2020 International Electronics Symposium (IES)

An Efficient Framework of Hybrid

Recommendation System based on Multi Mode
Yunaz Gilang Ramadhan
Tessy Badriyah Iwan Syarif
Politeknik Elektronika Negeri
Politeknik Elektronika Negeri Politeknik Elektronika Negeri
Surabaya, INDONESIA
Surabaya, INDONESIA Surabaya, INDONESIA
[email protected]
[email protected] [email protected]

Abstract— Recommendation systems have been widely product that they think is the best. The recommendation
applied in many areas, such as E-commerce, and so on. system is the right solution to solve the problem.
However, in some complex systems such as missed sparse data,
it will be increasingly difficult to build a model for user However, there is no method in the recommendation
recommendations. In this research we develop a system that is good to use at any time. There are several
recommendation system on E-Commerce. This system will be conditions in which a recommendation system method is good
able to adapt and provide the best recommendations for each to use than other recommended system methods. The reality
user dynamically even in sparse environment. The system will is not all E-Commerce has the same characteristics and the
be created in a web-based application to display the product possibility with the development of business carried out by an
recommendations to users. The recommendation system E-Commerce can also experience changes in characteristics.
developed is expected to be able to solve cold-start problem This will make a good recommendation system used before it
when there is no other relevant data to be recommended for the can turn out to be bad. For this reason, it is necessary to
new added product and also the sparsity problem. To overcome develop a recommendation system that is able to adapt
this problem, the system will implement multi-mode algorithm according to the characteristics of information contained in E-
that uses more than one search algorithm for the closest Commerce. The flexibility of the recommendation system will
characteristics in the recommendation system and can choose produce recommendations that are always good for each
one of the best algorithms to use in accordance with the existing situation. Research on the construction of a flexible
data and hybrid-filtering that can use a combination of
recommendation system is not only applicable in the case of
Collaborative Filtering is to make recommendations based on
information equations between users and Content-Based
product recommendations in E-Commerce but can also be
Filtering is to make recommendations based on information useful in developing a system of recommendations for other
representation of a content. Thus the system will be able to cases.
provide product recommendations on any state of data on E-
Commerce. II. THE PROPOSED METHOD
Keywords— Recommender system, sparsity, cold-start, hybrid In this study, we built a separate multi-mode method
filtering which is able to accept various kinds of data input (rating,
order, visit, & search). The proposed method called Multi-
Mode Hybrid Filtering (MMHF). Hybrid filtering method in
I. INTRODUCTION this research can formulate recommendations based on the
Technology is growing faster. Today we are entering a weighting average of the recommendations using both
digital age where many people who have access to the internet Collaborative and Content-Based Filtering. The weighting
and information are important in their daily lives. Realizing value is obtained from multiplying the calculation results from
this, business people began to adapt. In the past, business Association Retrieval Correlation (ARC) with the similarity
people carried out their business using conventional methods, value calculated by Content based filtering for each product.
however, today many business people use information The following will explain the system design of the research
technology to market their products or services. One of the that we did and the part that underlies the method we propose,
uses of information technology for business is E-Commerce. namely Association Retrieval Correlation and Content Based
E-Commerce itself is currently growing very rapidly because Filtering by Text -Mining on item's information.
of the ease of transaction offered. E-Commerce systems are
becoming increasingly complex because more products or A. A. System Design
services are offered, and more users are available. The The proposed system consists of data extraction features
complexity of information contained in this E-Commerce which are the pre-processing stage to obtain data and store
system will be very difficult to manage. Users will also find it important features on each data (Fig 1., number 1).
difficult to get the product or service he wants. Sometimes
many prospective buyers do not know what items to buy, they
will try to find information about what products are offered.
Problems arise because a lot of information must be
processed, imagine just to make a purchase, the user must do
a lot of page switching. Even using the search engine features
is not the solution, because search results will always be the
same for all users, even though users only want to buy the

978-1-7281-9530-8/20/$31.00 ©2020 IEEE 351

Authorized licensed use limited to: Carleton University. Downloaded on November 03,2020 at 04:57:33 UTC from IEEE Xplore. Restrictions apply.
 Figure 1. System Design

