recommendations. Some of the recommendation system methods hair care products. by 35.8% and followed by skin care
include Content Based Filtering, Collaborative Filtering and
products by 31.7%. These data indicate that Indonesia has
Hybrid. However, this method has weaknesses if applied
dominantly. Solutions using the Hybrid method by combining the
potential in the development of the cosmetic industry, both in
two, can help improve the system. The purpose of this study is to terms of products and users [3]. With many companies
implement Content Based Filtering and Collaborative Filtering producing skin care products with different brands, prices and
methods on an Android-based skin care product recommendation quality, it becomes a difficult consideration for a consumer in
system. System testing is carried out using Precision, MAE, and choosing an appropriate product [5].
Blackbox Testing techniques. The test was conducted on 20 skin For this reason, the recommendation system can assist
care users. The testing technique with the Precision of both users in providing product recommendations from the many
methods, obtained the results of Content Based Filtering 0.80 and products available. A recommendation system is a system that
Collaborative 0.88. MAE testing was carried out on the provides and recommends an item or goods to assist users in
Collaborative Filtering method with 3 scenarios getting results of making a decision [6]. There are several recommendation
0.96 in scenario 1 (3 users), and 0.92 in scenario 2 (5 users), then system methods commonly used in making the system,
0.88 in scenario 3 (20 users). So it can be concluded that the including Content Based Filtering, Collaborative Filtering and
method applied has the best accurate value and error rate on the Hybrid [7]. However, each of these methods has weaknesses
results of more recommendations. when applied dominantly. In this case, if the system only uses
Keywords— Recommended System, Skin care, TF-IDF, Pearson the Content Based Filtering method, the system only
Correlation, Content Based Filtering, Collaborative Filtering, recommends items based on user characteristics so that the
Hybrid Recommended System
recommendations displayed for users are not broad. While the
I. INTRODUCTION Collaborative Filtering method only provides
Along with the times, people are required to be more recommendations based on ratings from other users, so it is
modern in terms of appearance. The need for appearance has not certain that the items to be recommended will be in
become a top priority for both women and men. In this case, accordance with the criteria according to the user. The
appearance is not only about clothes or accessories, but can be solution using the Hybrid method by combining the two, can
seen from the face [1]. help improve performance on a recommendation system that
Facial skin is a very sensitive part. Compared to the skin only uses the dominant method [8].
of other parts of the body, facial skin has a very thin layer, so Several studies on skin care recommendation systems
it is not uncommon for many people to experience problems have been widely applied. As in the research conducted by
with facial skin. Problems with facial skin usually occur (Safitri, 2021) who implemented a skin care recommendation
because someone is lazy to care for and keep the skin healthy. system using the Content Based Filtering method with the
Facial skin care or commonly referred to as skin care is an Apriori Algorithm which can produce new product
action that can be taken to maintain skin condition [2]. recommendations based on the suitability of products that
Skin care is an activity carried out in treating or improving users have liked. In this case, there are still many studies that
the condition of facial skin. Skin care is done by using a have not carried out the application by combining the two
variety of supporting products that contain ingredients methods (hybrid) [9].
according to skin conditions. These products will later be Thus, the authors conclude to conduct research by
applied to the maintenance routine [3].
implementing the Hybrid method using two recommendation
Before taking care of facial skin, it's good someone needs
methods, namely Content Based Filtering and Collaborative
to know the condition of their skin. This is very important for
someone in choosing skin care products that suit their skin Filtering on an Android-based skin care product
conditions. Facial skin types are generally divided into four recommendation system.
types, namely dry, normal, oily and sensitive. Each product II. TEORETICAL FOUNDATION
has a different content and function in its use according to their A. Recommended System
respective needs [4]. The recommendation system is a system that aims to help
provide recommendations for users to determine an
interesting item or product. In other words, the system is

