Fuzzy Systems and Data Mining IX 843

A.J. Tallón-Ballesteros and R. Beltrán-Barba (Eds.)

© 2023 The authors and IOS Press.
This article is published online with Open Access by IOS Press and distributed under the terms
of the Creative Commons Attribution Non-Commercial License 4.0 (CC BY-NC 4.0).
doi:10.3233/FAIA231096

Game Recommendation System

Man-Ching Yuen a,1 , Chi-Wai Yung a , Wing-Fat Cheng a ,
Hon-Pong Tsang b, Chi-Ho Kwan b , Chun-Lok Chan b, Po-Yi Li b
a
Department of Information Technology, Vocational Training Council, China
b
Department of Applied Data Science, Hong Kong Shue Yan University, China

Abstract. Recommendation systems are widely adopted in many areas to provide

better services to customers. As there are many games stored in online games
platforms, people may be confused when choosing or buying the game that is
suitable for them. Many game platforms would like to have a game recommendation
system, so that it can automatically recommend the right games to their customers.
However, there are a lot of difficulties in developing game recommendation systems.
First, it is difficult to collect and organize data on customers’ behavior. Second, the
user interface needs to be easier to use for customers, such as the charts displayed
that customers are interested in. In this paper, we have developed a games
recommender system with complete functional recommending features. By using
machine learning techniques and applying data visualization on our system, we build
a recommendation system that can showcase flexible outcomes with the same
element as the user input, which can give the user more choice when finding the
games they want.

Keywords. Recommendation system, game, machine learning, web scraping and

data visualization.

1. Introduction

Nowadays, there are loads of games platforms such as PS4 store, Steam, Xbox etc.
However, not all the games platforms have a built-in recommender system to help users
to find games they desire [1, 2, 3, 4, 5]. It is because some games platforms do not have
enough data visualization to show the current trend of games clearly. Furthermore, some
people think that most of the existing recommender systems can not satisfy the users.
They may face difficulties when trying to find a suitable game to play using the existing
recommender systems.
Based on these reasons, we want to build a recommendation system to users who
want to play games which suit their taste. We will investigate the relationship between
each of the categories and the analytics of the trend of the selected game platform. So,
we can find out the most popular games for each type of genre. By entering different
elements of the game that the user is interested in, the recommender system can list out
the game that the user may prefer.
Our goal of this project is trying to build a recommender system that is uniquely
different from other recommender systems. We hope to improve the accuracy, efficiency
to about 10% better than the existing recommender system. And to make our system

1
Corresponding Author, Man-Ching YUEN, Department of Applied Data Science. Hong Kong Shue
Yan University, Hong Kong, China; E-mail: mcyuen@hksyu.edu.
844 M.-C. Yuen et al. / Game Recommendation System

uniquely different from the other existing work, we are trying to develop a system to
have unique features like showing trends by using data visualization and can showcase
flexible outcomes with the same element as the user input rather than the other system
to have the same output every time. We may consider choosing one of the online game
platforms for datasets. We will scrap the data including the basic information, categories
and system requirements to develop our game recommendation system.
The organization of this paper is as follows. Section 2 presents the related work.
Section 3 describes our game recommendation system with the data analytical
framework. Section 4 shows the experimental result analysis. Section 5 draws out the
conclusion.

2. Related Work

There are a number of existing works. Content-based filtering and collaborative filtering
are usually used for recommendation systems [6, 7, 8]. In the following, we describe
some existing works and highlight the improvement in our game recommendation
system.
1) Games Finder - The website is one of the databases of online games (Games
Finder) [9]. You can click the game icon. Then, it will display other similar games.
However, it does not have any data visualization. You can input your favorite game
on the website. Our model also has the function to get recommendations which
use content-based filtering. Moreover, we have designed a dashboard to show
distribution of reviews in different genres.
2) Quantic Foundry - By entering 3 game titles, this website will show a list of games
that are similar with those you enter, and the platform you can buy the game from
[10]. However, the input items are quite simple and without data visualization. Our
system adopts data visualization techniques to enhance customer experiences.
3) Deep Visual Semantic Multimedia Recommendation Systems (D_VSMR) –The
proposed approach employs content-based techniques to expand users’ profiles
based on the visual content of games [11]. However, the features extracted by the
system might not be equally representative for all users.
4) Social-aware Contextualized Graph Neural Recommender System (SCGRec) – It
proposed using user personalized data (such as social connections) to improve the
game recommendation [12]. However, due to the high complexity, frequent update
on social media data in the system is not possible.