The data obtained is then stored in a database (shown in

number 2). When there is a request to compile
recommendations, the system will access information from
the database. This information will be processed by a module
that determines recommendations that will use the
Association Retrieval Correlation (ARC) as Collaborative
Filtering and will combine the results among products using Figure 2. Workflow of Multi Mode Hybrid Filtering (MMHF)
Recommendation System
the Content based Filtering method based on the product
features, The product features in this case are a combination
of text-mining and TF-IDF calculations (indicated by number
The recommendation system designed has priority in
3) . The process will produce a recommendation module that selecting input data and methods used.
automatically selects the data to be processed, and other
Hybrid Filtering parameters to process data and arrange The following is the priority sequence of input data
recommendations (shown in number 4). The selected starting from the lowest to the highest:
recommendation mode will then formulate recommendations a. Search History
(shown in number 5). The output in the form of
recommendations will then be obtained by the user (shown in Retrieving data entered from the user's search history, will
number 6). only be used when other data is not available. The filtering
method used when using this data is Content-Base
Filtering. Search for the most relevant products with users
III. TRACKING USER BEHAVIOUR ON MULTI-MODE HYBRID based on their search history.
FILTERING b. Visit History
Tracking user behavior aims to provide dynamic
Retrieving data entered from the user's visit history, will
recommendations even though user information is not
only be used when rating data and orders are not available.
obtained from rating or order data, even by applying this
The filtering method used is Hybrid Filtering, with
module, users who are not logged in are still able to get
comparison of order data or other user rating data.
recommendations according to their habits. Currently in our
design, there are two user habits that are recorded. Namely the c. Rating
history of search and history of user visit to the product
Retrieve data entered from the user rating history of the
User history is valuable information for the product. Used according to the tendency of users who
recommendation system. But not all information is always prefer to give a rating on the product. The filtering method
relevant. For example, for searches, we can find out the used is Hybrid Filtering with a comparison of other user
tendency of current users to prefer products like enough from rating data
the latest search history, the old history is no longer relevant.
To achieve what is needed, we apply FIFO (First In First Out) d. Order
to the variable that stores the user's search history Retrieve data entered from the user's purchase history.
The following diagram explains the workflow from the Used according to the tendency of users who prefer to buy
tracking process of user behavior in MMHF. products. The filtering method used is Hybrid Filtering
with comparison of other user's order data.
The above information will be processed to find
recommendations and after that the results of
recommendations can be evaluated

352

Authorized licensed use limited to: Carleton University. Downloaded on November 03,2020 at 04:57:33 UTC from IEEE Xplore. Restrictions apply.
 IV. RESULT AND DISCUSSION 5. F-measure :
In this study, several stages of testing were carried out to An evaluation calculation in an information
prove the superiority of the proposed hybrid filtering method retrieval that combines recall and precision. The
and see whether the recommendation system that had been recall value and precision in a situation can
designed could work well. The reference for assessing how have different weights. The measure that
well the results obtained is determined by the performance displays reciprocity between recall and
measurement used in this study precision is a F-measure which is the mean
harmonic weight of recall and precision [13]
A. Performance Measurement
There are several ways to determine the performance of a
recommendation system. Following are the methods used to (5)
calculate system recommendation performance in this study:
1. Precision : Where:
Is the ratio of the number of recommendations p is the value of precision
relevant to the total number of r is the recall value
recommendations.
B. Testing Scenario
(1)
Comparative data scenarios are recommendations based
Where: on scenarios compared to a user's relevant data at that time to
tp is the number of products found and relevant measure the performance recommendations given. This
fp is the number of products found but not scenario is divided into two types, namely strict and non-strict.
relevant Strict mode means that the recommendation data will only be
compared with the original data from the user. Missing
2. Recall : recommendations get data sources from user rating data.
Is the ratio of the number of recommendations Then the recommendations will be compared to products that
relevant to the total relevant data. have been rated by that user only. While the second scenario
in the form of non-strict mode means that when
(2) recommendations get data sources from rating data, the
Where: results of the recommendations will be compared to products
tp is the number of products found and relevant from the top 5 categories that are rated best by users.
fn is the number of products not found but Before discussing the results of the method, we will
relevant display data visualization to split two data entries from user
orders and ratings:
3. Fallout :
Is the ratio of the number of recommendations
given irrelevant to the total irrelevant data.

(3)

Where:
fp is the number of products found but not
relevant
is the number of irrelevant products

4. missRate
is the ratio of the number of relevant items but
is not recommended with the total relevance of
the relevant items.
Figure 3. Distribution of order data

(4)
Where:
fn is the number of products that are relevant but
not recommended
is the recommended and relevant number of
products

353

Authorized licensed use limited to: Carleton University. Downloaded on November 03,2020 at 04:57:33 UTC from IEEE Xplore. Restrictions apply.
 Figure 5. User Visit Evaluation
Figure 4. Distribution of rating data

time
From the data distribution in Figure 3 and 4, it can be seen 50
40
that the order data is more sparse. This more sparse condition 30
will certainly affect the value of confidence. The more tightly 20
the value of confidence will be higher too, and the easier it is 10
to recommend other products for users 0

C. Performance Analysis
In this section, an experiment is conducted to compare the
performance between the hybrid method and the collaborative
filtering method. The Hybrid Filtering method used is a Figure 6. Computation time on the User Visit
combination of ARC and Content-Base Filtering methods
From Fig 5 and Fig 6 , it can be seen that collaborative
The comparison scenarios are divided into two types, performance is superior to Hybrid in fallout, recall, missrate
namely strict and non-strict. Strict mode means that the and time execution. But the Hybrid method is superior to the
recommendation data will only be compared with the original precision and f-1-score. It can be concluded that the more
data from the user, for example the recommendation gets the collaborative weights on hybrid filtering, the longer the
data source from the user rating data, so the recommendation execution time will be.
results are only compared to products that have been rated by
(ii). Using product rating data:
the user. While the second scenario is a non-strict mode,
meaning that when the recommendation gets the data source Meanwhile, the evaluation of the rating data test will
from the rating data, the recommendation results will be compare which algorithm is the most reliable in handling less
compared with the products from the top 5 categories that are sparse data
rated the best by the user.
The hybrid filtering module used has adjustable
parameters so that it can provide flexibility to determine the
number of recommendations. For example, with the value of
LIMIT_COLLABORATIVE = 100 and
LIMIT_CONTENT_BASE = 10, the maximum
recommendation obtained is 100 x 10 = 1000 products (this
number will later decrease because the system will ignore the
same recommended product).
In the experiment, we use three methods, namely
collaborative filtering method, hybridA with collaborative
weight 100 and content-base 10 weight, while hybridB with
collaborative weight 50 and content-base weight 10. All of
these scenarios are run on product visit history data and rating
Figure 7. Product Rating Data Evaluation
data given by the customer to the product.
(i). Using product visit data:
Experiments on product visit history data using new users
who have only visited one product. And will be compared the
results of recommendations from which algorithm can
recommend more products.