1
designed to provide predictions based on user preferences and Filtering by utilizing item ratings based on other users'
ratings given by any user. In recent years, the recommendation ratings, so that the performance of the recommendations
system has been widely implemented in several applications. produced is more widespread, but this method has several
Examples of implementing the concept of a well-known drawbacks, including:
recommendation system, such as recommendations for films, a. This method cannot provide recommendations if
books, news, food menus, products, matchmaking, or there is no value assigned to an item (cold-start-
friendship searches on Facebook or Twitter [10]. problem).
Everyone has different tastes or interests in choosing an b. Limitation problems or sparsity problems on new
item or product. In this case, the recommendation system can
items that have not yet received a rating, so there is a
work by providing certain suggestions to the user in choosing
possibility that the item cannot be recommended.
an item that might be of interest to the user. By using certain
techniques or methods that can produce a recommendation, it
can be based on the opinions of other users or through D. Hybrid
calculating the similarity of characteristics of an item that was Hybrid Recommender is a method of recommendation
liked by previous users. Therefore, a recommendation system system by combining two or more recommendation
was built to provide suggestions based on the prediction of an techniques to improve recommendation performance. The
item that suits the tastes of the user. main idea of this method is to cover the shortcomings of the
To run a recommendation system there are several single method. The techniques used in Hybrid
methods used to develop a system. In this case, there are three recommendation filtering are as follows [17]:
approaches to recommendation system methods commonly a. Weighted: the results of each recommendation
used, such as Content Based Filtering, Collaborative Filtering, technique are combined and processed to produce
and Hybrid Filtering [11]. the best recommendation.
b. Switching: this system can choose, switch or switch
B. Content Based Filtering recommendation techniques according to the
A content-based recommendation system is a situation. It's complex, because you have to
recommendation method with a technique of equating an item determine the switching criteria first. So different
attribute based on user preferences and recommending a new methods will produce different recommendations.
item based on similar characteristics or information contained c. Mixed: technique recommendations from various
in the item [12]. methods are displayed displaying them
Content Based Filtering recommends items with simultaneously.
attributes similar to what the user is currently viewing or likes d. Feature Combination: the result of one
and recommends them based on the item's information. In recommendation technique becomes the input of
other words, this method will extract the information another recommendation technique.
contained in the item, then it will be compared with items that e. Feature Augmentatio: features from different
have been seen or liked by the user [11]. recommendation data sources are combined
However, this content-based recommendation system has together into a single recommendation algorithm.
advantages and disadvantages, namely [13]: f. Cascade: one recommendation processes the
Advantages: recommendations provided by another.
a. The system performs calculations based on the g. Meta-level: the model is studied by one
similarity between user profiles and item profiles. recommendation that is used as input to another.
b. In this system, users do not require interaction with
other users (User independent), so they do not E. Algoritma TF-IDF
experience sparsity problems (data voids). The TF-IDF (term frequency-inverse document
Deficiency: frequency) algorithm can be used as a user profile to look for
a. Can only be used on attributes that match the item. similarities in keywords. This algorithm will calculate the
b. Limited recommendations are given, because they Term Frequency (TF) which is the frequency of occurrence
only recommend similar items, so there is no chance of a word in a document and Inverse Document Frequency
to get unexpected item recommendations. (IDF) the frequency of the document containing the word
[18].
C. Collaborative Filtering The TF-IDF value is obtained from the calculation of TF
The basic concept in the Collaborative Filtering method which symbolizes the frequency of subject terms appearing
is to utilize the opinions of other users to provide predictions in the document, and IDF which symbolizes the frequency of
on an item that may be of interest to a user [14]. The purpose documents containing subject terms in items [19]. The
of the Collaborative Filtering method itself is to provide equation formula in calculating TF-IDF can be explained in
suggestions for new items or to predict certain items for a user the following equation:
based on the preferences of previous users and opinions from 𝑖𝑑𝑓𝑖 = 𝑙𝑜𝑔 ( 𝑛/𝑑𝑓𝑖 ) (1)
other similar users [15].
In general, the recommendation process in this 𝑊 = 𝑇𝐹 𝑥 (𝐼𝐷𝐹 + 1) (2)
Collaborative Filtering method consists of two steps, namely Information:
calculating the similarity between items and predicting the 𝑊: the weight of the d-th document against the t-word
rating on the item [16]. 𝑡𝑓: multiple words searched in a document
The Collaborative Filtering method can cover the 𝐼𝐷𝐹: Inversed Document Frequency
shortcomings of the previous method, namely Content Based 𝑁: total number of documents