3. Our System

3.1. Overview

Our first step is to apply web scraping on STEAM (A video game digital distribution
platform), the scraped data will be stored in a .JSON file. Then, the data will be cleaned
and used for data visualization. After that, content-based filtering and collaborative
filtering may be used for designing the recommendation system. Finally, the Flask
application will be used for demonstration of our final prototype.
 M.-C. Yuen et al. / Game Recommendation System 845

3.2. System Architecture

Figure 1 shows the system architecture of our game recommendation system.

1. Build a model with data

¾ MongoDB Database - MongoDB is a source-available cross-platform document-
oriented database program. Classified as a NoSQL database program, MongoDB
uses JSON-like documents with optional schemas. MongoDB is developed by
MongoDB Inc. and licensed under the Server-Side Public License (SSPL).
¾ Machine Learning Algorithms (Content-based filtering) – We use both TF-IDF
and Cosine similarity in our system. We use Python to implement our system.
2. UI Client
¾ Web Application (Flask) - We use Flask API to run our recommendation system,
Data Visualization and the basic function of a client.
¾ Server

Figure 1. The system architecture of our game recommendation system.

846 M.-C. Yuen et al. / Game Recommendation System

3.3. Data Collection and Preparation

To acquiring data, we have created 3 programs to apply web scraping on the STEAM
platform to acquire data into datasets. Three of them are URLs of the game, game
information and comments from some games.

1. Scraping all the URLs of the video games. The input and the output are as shown in
Figures 2 and 3. Figure 2 shows the list of games on the webpage we want to scrap.
We need to scroll the page with “Page Down” button many times if we want to see
all the games. The code in Figure 3 shows we used the web driver from selenium to
scroll the page. Selenium is used mostly in this program. After that, we need to find
the URLs of all the games by XPath and store them into a list.

Figure 2. The input of scraping all the URLs of the video games.

Figure 3. The output of scraping all the URLs of the video games.

2. Scraping the information of the games. By reading the .pkl file, which have been
mentioned above, we have to scrap the information of the game that we need. Figure
4 shows a website of one of the games. Useful data for content-based filtering have
been scraped such as Game Title, Genre, User Tags, Overview and so on. A single
game data will be stored in a dictionary and all the data we acquire will be stored into
 M.-C. Yuen et al. / Game Recommendation System 847

“game_data” which is a list. All of the data is exported into a .JSON file
(output_2F.JSON). To speed up the process of scraping the game information, we
found that we can scrap the information by running eight programs at the same time.
This has benefits to update our datasets more frequently.

Figure 4: A website of one of the games with useful data for content-based filtering.

3. Scraping the comments of the games. The scraped data in this part is for building the
recommendation system using content-based filtering. We are not using collaborative
filtering because we found out that the scrapping time will increase sharply when
scraping loads the comments in each game. If we scrap less comments in each game,
it is hard to find the games which have positive comments from the same player. It is
because the STEAM platform has many players. Figure 5 shows one of the websites
about the game’s comments. Three types of data are scraped, the id of the game, users’
id which is highlighted in yellow color and the comments which are highlighted in
green color. By getting the game id from output_1F.pkl, we have to change all the
URLs to the comments page. Like scraping the URLs of the game, we need to scroll
down the page to try to scrap the greatest number of comments. A JSON file will be
created which stores the scraped comments.
848 M.-C. Yuen et al. / Game Recommendation System

4. Data Preparation. Some of the values will be modified during web scraping. For
example, to make the analysis process easier, ‘()’ and ‘,’ are replaced by empty space
so that the data type of this information is integer.

Figure 5. One of the websites about the game’s comments.

3.4. Data Modelling / Algorithm

In this model, we decided not to split data into training parts and testing. As the content-
based filtering is calculating their similar value by using matrix. We use cosine
similarity as a model to build the game recommender system with Python. cosine
similarity measures the similarity between two vectors of an inner product space. It is
measured by the cosine of the angle between two vectors and determines whether two
vectors are pointing in roughly the same direction.