354

Authorized licensed use limited to: Carleton University. Downloaded on November 03,2020 at 04:57:33 UTC from IEEE Xplore. Restrictions apply.
 Intelligent Information Technology and Security Informatics, pp.
time 530-533, 2010.
40
35 [4] I. H. Witten, “Text Mining,” University of Waikato, Hamilton,
30 2002.
25
20
15 [5] E. T. W. I. S. Tessy Badriyah, “A Hybrid Recommendation System
10 for E-Commerce based on Product Description and User Profile,”
5
0 2015.

[6] F. R. A. V. Tariq Mahmood, “Improving Recommendation

Effectiveness by Adapting the Dialogue Strategy in Online Travel
Planning,” ECTRL Solutions, Trento, 2009.

[7] S. Robertson, "Understanding Inverse Document Frequency: On

Figure 8. Computation time on Product Rating Data
theoretical arguments for IDF," Journal of Documentation, vol. 60,
no. 5, pp. 503-520.
As previous experiments, collaborative has good recall
values, fallout, and miss rate. But collaborative is not
[8] V. S. Prem Melville, “Recommender Systems,” Encyclopedia of
consistent in strict and non-strict modes. While hybrids are
Machine Learning, pp. 829-838, 2010.
more consistent with results. Besides that, it was found, the
collaborative algorithm is very wasteful of time for large [9] “Recommender system,” wikipedia.org, 31 December 2017.
amounts of data (such as rating data) [Online]. Available:
https://en.wikipedia.org/wiki/Recommender_systeHym. [Diakses
V. CONCLUSION 23 January 2018].

This study proposes an efficient Hybrid Recommendation [10] M. T. W. L. Gunnar Schröder, "Setting Goals and Choosing
System based on Multi Mode. The proposed method called Metrics for Recommender System Evaluations," European Social
Multi Mode Hybrid Filtering (MMHF) is based on user
Fund and the Free State of Saxony, Dresden.
behavior tracking, which is about history of searching and
history of user visits to products. In the experiment of the [11] C. W. M. X. X. G. YiBo Chen, “Solving the Sparsity Problem in
performance of features, the track module of user behavior has
Recommender Systems Using Association Retrieval,” Journal of
succeeded in efficiently monitoring user habits and
Computers, vol. 6, no. 9, pp. 1896-1902, 2011.
formulating recommendations based on these data.
From the results of the experiment, it was found that the [12] F. O. A. H. J. B. Jesus Bobadilla, “A Collaborative Filtering
system can generate tags / features automatically through the Approach To Mitigate The New User Cold Start Problem,”
text mining method and similarity calculations using TF-IDF Knowledge-Based Systems, p. 225, 2012.
smooth and with excellent content-based quality.
[13] Y. C. Y. J. M. S. Tan Yang, “A New Algorithm For Multi Mode
The track user behavior module has succeeded in
Recommendation In Social Tagging Systems,” 2012.
overcoming the cold start problem in the system. To improve
the performance of the feature, new information can be added
[14] J. T. X. Z. X. X. Mi Zhang, “Addressing Cold Start in
which is monitored and integrated into the recommendation
module. Besides that, the slow computation time is an obstacle Recommender Systems: A Semi-supervised Co-training
that arises. Further research is needed on architectural design Algorithm,” SIGIR, 2014.
or system features in order to speed up the process of making
recommendations. [15] A. M. Hridya Sobhanam, “Addressing cold start problem in
recommender systems using association rules and clustering
technique,” dalam International Conference on Computer
Communication and Informatics, Coimbatore, INDIA, 2013.
REFERENCES

[1] H. R. V. Paul Resnick, “Recommender System,” Communications

of the ACM, vol. 40, no. 3, pp. 56-58, 1997.

[2] V. Beal, “What is Recommender System?,” www.webopedia.com,

2013. [Online]. Available:
http://www.webopedia.com/TERM/R/recommender_systems.htm
l. [Diakses 6 6 2017].

[3] M. C. J. H. Zhao Qiu, “Design of Multi-mode E-commerce

Recommendation System,” Third International Symposium on

355

Authorized licensed use limited to: Carleton University. Downloaded on November 03,2020 at 04:57:33 UTC from IEEE Xplore. Restrictions apply.