2
F. Algoritma Pearson Correlation Based Similarity to observe what brands and types of products are
The Pearson Correlation Based Similarity algorithm is being produced in the market. Then record the
one of the most widely applied similarity calculation methods product information on the website which the
in the Collaborative Filtering method. This algorithm uses researcher will use as a data set in this study. The
calculations based on correlations. Pearson correlation information is in the form of a description, category,
assesses how big the linear relationship between two price, name, brand, and criteria for the product.
variables. Pearson correlation coefficient is derived from a c) Questionnaire
linear regression model which has the assumption that the In the method of collecting data with a questionnaire,
it is used to strengthen the existing data on the
relationship between two variables must be linear, with errors
background and on the test data. The questionnaire
must be independent and have a probability distribution with
was conducted by distributing a form containing a
a mean of 0 and a variance [20]. list of questions and statements to respondents.
One of the correlation measurement techniques is to use
the Pearson Product Moment Correlation or commonly B. Sampling Method
abbreviated as Pearson Correlation. The results of the In determining the number of samples in this study the
Pearson Correlation measurement are in the form of authors used probability sampling method (random sampling).
similarity criteria classification with a value range of -1 (very Probability sampling is a sampling technique that can provide
dissimilar) to 1 (very similar) with 0 as the middle value equal opportunities for each member in a population [22]. In
(neutral). The following is the equation of the Pearson probability sampling technique there are several methods. In
Correlation Based Similarity method [13]. this research, the writer uses stratified random sampling
∑ 𝑢∈𝑈(𝑅𝑢,𝑘 −𝑅̅𝑘 )(𝑅𝑢,𝑙 −𝑅̅𝑙 ) method.
𝑠𝑖𝑚 (𝑘, 𝑙) = (3)
2 2
√∑ 𝑢∈𝑈(𝑅𝑢,𝑘 −𝑅̅𝑘 ) √∑ 𝑢∈𝑈(𝑅𝑢,𝑙 −𝑅̅𝑙 ) a) Stratified Random Sampling
Information: Stratified random sampling method is a method of
𝑠𝑖𝑚 (𝑘, 𝑙): item-k and item-l . similarity values selecting samples by dividing the population into
𝑅𝑢,𝑘 & 𝑅𝑢,𝑙 : rating by User u on item-k and item-l homogeneous groups known as stratum, then the sample is
taken randomly in each stratum and an estimate of the
𝑅̅ 𝑘 & 𝑅̅ 𝑙: average rating on k-item and l-item.
sample representing the stratum is made.
In this study, the groups were divided into several skin
G. Algoritma Weighted Sum types. According to (Riani S & Setyowati, 2014) skin
Prediction calculation method with Weighted Sum, is types are generally divided into 4 parts, namely, normal,
used to calculate the predicted rating value on new items. dry, oily and sensitive [4]. For that the author makes the
This algorithm calculates the rating predictions on these items population in determining the sample is a population that
by comparing the ratings of the items that have been given by has these 4 skin types.
the user with the similarity results between these items and b) Population
other items [21]. Here's the formula for the Weighted Sum In the Skin care Product Recommendation System
algorithm: research, the population is skin care users with several
∑𝑖∈𝑗 𝑟𝑢,𝑖 ∗𝑠𝑖𝑚(𝑖,𝑗) different skin types.
𝑃(𝑢, 𝑗) = ∑𝑖∈𝑗 𝑠𝑖𝑚(𝑖,𝑗)
(4)
c) Sample
Information: Sampling according to (Arikunto, 2011) is if the subject is
𝑃(𝑢, 𝑝): Prediction for User u in item j. less than 100 people, then all should be taken. If the subject
∑𝑖∈𝐼 : The set of items similar to item j. is large or more than 100 people, it can be taken 10-15%
𝑟𝑢,𝑖 : User Rating u on item i. or 20-25% or more [23].
𝑠𝑖𝑚(𝑖, 𝑗): The similarity value between item i and item j. In this study, the authors took 20% of the total population with
a sample of 20 people. The samples in this study were divided
III. RESEARCH METHODS into several groups, namely groups with normal skin types,
A. Method Of Collecting Data groups with oily skin types, groups with dry skin types, and
In compiling a study, data collection techniques are useful groups with sensitive skin types. Samples taken from each
for analyzing data related to the research. Researchers need group were calculated using the Slovin formula and the
data and information in order to find background problems following results were obtained:
and theories on methods related to the problem being Tabel 2. Data Sampel
discussed. In this case, the researcher used several techniques,
Jumlah Penghitungan
namely: No. Kelompok
Anggota Sampel
Hasil
a) Literature Study 25
1. Normal 25 𝑥 20 5
In this study the authors collect data relating to the 100
30
object being discussed. By reading and studying 2. Berminyak 30 𝑥 20 6
100
various books, journals, and theses related to 20
3. Kering 20 𝑥 20 4
research. And also collect data digitally via the 100
25
internet by looking for related websites to help 4. Sensitif 25 𝑥 20 5
100
support the discussion in this research. References Total 20
can be seen on the Bibliography page.
b) Observation
C. System Development Method
In observation, the author makes observations of the
In developing a system, researchers need a method that
existing systems on several skin care websites such
as Sociolla.com, Beautyhaul.com, and Sephora.com can explain the stages of research so that it can be understood
easily. One of the system development models that I will use

3
is Rapid Application Development (RAD). The purpose of In this study, the data that the author has collected is the
this model selection is that the system being developed is in data that the author will use as material to implement the
accordance with the needs and desires of the user. There are system.
several stages in the implementation, as follows: The product data that the author took from the results of
A. Needs Planning Stage observations or observations were as many as 60 products.
At this stage the author divides into several parts, including: Product information consists of 5 categories (facial wash,
1. Problem Analysis toner, moisturizer, sunscreen and mask), 4 skin types
At this stage, an analysis of the problem is carried (normal, dry, oily and sensitive) and 27 brands. And other
out which is the reason for making the system. information for additional reference such as name, price,
2. Data Needs Analysis description and product image.
At this stage the author analyzes what data needs are Then for user or user data, the author has selected 20 users
needed in the system to be made. as the test sample that the author has obtained using the
3. System Requirements Analysis sampling method. Each user has done a rating on each
At this stage, an analysis of the software and product they have used. So that obtained 363 rating items that
hardware specifications needed in making the can be used as a reference in the study.
system is carried out.
IV. ANALYSIS AND DESIGN
B. Stage Design System The Hybrid-based skin care recommendation system was
At this stage, the system design process is carried out to built using two methods in one system, which in this method
determine the suitability between the system and the user. uses the Switching technique in implementing the
1. Process Design recommendations. The Switching technique will provide
At this stage, the algorithm calculation process is recommendation results by choosing one of the methods used
designed for the system to be created. By applying according to available needs, and will not display other
the Hybrid method (Content Based Filtering and methods.
Collaborative Filtering) with Switching technique.
2. UML Design
At this stage, it will be explained what processes will
be run by the system. Built using UML diagrams,
such as Use Case diagrams, Activity diagrams,
Class diagrams, and Sequence diagrams.
3. Database Design
Perform database design that contains Design data
tables needed in data processing which will later be
implemented into the database. The database that
will be used is Realm.
4. User Interface Design Gambar 4. Alur Proses
This stage is designing the appearance of the system
that will be built in order to connect the user with the A. Content Based Filtering Method Process
system. The stages in the calculation process in the Content Based
5. Coding Filtering method are as follows:
This stage is the result of the design into the system. 1. Define queries.
Coding is done using a language that the system can 2. Compare item attributes
understand and is represented in the form of a 3. Calculation of the TF-IDF Algorithm.
4. The recommendation results are obtained from the
program. The software used is Android Studio, Java
highest-lowest weight value.
as the programming language and Realm as the In this stage, attributes or keywords that are considered in
database. determining the similarity criteria between products are skin
type and product category, which can be seen in the following
C. Implementation Phase table:
1. System implementation
Tabel 4.1 Atribut
At this stage the system is ready to operate, so it will
be known whether the system is in accordance with Kategori Produk Jenis Kulit (Skin Type)
Facial Wash Normal
the interface design that has been made.
Toner Kering
2. Testing
Moisturizer Berminyak
Furthermore, the authors conducted tests on the
Mask Sensitif
system to state the accuracy between manual
Sunscreen
calculations and system calculations. The testing
system will be calculated using Precision, Mean As an example of calculation, by taking 1 sample of skin
absolute error (MAE) and Blackbox Testing. care product data, which becomes a query from the counter,
recommendations will be sought according to the similarity
D. Research Data between the product attributes. In this case, the author will
take a sample data scenario using 20 skin care product data as
calculation data.