Compute Term Frequency-Inverse Document Frequency (TF-IDF) vectors for

each document. TF-IDF is the frequency of a word occurring in a document, down-
weighted by the number of documents in which it occurs. This is done to reduce the
 M.-C. Yuen et al. / Game Recommendation System 849

importance of words that frequently occur in plot overviews and, therefore, their
significance in computing the final similarity score.

3.5. Data Presentation

Figure 6 shows the sunburst chart with no selection input. Figure 7 shows the sunburst
chart with the ‘Very Positive’ cell in ‘Recent Reviews’ path selected. Figure 8 shows the
sunburst chart with the ‘Action’ cell in ‘Genre’ path selected AFTER ‘Very Positive’
cell in ‘Recent Reviews’ path was selected.

Figure 6. The sunburst chart with no selection inputted.

850 M.-C. Yuen et al. / Game Recommendation System

Figure 7. The sunburst chart with the ‘Very Positive’ cell in ‘Recent Reviews’ path
selected.

Figure 8. The sunburst chart with the ‘Action’ cell in ‘Genre’ path selected AFTER
‘Very Positive’ cell in ‘Recent Reviews’ path was selected.
 M.-C. Yuen et al. / Game Recommendation System 851

We design and develop interactive dashboards by using Python and Tableau. Figure
9 shows the dashboard that is built with four basic graphs with the data we obtained.
Figure 10 shows the dashboard with one cell selected. Bottom left shows the top-rated
game which matches the element selected in that graph. Figure 11 shows the dashboard
with two cells selected. Bottom left shows the top-rated game which matches the
elements selected in those graphs. Figure 12 shows the dashboard with only the game’s
cell selected.
In this project, we used the Flask framework to run our recommendation system. It
includes the basic UI for user input and out and some data visualization charts. Our
business value is how much profit can be earned in the game selling. If the company
wants to launch a first-person shooting (FPS) game in the STEAM, we can help this
game evaluate whether it can make money. We can use the dashboard to check whether
the FPS game is one of the most popular games which is played by many people. The
company can analyze the dashboard and have a better consideration on launching the
games.

Figure 9. The dashboard.

Figure 10. The dashboard with one cell selected.

852 M.-C. Yuen et al. / Game Recommendation System

Figure 11. The dashboard with two cells selected.

Figure 12. The dashboard with only game’s cell selected.

3.6. Website

3.6.1 System Design

A website is created on localhost server. By running the program of the Flask framework,
the browser will be opened and go to the home page which is mentioned in the following
part. HTML, CSS are used to create the content and layout.
 M.-C. Yuen et al. / Game Recommendation System 853

3.6.2 Interface Design

In our website, three interfaces are designed as shown in Figure 13. Figure 14 shows the
home page of our game recommendation system. The home page provides a simple
background and mainly in blue color. A search box is created for the users to input the
full name of games. By clicking the submission button, the users page shows the
recommended results or the page which shows the results are not found. A bottom left
container shows our summary of the website. Figures 15 and 16 show the sunburst chart
by scrolling down the home page. By scrolling down the page, you can see the sunburst
chart which shows the percentage of positive reviews by recent reviews as shown in
Figure 15. The chart is interactive to users by clicking the genre of the game and the
recent review like the figure shown in Figure 16. Figure 16 shows the sunburst chart after
a user selected the ‘Very Positive’ in the ‘Genre’ path. By selecting this path, it shows
all the games and its genre which has a ‘Very Positive’ recent review in overall review.

Figure 13. Three interfaces in our website.

Figure 14. Home page of our game recommendation system.

854 M.-C. Yuen et al. / Game Recommendation System

Figure 15. The sunburst chart which shows the percentage of positive reviews by recent
reviews.

Figure 16. The chart is interactive to users by clicking the genre of the game and the
recent review.
 M.-C. Yuen et al. / Game Recommendation System 855

4. Preliminary Result on Performance Evaluation

4.1. Performance Metrics

The model can recommend games with the same taste as the user based on the game's
attribute information. For example, the user chooses “Counter-Strike: Global Offensive”
(objective-based, multiplayer first-person shooter) as input data of the model.
There is a similarity score to compare similarities between different
games and “Counter-Strike: Global Offensive”. If the similarity of two games is high,
the score is closer to 1. We expect our model to accurately recommend a similar game
according to the similarity score (1).

