An Area Recommendation Method Using Similarity Analysis for Play Patterns in MMORPG

Abstract

Recently, game companies have been increasingly offering a variety of content in their games. The more this happens, the more players will need to consider what is best for them. Players who have played such a game may not find it difficult to play, but those who are not used to play may have a hard time finding content. Therefore, in this paper, we try to give a customized guide to players in Massively Multiplayer Online Role-Playing Games (MMORPGs). We compare the similarity of growth speeds and visited areas, and then utilize this information to recommend the most similar characters. In this work, the K-means algorithm is used for clustering based on location, the Euclidean distance is calculated to recommend similar characters with similar growth speeds. In addition, Jaccard Similarity is introduced to recommend similar characters with similar access areas. Finally, we propose a method to recommend suitable areas by applying the access speed to the recommended characters in the previous steps. Our method achieves Precision and Recall of 0.74 and 0.81, respectively, on the real-life PvE (Player VS Environment) dataset.

1. Introduction

Humans are amused by perceived control while playing games. Perceived control is a belief that people can control their surroundings. Perceived control usually increases when one does things with their will [1]. Conversely, perceived control is lost if a person is upset by fear or being forced by someone or something. For example, in this study, players may be confronted with many contents and systems while playing the game. However, there are so many things to do, but they may not know what to do. This may leave the player feeling burdened. This is known as Hick’s law, a phenomenon of feeling overwhelmed by too many choices.

Ironically, humans prefer to have a lot of choices, but if there are a lot of choices, they look confused [2]. In the game world, it is the same way. Players want a variety of choices. They do not like to shrink their world. However, as time passes and the game gets bigger and bigger, players become overwhelmed. Then, the game stops being fun and players may feel like it is a daily chore they must do.

People play games because it is fun. It is not normal for players to feel pressured to play a game. Therefore, we tried to find ways to help ease the burden of choice for players in the game and give them some options which can be accepted by them with a high probability.

According to the 2021 “Korea Game White Paper”, the revenue of game industry has grown from KRW 15.575 trillion in 2019 to KRW 18.8855 trillion in 2020, an increase of 21.3%. Among them, Role-Playing Game (RPG)accounts for 67.9% accounting for the largest share. Despite the fact that the preferred platform is shifting from PC to mobile, the popularity of RPGs in Korea remains [3]. In addition, according to the “2021 Game User Survey”, RPGs are almost the most preferred game on both PC and mobile platforms [4].

However, in contrast to their popularity, massively multiplayer online role-playing games (MMORPGs) have a high barrier to entry. It is not easy, especially for new players, due to the accumulation of content. Compared to other players who have been playing for a long time, new players feel a big gap in everything. Just listing each piece of content is enough to exhaust new play. This can also happen with light users (called casual users, as opposed to hard-core users). In addition, hard-core users may sometimes feel this way. Some may follow general steps to develop their characters rather than their preferences or personal speed. Today, many games offer tutorial systems for users. However, there is a kind of trap in it. It is just an ideal step to grow, if a player wants to do another thing, that player has to move to find other things to do. Players could be used to it, but it is not a way to reduce their burden.

Therefore, we set the research goal as reducing burden of the players by focusing on searching the areas to recommend according to the method which is predicted to prefer.

In 1996, Richard Bartle (British game designer and creator of Multi-User Dungeon) wrote “Hearts, Clubs, Diamonds and Spades: Players for the Mud” and divided players into four categories in the book [5], as shown in Figure 1.

Killer is a type of player who takes action against someone. This type of player likes to compete, sometimes ahead of others.

Achievers are players who take action against the game world. This type of player tends to achieve what they are supposed to do.

Socializer is a player who is interested in building relationships with other players.

Explorer is a player who interacts with the game world. This type of player likes to explore the world and learn a lot about the content.