4
Tabel 4.2 Data Produk Kode : P1
Kode Brand Produk Nama Produk Kategori Skin Type Brand : COSRX
AC Collection Nama : AC Collection Calming Foam Cleanser
Facial
P1 COSRX Calming Foam Berminyak Kategori : Facial Wash
Wash Skin Type : Berminyak
Cleanser
Facial
P2 ELSHESKIN Oily Cleanser Wash Berminyak
Wash 2) Compare attributes or keywords
Acne Clarifying 2in1 Facial Tabel 4.3 Atribut Item Pereferensi Pengguna
P3 SAFI Berminyak
Cleanser Wash Kode Kategori Skin Type
White Expert 2in1 Facial
P4 SAFI Kering Q P1 Facial Wash Berminyak
Cleanser & Toner Wash
P2 Facial Wash Berminyak
BIO BEAUTY
P5 Luxurious Face Oil Moisturizer Berminyak P3 Facial Wash Berminyak
LAB
P4 Facial Wash Kering
Hydra Boost Oil Free
P6 ELSHESKIN Moisturizer Berminyak P5 Moisturizer Berminyak
Moisturizer
P6 Moisturizer Berminyak
PONDS BRIGHT
P7 POND'S BEAUTY PERF Moisturizer Berminyak P7 Moisturizer Berminyak

CRM OILY P8 Moisturizer Kering


Hydrabio Gel-Creme P9 Mask Berminyak
P8 BIODERMA Moisturizer Kering
40 ml P10 Mask Berminyak
Cica Chocomint Clay P11 Mask Berminyak
P9 NPURE Mask Berminyak
Mask P12 Mask Kering
Charcoal Mineral
P13 Toner Sensitif
P10 SNP Black Ampoule Mask Berminyak
P14 Toner Kering
Mask
P15 Toner Normal
Mugwort Pore
P11 WHITELAB Mask Berminyak P16 Toner Normal
Clarifying Mask
P17 Sunscreen Kering
PHILOCALY
P12 Peel Off Mask Peach Mask Kering P18 Sunscreen Kering
SKIN
Noni Probiotics P19 Sunscreen Kering

P13 NPURE "Balance Me" Fine Toner Sensitif P20 Sunscreen Normal
Toner
Blueberry 3) Calculate TF-IDF
P14 INNISFREE Toner Kering
Rebalancing Skin The first step in calculating the TF-IDF algorithm is to
JARTE weight the word or term frequency (tf) between the query and
P15 CICA Care Toner Toner Normal other products.
BEAUTY
White Willow Toner Tabel 4. 4 Term Frequency (TF)
Essence - Centella Kode Kategori Skin Type
PURIVERA
P16 BHA Willow Bark Toner Normal Q P1 Facial Wash Berminyak
BOTANICALS
2% As Salicylic Acid
P2 1 1
Alternative
P3 1 1
Perfect Shield
P4 1 0
P17 ERHA Sunscreen Normal to Sunscreen Kering
P5 0 1
Dry Skin
P6 0 1
Ultra Sheer Dry -
NEUTROGEN P7 0 1
P18 Touch Sunscreen Sunscreen Kering
A P8 0 0
SPF 50+
P9 0 1
Bright Beauty Skin
P10 0 1
P19 POND'S Perfecting Cream Sunscreen Kering
P11 0 1
SPF30
P12 0 0
Hydrasoothe
P20 AZARINE Sunscreen Normal P13 0 0
Sunscreen Mist
P14 0 0
P15 0 0
1) Specifying Query
P16 0 0
The author will take the data scenario on product-1
P17 0 0
which is used as a query that will look for product
recommendations based on the calculation of the TF-IDF P18 0 0