(1)

Figure 17 shows the similarity score to compare similarities between different

games and “Counter-Strike: Global Offensive”. If the score is closer to 1, it is more
similar between two games.

Figure 17. The similarity score.

4.2. Preliminary Results

In Figure 18, the page of recommendation results provides a simple background and
mainly in purple color. ‘Cities: Skylines’ is chosen as an example; it will show the top
five results related to ‘Cities: Skylines’ with their genres. And display like a food menu.
The link of the title of the game can be clicked to go to the website from STEAM to
check further details. The page displays the error message if the recommended results
are not found.
856 M.-C. Yuen et al. / Game Recommendation System

Figure 18. The page of recommendation results.

5. Conclusion

In this project, we have developed a games recommender system with complete

functional recommending features. We applied data visualization to our solution and can
successfully make the recommending system more appealing to the user, which
completes our goal of making a recommending system that feels more refreshing to the
existing one. Also, we achieve our goal of making a system that can showcase flexible
outcomes with the same element as the user input, which can give the user more choice
when finding the game they want.
We successfully apply machine learning, data science skills such as web
scraping, data preparation, machine learning algorithms, data visualization and Flask
application and build a working recommendation system, these skills will surely help us
furthermore in the future with our data science work. And we hope that our recommender
system can successfully help those people who want to find their favorite games.

References

[1] Cheuque, G., Guzmán, J., & Parra, D. (2019, May). Recommender systems for Online video game
platforms: The case of STEAM. In Companion Proceedings of The 2019 World Wide Web
Conference (pp. 763-771).
[2] Hannula, R., Nikkilä, A., & Stefanidis, K. (2019, September). GameRecs: Video Games Group
Recommendations. In European Conference on Advances in Databases and Information Systems (pp.
513-524). Springer, Cham.
[3] Pathak, A., Gupta, K., & McAuley, J. (2017, August). Generating and personalizing bundle
recommendations on steam. In Proceedings of the 40th International ACM SIGIR Conference on
Research and Development in Information Retrieval (pp. 1073-1076).
 M.-C. Yuen et al. / Game Recommendation System 857

[4] Yuen M. C., Chan S. L., Leung H. T., Wu P. L., Yip P. Y. (2019). “A System for Collecting and
Analyzing Data from Existing Game Selling Platforms”, Proceedings of The 2019 15th International
Conference on Natural Computation, Fuzzy Systems and Knowledge Discovery (ICNC-FSKD 2019),
July 2019, Kunming, China.
[5] Dokoupil, P., & Peska, L. (2023, June). The Effect of Similarity Metric and Group Size on Outlier
Selection & Satisfaction in Group Recommender Systems. In Adjunct Proceedings of the 31st ACM
Conference on User Modeling, Adaptation and Personalization (pp. 296-301).
[6] Yang S., Korayem M., AlJadda K., Grainger T., Natarajan S. (2017). Combining content-based and
collaborative filtering for job recommendation system: A cost-sensitive statistical relational learning
approach, Knowl.-Based Syst. 136 (2017) 37–45.
[7] Yuen M. C., King I., Leung K. S. (2015) “TaskRec: A Task Recommendation Framework in
Crowdsourcing Systems”, Neural Processing Letters Volume 41 Number 2, 2015.
[8] Yuen M. C., King I., Leung K. S. (2021) “Temporal Context-Aware Task Recommendation in
Crowdsourcing Systems”, Knowledge-Based Systems, Elsevier.
[9] Games Finder. https://gameslikefinder.com/
[10] Quantic Foundry. https://apps.quanticfoundry.com/recommendations/gamerprofile/videogame/
[11] Ikram F., Farooq H., and Nawaz W.. (2022). Multimedia Recommendation System for Video Game
Based on High-Level Visual Semantic Features. Sci. Program. 2022 (2022).
https://doi.org/10.1155/2022/6084363
[12] Yang L., Liu Z., Wang Y., Wang C., Fan Z., and Yu P. S. (2022). Large-scale Personalized Video Game
Recommendation via Social-aware Contextualized Graph Neural Network. In Proceedings of the ACM
Web Conference 2022 (WWW '22). Association for Computing Machinery, New York, NY, USA, 3376–
3386. https://doi.org/10.1145/3485447.3512273