There are also many studies on classifying characters into several types. In 2008 there was research on classifying players based on immersion, which is related to character design. The player’s interaction with the game world is said to be the key to immersion. It produced character models that seemed to be more immersive to players. In the same year, another study immersed players in [6]. It deals with one aspect of game design. After eliminating features that interfere with immersion based on the “fun theory of game design”, it may elicit a better response from users [7]. In 2014, a study was conducted on Aeon of Strife (AOS) also known as Multiplayer Online Battle Arena (MOBA) to classify players into four types [8]. It classified players as extroverted or introverted, and individual or social. In 2018, an improved purchase prediction study was held [9]. This study focused on the fact that most players do not spend their money on games. In order to satisfy players and give them better service, this study was based on the cognitive-psychological characteristics of players. In 2020, a study was held to control the difficulty of user customization using artificial neural networks [10]. It analyzes the player’s patterns and dynamically controls the situation.

In the gaming industry, many attempts are constantly trying to meet the needs of players, such as genre fusion or grafting new technologies. Often, genre fusion is used because it is difficult to survive with only one genre, as one world-famous game has already taken over the market. Play-to-Earn (P2E) which is a game that allow players to earn money while playing and Play-to-Own (P2O) which allows the user to own the result obtained through the game are new genres that have been newly created. They are new genres that are the opposite of the existing Pay-to-Win (P2W) genre. The Virtual Reality (VR), which is a technology that allows users to experience the virtual world), Augmented Reality (AR) which is a technology that shows some information as virtual information based on the real world and Metaverse which is a virtual world where all activities are like the real world, have constantly been used to make a difference to other games. In addition, there is a growing interest in improving the quality of services through data management, data collection and statistical analysis processing [11]. Yang et al. [12] proposed a Socially-aware Contextual Graph Neural Recommendation system (SCGRec), which exploits three perspectives to improve game recommendation, namely, personalization, game contextualization and social connections. Moreover, there is a study of recommending items in the game. Duan et al. [13] presented a relationship-aware graph attentional item recommendation system. They considered the relationship between characters and items. Moreover, with the rapid development in the field of computer vision in deep learning, there is a recent game recommendation system based on the semantic information of visual information. Ikram and Farooq [14] introduced an approach called Deep Visual Semantic Multimedia Recommender System (D_VSMR) to process high-level visual features of multimedia and make video game recommendations based on visual semantics. However, in this work we aim to recommend areas in the game to users based on their profiles.

Our goal is to find a way to recommend a direction for a character to help his or her play, other than following a tutorial or basic quest flow. When a region is visited, where to choose the next region to visit can vary depending on the needs at that time, so we decided that collecting and analyzing data would provide satisfactory results to players.

2. Methods Overview

In this study, we recommend areas in the game that are predicted to be preferred using a clustering model and similarity models. The overview of this study is depicted in Figure 2.

The main data were level of character, date that players play the game and the visited area where players visited to do something.

Since level data are one of the basic standards to check players’ characteristic, level data was used for both level similarity and area similarity. Figure 3 shows the features of data and models we used in this study.

3. Design and Implementation

3.1. Area Clustering

The regional data are the most important data in this study since the goal of this work is to recommend area in a game. It is ubiquitous throughout the study. Therefore, if there is a problem with the regional data, the results may be completely different. Therefore, when using K-means for clustering, it is important that it is accurate.

When the clustering is complete, the region data can be used. There are many different coordinates scattered in the game world. In order to use these coordinates for recommendations, we have to group them into some clusters. We use K-means to cluster the coordinates into regions with labels. Usually, the number of K is calculated by some formula, such as the Elbow method or the Silhouette method. However, we set K manually to be the same as the real data. In addition, the centroid is also set manually to be the same.

In general, centroids are continually updated according to the data in cluster. However, there are some places that less players visit. In this case, this area might be belonged totally different cluster when it starts clustering if centroids set automatically.

In Figure 4, the left one is an example of real area, and the right one is an example of clusters according to the way of setting centroids. The first row assumes to make cluster manually, and the second row assumes to make cluster automatically. For the first one, since it is same as the real centroid, it is highly likely to be accurate. For the second one, it could be different from the real world. The centroids could be calculated totally different when the different areas face to each other which space has high density even they are different or when the areas are in one sector, but the density space is dispersed into some places. To reduce the possibility of such errors and build a similar environment to reality, we used fixed K and centroids to make clusters.