algorithm. The profiles on product-1 are as follows: P19 0 0

5
P20 0 0 2. Prediction: calculate the predictive value of items
that will be recommended to the User by using the
The next step is to calculate the inverse document frequency Weight Sum algorithm.
(IDF). As an example of a counter, in this process the author will
provide a table of user scenarios and rating of skin care
Tabel 4. 5 Penghitungan IDF product items. The table contains 5 skin care products, and 5
Kode Kategori Skintype users who have rated items.
Q
P1 Facial Wash Berminyak Tabel 4.7 Rating Item
TF 1+1+1 1+1+1+1+1+1+1+1 Produk Produk Produk Produk Produk
Rata-rata
1 2 3 4 5
DF 3 8
𝑛 Nadhila 4 5 4 4,3
20 20
𝑖𝑑𝑓𝑖 = 𝑙𝑜𝑔 ( ) 𝑙𝑜𝑔 ( ) 𝑙𝑜𝑔 ( ) April 3 3 3
𝑑𝑓𝑖 3 8 Hafshah 3 3 4 3,3
IDF 0.82 0.39 Ulfa 3 4 4 3,6
IDF + 1 1.82 1.39
1) Similarity
Next is the calculation of the weight (W) on each product to The similarity process is the process of calculating the
find the best recommendation results. correlation value between items. In this study, the authors
calculated using an item-based approach, which means
Tabel 4. 6 Nilai Bobot (W) looking for similarities between 2 items that have been equally
Kode Kategori Skin Type rated by the user. In this study, similarity is used to calculate
Jumlah
Bobot the similarity of values between one product and another.
Q P1 Facial Wash Berminyak
For example, the researcher will calculate the correlation
P2 1.82 1.39 3.21
between Product-1 and Product-2 using the Pearson
P3 1.82 1.39 3.21 Correlation Based Similarity algorithm.
1.82 1.82 ∑ 𝑢 ∈ 𝑈(𝑅𝑢,𝑘 − 𝑅̅𝑘 )(𝑅𝑢,𝑙 − 𝑅̅𝑙 )
P4 0 𝑠𝑖𝑚 (𝑘, 𝑙) =
2 2
P5 0 1.39 1.39 √∑ 𝑢 ∈ 𝑈(𝑅𝑢,𝑘 − 𝑅̅𝑘 ) √∑ 𝑢 ∈ 𝑈(𝑅𝑢,𝑙 − 𝑅̅𝑙 )

P6 0 1.39 1.39 (3 − 3,3)(3 − 3,3) + (3 − 3,6)(4 − 3.6)


𝑠𝑖𝑚 (𝑃1, 𝑃2) =
P7 0 1.39 1.39 √(3 − 3,3)2+(3 − 3,6)2 𝑥 √(3 − 3,3)2 + (4 − 3,6)2
0 (−0,3)(−0,3) + (−0,6)(0,4)
P8 0 0 =
1.39 1.39 √(−0,3)2 +(−0,6)2 𝑥 √(−0,3)2 + (0,4)2
P9 0
0,09 + (−0,24)
P10 0 1.39 1.39 =
√(0,09) + (0,36)𝑥 √(0,09) + (0,16)
P11 0 1.39 1.39
0,09 + (−0,24)
=
P12 0 0 0 √ (0,45) 𝑥 √(0,25)
P13 0 0 0 0,09 + (−0,24)
=
P14 0 0 0 0,67𝑥 0,5
−0,15
P15 0 0 0 = = −0,44
0,33
P16 0 0 0
P17 0 0 0 After calculating the correlation between items using the
P18 0 0 0
Pearson Correlation algorithm, the similarity results are
obtained. The results of the Pearson Correlation measurement
P19 0 0 0
are in the form of similarity criteria classification with a value
P20 0 0 0 range of -1 (very dissimilar) to 1 (very similar) with 0 as the
middle value (neutral).
4) Recommendation Result Tabel 4.7 Hasil Similarity
Based on the table, the recommendations can be sorted
Produk 1 Produk 2 Produk 3 Produk 4 Produk 5
from the highest to the lowest values, namely the results
obtained are 3.21, 1.82, and 1.39. Produk 1 1 -0,44 -0,83 -1 1
From the calculation results above, the results for Produk 2 -0,44 1 1 -0,1 0
recommendations based on the Content Based Filtering Produk 3 -0,83 1 1 0 -1
method using the TF-IDF algorithm with the highest weight Produk 4 -1 -0,1 0 1 0
are found in P2 "ELSHESKIN Oily Cleanser Wash" and P3
Produk 5 1 0 -1 0 1
"SAFI Acne Clarifying 2in1 Cleanser" with a value of 3.21
each. So it can be concluded that the recommendations that
will appear first when the user chooses P1 are P2 and P3. 2) Prediction
In this scenario, the author will calculate the predictive
value of items that have not been rated by the User, which will
B. Collaborative Filtering Method Process
later be used as a recommendation item for the User.
The stages in this process are as follows:
Next, calculate the predictive value for Product 3 against
1. Similarity: find the similarity value between items
these two users. Here's how to calculate predictions using the
using the Pearson Correlation Based Similarity
Weighted Sum algorithm calculation.
algorithm.