3.2. The Similarity of Levels

Players tend to become immersed in the game as their character levels increase and the amount of time they spend playing increases [15]. Character level is one of the most basic characteristics of character growth. Higher levels over the same period of time indicate that they are playing more than other players. In this way, it can be seen that these types of players enjoy the game more. Therefore, based on the premise that growth speed is influenced by immersion, we devised a method to divide each character into a hard-core player group and a lighter player group so that the method can recommend similar characters to the target character.

In this study, there was no process of grouping players by specific types. This is because there is a premise that the game pattern can be changed depending on the situation they face.

To start the study, we checked the daily maximum level for each character. As a preprocessing, if for some reason someone did not have level information for a specific date, we set and used the maximum level from the previous day to make the dataset the same as shown in

Table 1. Preprocessing of level data.

Day 1	Day 2	Day 3	Day 4	Day 5	Day 6	Day 7
36	51	53	55	55	NaN	NaN
50	67	70	71	73	75	75
37	56	60	70	73	75	NaN
54	69	77	79	80	80	81
Day 1	Day 2	Day 3	Day 4	Day 5	Day 6	Day 7
36	51	53	55	55	55	55
50	67	70	71	73	75	75
37	56	60	70	73	75	75
54	69	77	79	80	80	81

. Like this, we consist of datasets that can be compared with each other.

The growth speed of each character is measured based on their daily rank. The similarity using Euclidean distance is based on the growth speed of the characters for a week. When the method recommends 10 similar characters based on the growth speed, the level information of the recommended characters is shown in

Table 2. Maximum level of characters who have high similarity of growth speed.

Day 1	Day 2	Day 3	Day 4	Day 5	Day 6	Day 7
52	72	80	81	82	83	83
50	72	80	81	82	82	83
51	74	79	80	82	82	83
50	71	79	80	81	82	82
51	75	80	81	82	82	83
51	75	80	81	81	82	82
51	74	79	80	81	81	82
49	72	79	80	81	81	82
54	73	78	80	80	81	82
53	76	80	82	83	83	84
49	72	79	80	80	81	81

.

In Table 2, the first row is the information of the target characters, and from the second to the last row is the information of similar characters. Their data cannot be exactly the same. However, although there are some deviations, it seems that they do not differ significantly across the range.

The similarity levels are tabulated in

Table 3. Similarity between a target character and recommended characters in Table 2.

Day 1	Day 2	Day 3	Day 4	Day 5	Day 6	Day 7	Similarity
52	72	80	81	82	83	83	-
50	72	80	81	82	82	83	2.23
51	74	79	80	82	82	83	2.82
50	71	79	80	81	82	82	3.16
51	75	80	81	82	82	83	3.31
51	75	80	81	81	82	82	3.60
51	74	79	80	81	81	82	3.60
49	72	79	80	81	81	82	4.12
54	73	78	80	80	81	82	4.35
53	76	80	82	83	83	84	4.47
49	72	79	80	80	81	81	4.80

and shown in Figure 5. The left graph in Figure 5 is the result of the level information of target character and the right graph in Figure 5 is the result of the level information of similar characters. Those two graphs are shown the similar shape from the first day to the last day.

3.3. The Similarity of Areas

In the real world, there are personality type tests, such as the Myers-Briggs Type Indicator (MBTI) and Enneagram. In the gaming world, there are also some player types, such as Richard Bartle’s four Player Types. This is because there are some common phenomena based on personality. However, in this study, we did not group players. Kurt Lewin (an American psychologist) says that humans perceive themselves in their environment. Depending on the given situation, a de-differentiation of the living space (a situation where the living space is dominated by one person and there is no cognitive ability to think about others) may occur [16]. Based on this theory, we hypothesize that the situation has a greater impact on the player regardless of a person’s personality.

Figure 6 is a successful interest curve in “The Art of Game Design” by Jesse Schell. We tried to match this curve to the player’s interest in game play. The interest at the beginning (A), feel fun due to the rapid growth at low level (B), repetition of the process of finding a new interest after losing interest as growth speed decreases after reaching at a certain level (C∼G), and ended the game with satisfaction as a result (H). For example, a step C to G, the players are trying to find new interest, but if they cannot they lose interest and might quit the game.

Based on the above, we analyzed the regional preferences after the rank change. Region preference was based on the number of visits to the region. In measuring similarity by analyzing each player’s preferred region, we had to decide how to study between two methods. The first method measures the entire level range; it recommends characters with high similarity for a particular level region. It divides the entire level range into sections and recommends characters with high similarity to the previous level section based on the target level. We have assumed several cases to decide which one we must use.

• In MMORPG, every character has a role. There might be the areas that must be visited at a specific level for a quest, or something needed for each role, or contents that must be progressed.
• The beginner’s section is highly likely to show a similar pattern regardless of the role. Even if the similarity of the beginner’s section is close to 100%, the after steps could not be similar.
• Even the same player would not show consistent play patterns in all ranges. The play pattern might be changed depending on the situation the player is in.

We decided to use the second method. Depending on the way the player tries to achieve the goal, the game mode can be changed frequently. After studying it, we found that it turned out the way we thought it would.

Figure 7 is a graph of the similarity of level sections which is divided through the entire level as shown in

Table 4. Difference between the similarity of entire level and level Sections (1).

Similarity of Entire Level	Similarity of Level Section 1	Similarity of Level Section 2	Similarity of Level Section 3	Similarity of Level Section 4	Similarity of Level Section 5
0.7	0.0	1.0	0.5	0.0	0.0
0.7	0.0	1.0	0.6	0.0	0.0
0.7	0.0	1.0	0.5	0.0	0.0
0.7	0.0	1.0	0.2	0.0	0.0
0.7	0.0	1.0	0.0	0.0	0.0
0.7	0.0	1.0	0.6	0.1	0.0

. It is based on the similarity of entire level when it is $0.7$. According to the Figure 7, level Section 2 shows the highest similarity of 100% but others do not. We guess that high similarity at entire level does not mean it is similar to all divided level sections.

Figure 8 is a graph of the similarity of level sections which is divided through the entire level as shown in

Table 5. Difference between the similarity of entire level and level Sections (2).

Similarity of Entire Level	Similarity of Level Section 1	Similarity of Level Section 2	Similarity of Level Section 3	Similarity of Level Section 4	Similarity of Level Section 5
0.1	0.0	1.0	0.0	0.1	0.2
0.2	0.3	1.0	0.0	0.0	1.0
0.5	0.0	1.0	0.4	0.1	0.5
0.5	0.0	1.0	0.4	0.1	0.0
0.6	0.0	1.0	0.6	0.1	0.0
0.6	0.0	1.0	0.6	0.1	0.0

. It is based on the similarity of level Section 2, when it is $1.0$. According to the Figure 8, level Section 2 shows the highest similarity as 100% but others do not include the similarity of entire level. We guess that high similarity at specific level sections does not mean it is similar to the entire level range or other level sections.

Naturally, there may be some relationship between each part and the whole range, and the accuracy of the similarity based on this relationship is unlikely to be very high. Therefore, we divided the entire range into sections and measured the similarity between each character.

The level is roughly divided into three sections; low, middle and high. Each section was further divided into subsections. Subsections are divided according to the number of level experiences.

In the low-level section, the beginner level is excluded. This is because it is similar for every player. It seems difficult to have various patterns in the low-level section, so we grouped the rest of the levels in the low-level section into one section.

In the middle-level section, all sub-sections are more narrowly divided than in the lower section. This is a section where you can see many aspects. Some players play safely in the lower-level sections, but some players challenge themselves to the higher level sections. This is also the period when players get used to the system. Players will know what to do based on how they have played, but it has a downside and may be limited to what they have already played.

In the high-level section, all subsections are more narrowly divided than in the middle-level section. In this section, the experience needed to grow their character is very high. This means that players need to spend more time in this section. Since it requires relatively more time, we designed the high-level section to be divided into narrower subsections than the mid-level subsections. This makes it easy to observe changes.

We tried to set weights for each subsection, and depending on the balance we set may produce completely different results. Therefore, weights were not given in the study in order to get more accurate results.

3.4. The Similarity of Visited Areas

Each player can change the game mode any time they need to. Therefore, even if it looks similar, its similarity may not quite be that similar. However, it is not possible to have players with 100% identical patterns. Therefore, if it looks similar, then it may have a higher degree of similarity than other players. As with DTW, it does not have to be one-to-one, it only needs a range of similarity in the near segment.