6
Prediksi Produk-3 pada April: 3) Product View
∑𝑖∈𝑗 𝑟𝑢,𝑖 ∗ 𝑠𝑖𝑚(𝑖, 𝑗)
𝑃(𝑢, 𝑗) =
∑𝑖∈𝑗 𝑠𝑖𝑚(𝑖, 𝑗)
3𝑥1
𝑃(𝐴𝑝𝑟𝑖𝑙, 𝑃3) = =3
1
Prediksi Produk-3 pada Hafshah:
∑𝑖∈𝑗 𝑟𝑢,𝑖 ∗ 𝑠𝑖𝑚(𝑖, 𝑗)
𝑃(𝑢, 𝑗) =
∑𝑖∈𝑗 𝑠𝑖𝑚(𝑖, 𝑗)
3𝑥1
𝑃(𝐴𝑝𝑟𝑖𝑙, 𝑃3) = =3
1
After calculating the predictions using the Weighted Sum
equation, the results obtained are Product 3 recommendations
for April Users with a rating prediction of 3 and Hafshah Users
on Product 3 with a rating prediction of 3.

V. IMPLEMENTATION AND TEST RESULTS


A. Implementation Results Gambar 5. 2 Tampilan View Produk
The following are the results of implementing an 4) Display of Similar Product Recommendations
Android-based skin care product recommendation system.
1) Login View.

Gambar 5. 3 Tampilan Rekomendasi Produk Serupa


5) Product Rating Display
Gambar 5. 1 Login
2) Main Page View

Gambar 5. 4 Tampilan Rating Produk

Gambar 5. 1 Tampilan Halaman Utama

7
6) Rating Recommendation Display April 6 5 1 0,83
Hafshah 6 4 2 0,67
Ulfa 6 4 2 0,67
Miranda 6 6 0 1
Pinka 6 6 0 1
Arindiar 6 6 0 1
Ayu 6 5 1 0,83
Isma 6 4 2 0,67
Nichyta 6 6 0 1
Ika 6 4 2 0,67
Annisa 6 6 0 1
Addini 6 4 2 0,67
Rizka 6 6 0 1
Pramesti 6 2 4 0,33
Elvia 6 3 3 0,5
Chica 6 4 2 0,67

Gambar 5. 5 Tampilan Rekomendasi Rating Sabrina 6 6 0 1


Meida 6 6 0 1
7) History Ranny 6 6 0 1
Rata-rata 4,8 1,2 0,80

2. The results of the accuracy value in the


Collaborative Filtering method
Testing on this method is carried out after the user has
rated the product and received recommendation results
in the form of predicting the rating value. The result of
precision value is 0.88.
Tidak
User Rekomendasi Relevan Precision
Relevan
Nadhila 35 34 1 0,971
April 42 30 12 0,714
Hafshah 43 40 3 0,930
Ulfa 42 37 5 0,880
Miranda 43 39 4 0,906
Gambar 5. 6 Tampilan History
Pinka 42 38 4 0,904

B. System Test Result Arindiar 44 38 6 0,863

a) Accuracy Value Test Results with Precision Ayu 39 31 8 0,794

The precision ratio is the ratio between relevant Isma 40 38 2 0,95


documents and the number of documents found in the Nichyta 42 35 7 0,833
search. Precision is related to the system's ability not to Ika 42 32 10 0,761
call irrelevant documents [24]. Annisa 45 40 5 0,888
By testing the user acceptance test (UAT) by filling
Addini 40 37 3 0,925
out a questionnaire, and then the results will be used as
calculations for precision testing. Questionnaire in the Rizka 39 38 1 0,974

form of user responses to the results of the Pramesti 43 35 8 0,813


recommendations obtained. Whether the Elvia 40 35 5 0,875
recommendations generated by the system are relevant Chica 44 42 2 0,954
or not.
Sabrina 41 38 3 0,926
The test is divided into two parts, namely the
Meida 36 35 1 0,972
Content Based Filtering method and the Collaborative
Filtering method. The following results were obtained: Ranny 34 30 4 0,882
Rata-Rata 36,1 4,7 0,88
1. The results of the accuracy value in the Content
Based Filtering method b) Error Value Test Results with MAE
Testing on this method is done by looking at the results The error rate of the recommendation system is seen
of the recommendations of each product selected by the based on the mean absolute error (MAE), which is the
user. The result is a precision value of 0.80. average of the absolute errors. Where the error is the
difference between the actual rating value and the predicted
Tidak
User Rekomendasi Relevan
Relevan
Precision rating value [25].
Nadhila 6 3 3 0,5 In this test, it is only carried out at the stage of the
Collaborative Filtering method. MAE is calculated by

8
comparing the predicted rating results on the system with the c) System Test Results with Blackbox Testing
original rating made by the user. This test is carried out with At this stage, testing is carried out to observe the results
3 test scenarios. To find out whether the number of users of the input and output of the system. This test is carried out
affects the error rate in the system. If the value of the error at the end of making the system to find out whether the
rate on the MAE is getting smaller, the results of the system can function properly.
recommendations given are getting better. The following Testing is done by running thoroughly all the functions
results were obtained: in the program. To find out whether the system or application
1. Hasil uji 3 user is able to respond well, so that the resulting output is in
User Data Rating Jumlah Nilai Selisih MAE accordance with the needs.
Nadhila 15 14,45 0,963
April 22 15,36 0,698 VI. CONCLUSIONS & SUGGESTION
Hafshah 23 28,37 1,233 A. Conclusions
Rata-rata 0.96 Based on the discussion that the author has described, it
can be concluded that:
1. Hybrid method by utilizing two recommendation
2. Hasil uji 5 user
User Data Rating Jumlah Nilai Selisih MAE methods (Content Based Filtering and Collaborative
Nadhila 27 24,77 0.917
Filtering) used can provide good recommendation
performance for users who are able to cover the
April 34 22,38 0.658
shortcomings of each method.
Hafshah 35 38,97 1.113
2. The system is able to provide recommendations in
Ulfa 34 28,75 0.845
terms of product similarity and product assessment
Miranda 36 38,56 1.071 to other users.
Rata-rata 0.92 3. By using a hybrid switching technique combining
the Content-Based Filtering and Collaborative
3. Hasil uji 20 user Filtering methods, it is possible to produce
User Data Rating Jumlah Nilai Selisih MAE recommendations according to user needs.
Nadhila 35 29,37 0,839 4. The recommendation system with the Hybrid
April 42 25,45 0,605 method using these two methods is proven to
Hafshah 43 46,9 1,090 produce precision values with equally good
Ulfa 42 64,04 1,524
performance. In the Content Based Filtering method
Miranda 43 46,48 1,080
with a result of 0.80 and the Collaborative Filtering
method with a result of 0.88. In this case, the
Pinka 42 48.37 1,151
Collaborative Filtering method is more accurate in
Arindiar 44 44,87 1,019
providing skin care recommendations to the user.
Ayu 39 36,63 0,939 This is because the Collaborative Filtering method
Isma 40 23,92 0,598 is able to provide recommendations for more
Nichyta 42 39,97 0,951 products because the rating given is increasing.
Ika 42 42,03 1,00 5. And in the error rate test of the recommendations on
Annisa 45 29,48 0,655 the Collaborative Filtering method, it is proven to
Addini 40 43,95 1,098 produce a smaller error value if the
Rizka 39 31,06 0,796 recommendations given are increasing. By doing 3
Pramesti 43 26,37 0,613
test scenarios, with stages 3 users, 5 users and All
Users (20 users), the following values are obtained:
Elvia 40 24,74 0,618
• Scenario 1 (3 users) = 0.96
Chica 44 35,71 0,811
• Scenario 2 (5 users) = 0.92
Sabrina 41 25,56 0,623
• Scenario 3 (20 users) = 0.88
Meida 36 22,47 0,624
It can be seen that the more ratings given by the user,
Ranny 34 33,18 0.975
the more recommendations are generated. The
Rata-rata 0,88 smaller the error value generated, the better the
recommendations given by the user.
From the table of scenario results above, it can be seen 6. In the Content Based Filtering method, it is very
that, when the scenario test was carried out with 3 users, the useful in providing recommendations for new users
average MAE value was 0.96. And when done with 5 users who have never rated an item, and want to get
get 0.92 results. And when the test scenario is carried out on recommendations. This can solve the cold-start
all users, the results are 0.88. problem in the Collaborative Filtering method.
So it is known that when testing all users, the average
error value is the smallest compared to other scenarios. This B. Suggestion
is because the more users the more ratings are generated, the 1. For further research, it is expected to use a different
more recommendations will be given. algorithm in recommending an item to produce even
better recommendations.