We used Jaccard Similarity to compare the similarity of visited areas in each level section. Set union as entire visited area between two target characters, and intersection as same areas they visited. Divide intersection with union is the result of Jaccard Similarity.

We used it to research to find out how accurate it is. Measure the similarity for each level subsection first, and arrange characters descending towards the similarity.

Table 6. Similarity of character with high similarity at level Section 5.

Level Section 1	Level Section 2	Level Section 3	Level Section 4	Level Section 5	Level Section 6	Level Section 7	Level Section 8
0.25	0.36	0.29	0.17	0.80	0.75	0.44	0.67
0.00	0.13	0.35	0.12	0.60	0.75	0.09	0.13

shows the information of characters with high similarity at level section 5. They seem to have relatively high similarity at level section 6.

Table 7. Similarity of character with high similarity at level Section 6.

Level Section 1	Level Section 2	Level Section 3	Level Section 4	Level Section 5	Level Section 6	Level Section 7	Level Section 8
0.00	0.42	0.24	0.00	0.08	0.75	1.00	0.00
0.25	0.22	0.00	0.00	0.38	0.60	1.00	0.00
0.25	0.40	0.55	0.28	0.00	0.57	0.57	0.25

shows the information of characters with high similarity at level section 6. They seem to have relatively high similarity at level section 7 as well.

In some cases, the similarity in the subsection which comes after the target subsection might be seemed not high even their previous subsection has high similarities. This is caused because it is just a simple area comparison between characters.

3.5. Visited Ratio Applied

There are two ways to compare the similarity between characters. One is to compare its rank and the other is to compare its access area. In both cases, the performance evaluation is relatively low. We consider the problems with these methods as follows.

• The similarity of growth speed: no other way than killing was considered. The data is too limited.
• The similarity of visited area: Even though it was analyzed with level and area, the frequency was not considered. Therefore, regions with very low visitation are also recommended, which makes the similarity low.

In order to obtain more accurate results, we designed the method in three steps as shown below.

• Limit the character pool through two previous methods. (Similarity of growth speed and visited area).
• Measure data about visited areas.
• Filter data from number 2 with visited ratio.

We assume that there are ‘passing by’ area and ‘destination’ area in the game. ‘Passing by’ area means an area that is not important; sometimes it is just one of the routes to pass. However, ‘destination’ area means it could be one of the final goals for the player. The play ratio between both of them would be much higher in ‘destination’ content than ‘passing by’ one.

Table 8. Example data for Jaccard similarity.

	Visited Area at Specific Level	Jaccard Similarity
Character A	1, 2, 3, 4, 5, 6	$\frac{\{2,4,6\}}{\{1,2,3,4,5,6,8\}}=0.43$
Character B	2,4,6,8
	Visited Area at Specific Level	Jaccard Similarity
Character A	1, 2, 3, 4, 5, 6	$\frac{\{1,2,3,4\}}{\{1,2,3,4,5,6\}}=0.67$
Character C	1, 2, 3, 4
	Visited Area at Specific Level	Jaccard Similarity
Character B	2, 4, 6, 8	$\frac{\{2,4\}}{\{1,2,3,4,6,8\}}=0.33$
Character C	1, 2, 3, 4

shows the example of Jaccard similarity and visited ratio of these examples are shown in

Table 9. Examples of adapting visited ratio in Table 8.

	1	2	3	4	5	6	8
Character A	0.07	0.50	0.10	0.30	0.02	0.01	0.00
Character B	0.00	0.60	0.00	0.20	0.00	0.05	0.15
	1	2	3	4	5	6	8
Character A	0.07	0.50	0.10	0.30	0.02	0.01	0.00
Character C	0.70	0.10	0.10	0.10	0.00	0.00	0.00
	1	2	3	4	5	6	8
Character B	0.00	0.60	0.00	0.20	0.00	0.05	0.15
Character C	0.70	0.10	0.10	0.10	0.00	0.00	0.00

. In Table 8, it looks like the similarity of Character A and Character C is high, but in Table 9, it looks like the similarity is higher between Character A and Character B. It was happened because the previous method adapted in Table 8 only compared the visited area.

By filtering the outliers, we can recommend the prediction as the preferred region. The results of this method are as follows.