9
2. For further research, it is hoped that it can be done System: Research Challenges. International Journal of Engineering
Trends and Technology (IJETT). Retrieved from
implicitly, that is, where the user does not realize he
http://www.ijettjournal.org
has provided input to the system. [13] Lubis, Y. I., Napitupulu, D. J., & Dharma, A. S. (2020).
Implementation of Hybrid Filtering (Collaborative and Content-based)
VII. BIBLIOGRAPHY Methods for the Tourism Recommendation System. 12th Conference
[1] Maarif, V., Nur, H. M., & Septianisa, T. A. (2019). Sistem Pendukung on Information Technology and Electrical Engineering, 6–8.
Keputusan Pemilihan Skincare Yang Sesuai Dengan Jenis Kulit Wajah [14] Ricci, F., Rokach, L., Shapira, B., Kantor, P. B., & Ricci, F. (2011).
Menggunakan Logika Fuzzy. EVOLUSI : Jurnal Sains Dan Recommender Systems Handbook. Recommender Systems Handbook.
Manajemen, 7(2), 73–80. https://doi.org/10.31294/evolusi.v7i2.6755 https://doi.org/10.1007/978-0-387-85820-3
[2] Widjaja, A. W. (2011). PENGEMBANGAN MODEL PERILAKU [15] Susilo, B., Setiawan, Y., & Pranata, I. (2018). Pendekatan Metode
KONSUMEN PRIA DALAM MEMBELI PRODUK PERAWATAN Collaborative Filtering pada Sistem Rekomendasi Pariwisata Kota
KULIT ( SKIN CARE ) DI INDONESIA, 16–30. Bengkulu, 1, 356–361.
[3] Kusumaningrum, S. D. (2021). Kajian pustaka dalam penentuan tipe [16] Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001). Item-based
dan permasalahan kulit wajah. Kajian Pustaka Dalam Penentuan Tipe collaborative filtering recommendation.pdf. Proceedings of the 10th
Dan Permasalahan Kulit Wajah, 1(1), 17–21. International Conference on World Wide Web (WWW’01), 285–295.
[4] Riani S, N., & Setyowati, E. (2014). Journal of Beauty and Beauty [17] Gangan, S., Pawde, K., Purbey, N., & Nair, P. S. (2014). Study of
Health Education, 3(1), 1–7. Recommender Systems Techniques, (10), 38–41.
[5] Cahya Purnomo, D., Yanti, M., & Widyassari, A. P. (2021). Pemilihan [18] Nurjannah, M., & Fitri Astuti, I. (2013). PENERAPAN ALGORITMA
Produk Skincare Remaja Milenial dengan Metode Simple Additive TERM FREQUENCY-INVERSE DOCUMENT FREQUENCY (TF-
Weighting (SAW). Jurnal Ilmiah Intech : Information Technology IDF) UNTUK TEXT MINING Mahasiswa S1 Program Studi Ilmu
Journal of UMUS, 3(01), 32–41. Komputer FMIPA Universitas Mulawarman Dosen Program Studi
[6] Ferio, G., Intan, R., & Rostianingsih, S. (2019). Sistem Rekomendasi Ilmu Komputer FMIPA Universitas Mulawarman. Jurnal Informatika
Mata Kuliah Pilihan Menggunakan Metode User Based Collaborative Mulawarman, 8(3), 110–113.
Filtering Berbasis Algoritma Adjusted Cosine Similarity. Jurnal Infra, [19] Nagarnaik, P., & Thomas, A. (2015). Survey on recommendation
7(1), 1–7. system methods. 2nd International Conference on Electronics and
[7] Prasetya, C. S. D. (2017). SISTEM REKOMENDASI PADA E- Communication Systems, ICECS 2015, (Icecs), 1496–1501.
COMMERCE MENGGUNAKAN K-NEAREST, 4(3), 194–200. https://doi.org/10.1109/ECS.2015.7124835
[8] Tommy, L., Novianto, D., & Japriadi, Y. S. (2020). Sistem [20] Li, Q., & Kim, B. M. (2003). Clustering approach for hybrid
Rekomendasi Hybrid untuk Pemesanan Hidangan Berdasarkan recommender system. Proceedings - IEEE/WIC International
Karakteristik dan Rating Hidangan. Journal of Applied Informatics and Conference on Web Intelligence, WI 2003, 33–38.
Computing, 4(2), 137–145. https://doi.org/10.30871/jaic.v4i2.2687 https://doi.org/10.1109/WI.2003.1241167
[9] Safitri, D. A. N., Halilintar, R., & Wahyuniar, L. S. (2021). Sistem [21] Adomavicius, G., & Kwon, Y. (2007). for Multicriteria. IEEE
Rekomendasi Skincare Menggunakan Metode Content-Based Filtering Intelligent Systems. Retrieved from www.computer.org/intelligent
dan Algoritma Apriori. Seminar Nasional Inovasi Teknologi (SEMNAS [22] Sugiyono, P. D. (2015). Metode penelitian dan pengembangan. Res.
INOTEK), 242–248. Retrieved from Dev. D, 2015, 39–41.
https://proceeding.unpkediri.ac.id/index.php/inotek/article/view/1136 [23] Hasugian, J. (2006). Penelusuran Informasi Ilmiah Secara Online:
[10] Jain, S., Grover, A., Thakur, P. S., & Choudhary, S. K. (2015). Trends, Perlakuan Terhadap Seorang Pencari Informasi Sebagai Real User.
problems and solutions of recommender system. International Pustaha, 2(1), 1-13–13.
Conference on Computing, Communication and Automation, ICCCA [24] Irfan, M., Dwi, A., & R., F. H. (2014). Sistem Rekomendasi: Buku
2015, 955–958. https://doi.org/10.1109/CCAA.2015.7148534 Online dengan Metode Collaborative Filtering. Implementation
[11] Malik, S., Rana, A., & Bansal, M. (2020). A survey of recommendation Science, 39(1), 1–15. Retrieved from
systems. Information Resources Management Journal, 33(4), 53–73. http://dx.doi.org/10.1016/j.biochi.2015.03.025%0Ahttp://dx.doi.org/1
https://doi.org/10.4018/IRMJ.2020100104 0.1038/nature10402%0Ahttp://dx.doi.org/10.1038/nature21059%0Ah
[12] Singh, L., Madia, N., Makvana, K., Alhijawi, B., Hanafi, M., Sharma, ttp://journal.stainkudus.ac.id/index.php/equilibrium/article/view/1268/
L., & Gera, A. (2013). A Survey of Recommendation System: 1127%0Ahttp://dx.doi.org/10.1038/nrmicro2577%0Ahttp://.
Research Challenges Related papers Survey on Recommendat ion Syst
em Using Semant ic Web Mining A Survey of Recommendation

10

You might also like