Table 10. Adapting visited ratio in Table 7.

Item	Value for the Item
Target Character	Character no. 1
Target Level	76
Level subsection to be compared	Level Section 6
Standard visited ratio	50%
Recommended Character	Recommended Area	Visited Ratio
Character no. 108	44	100%
Character no. 67	44	100%
Character no. 129	9	73%
Character no. 47	9	66%
Character no. 77	147	62%

shows a recommended area to character number no.1 at level 76. By comparing the similarities of growth speed and visited areas, it found out similar some characters. Then, among those similar characters, found out the areas where each character visited more than 50%. As above, if only the areas where have visited ratio more than 50% recommended, the target player can receive area recommendation based on high preference.

4. Performance Evaluation

In this study, performance evaluation was conducted in two categories. Evaluate for each pre-step and evaluate for final result. The process is shown in Figure 9.

Precision shows its performance evaluation with values that is actually true among the predicted true. Recall shows its performance evaluation with values with true positive and false negative, which means predicted true among actually true.

In the study, Precision was used as (the area where the character actually visited/the area where we recommended) and Recall was used as (the area where the character visited among the recommended area/the area where the character actually visited).

4.1. Evaluation for Each Pre-Step

The performance evaluation with Precision and Recall was about the similarity of growth speed, visited area and visited area with visited ratio as 50%. The result of both growth speed and visited areas showed lower than the result of visited areas with visited ratio. This result demonstrates that by considering the visit ratio we can improve the recommended performance of our method. In Figure 10, all shows higher performance in Recall in common.

4.2. Evaluation for Final Result

The second performance evaluation was about the similarity of visited area with visited ratio as 30%, 50%, and 70%. Results are shown in Figure 11. Wen set visited ratio as 30%, The recommended area range is larger than expected, but the denominator also became larger, and the Precision value is lower than when the visited ratio was 50%. It was expected high-accuracy recommendations will be possible because the area which has a visited ratio relatively low could be recommended either in a wider range. The recommended area range is larger than expected, but the denominator also became larger, and the Precision value is lower than when the visited ratio was 50%. When set visited ratio as 70%, the gap between players with similar tendencies was narrowed. The Recall was higher than when it was 30%, but lower than when it was on 50%. The performance evaluation was expected to be better than it was set in 30% even though it can be classified extremely through character’s characteristic.

4.3. Disscussion

As mentioned in the title, this study is an approach that can be applied due to the characteristics of MMORPGs rather than all genres, and there is a limitation in that it is difficult to apply the same value at all in that the classification criteria must be divided according to the characteristics of each game. It can be effective when you divide it from the position of knowing the level standards well, and if there is a growth method that is completely unrelated to the level or deviated from the existing MMORPG method, the accuracy may be further reduced. If detailed data such as which items can be obtained from a specific area and which areas are visited by characters of a specific level to obtain the item are added, the accuracy will be slightly higher.

5. Conclusions

Players may be happy or bored while playing a game. In this study, we focused on finding a way to improve the fun of a potentially boring game. There are many ways to lose the fun. Raph Koster, who developed Ultima Online, Star Wars Galaxy, and Everquest, lists several cases where players may lose fun in “Theories of Fun in Game Design”. In some cases, the fun can be lost. The game is boring since it is too easy, there is a lot of content but not enough fun, the game is boring because it is hard for the player to find patterns, the game is too hard or too slow, and there is no more fun because all patterns are identified [18]. Among them, this study focuses on the situation where players are bored because they cannot find the patterns.

In this study, since only area and level data were mainly used without other additional data left little disappointment in the performance evaluation results. As a future study for area recommendation or other contents recommendation, it is expected that the research would be better if there are various contents used. If not only just visited area or growth speed, but also reflecting the preference of the players, it is expected to conclude the higher accurate results could be obtained. The data used in this study is only simple area information and level information, but if a little more variety of data is available, we think it will be possible to make a customized proposal process. If we predict and recommend what players want based on data from other players, players will be more interested in the game and will be able to find their own play patterns. As all services become more and more personalized, it is expected that even in-game, players will be able to focus on and recommend them to have a better effect. Although this study is limited to regional recommendations, it is expected that a more sophisticated guide can be provided by considering the proportion of content participation.