Mobile Game Recommendation using Touch Gestures

Hao-Tsung Yang, De-Yu Chen, Ya-Xuan Hong, and Kuan-Ta Chen
Institute of Information Science, Academia Sinica

PDF Version | Contact Us


Today, any Internet user can find out and download more than one hundred thousands of games on mobile app marketplaces; nevertheless, how to pick the best games out of the large pool without spending much time on tryout is very challenging. The common rank list and social recommendation approaches for discovering new games do not work well when a player wants to search for games with a particular gameplay, or he may want to find out all the slow-paced games suitable for his grandparents. There is currently no feasible way to do that because these requirements cannot be formulated in a proper text search query.
In this paper, we propose a scheme that can discover mobile games with similar gameplay based on players' touch gestures while playing a game. We select 22 mobile games from 5 genres and recorded 94 touch gesture traces from 5 subjects. The evaluation results show that 1) touch gestures can serve robust signatures of gameplay as the key traits of the touch gestures from different subjects remain consistent; 2) our scheme can give reasonably accurate recommendations of similar games simply based on touch gestures.

1  Introduction

Today, any Internet user can find out and download more than 150,000 iOS games on App Store2 and more than 120,000 Android games on Google Play3. These figures evidence the booming of the mobile game industry in recent years. The launch of mobile app marketplaces such as App Store and Google Play is good and bad to mobile game developers. The pros come from the unprecedented low publishing cost for a new mobile game; meanwhile, as there are simply too many choices on the market, getting the attention of potential players for new game titles is more difficult than ever. The same situation applies to mobile game players as well. While it is certainly nice to have more than 100,000+ games for download and play, how to pick the best games (in terms of personal preference) out of the large pool without spending much time on tryout is very challenging, if not impossible.
To discover fun mobile games, it is a common practice for gamers to follow the rankings provided by marketplaces. For instance, Google Play provides rankings for the "top" games in different categories, including the top paid games, the top free games, and the most popular games, and so on, and provides a handy user interfaces for users to install games on the top lists. Players can also get recommendations of particular games on Internet forums or from their own social circles. The billboard and social recommendation approaches work perfect when a player is not looking for any specific type of games; however, these approaches do not work well when this is not the case. A player may want to search for games with a particular gameplay4, or he may want to find out all the slow-paced games suitable for his grandparents. For example, if a player favors the gameplay of Angry Bird and would like to try out all the Angry-Bird-variants on a marketplace, there is currently no feasible way to do that because these requirements cannot be formulated in a proper text search query.
Marketplaces of mobile games may provide certain recommendation algorithms, but such algorithms are normally based on association rules [1], rather than on the gameplay of mobile games. Furthermore, as gamers' download behavior is highly dependent on the ranking lists and the cost of downloading a game is virtually zero, the association rules between game downloads would be less correlated with the categories of the games (comparing with the association rules between book sales in bookstores). Therefore, a quick check of the "recommended games" provided by mobile game marketplaces would reveal that the current mobile game recommendations either by 1) association rules, that will recommend top games; or by 2) text mining, that will recommend games with similar titles and descriptions. Either approaches give recommendations regardless of game genres. To the best of our knowledge, this work is the first to take account of gameplay in the personal recommendations of mobile games.
In this paper, we propose a mobile game recommendation system based on players' touch gestures while playing a game. Our rationale is that, since it is difficult to clearly describe the gameplay for a particular game, i.e., the way in which a player interacts with the game, we consider that the touch gestures, i.e., the time series of players' (multi-)touch events on the screen of a mobile device, can well represent the gameplay of a particular game and thus can serve as a signature of the game. Therefore, in our recommendation system, we treat a trace of touch gestures as a signature of a particular game and recommend similar games based on the signature. As long as we have a database of gameplay signatures for a variety of mobile games, we can compute the similarity between any two games and recommend games with similar play styles to users based on the similarity metrics. To verify the efficacy of our proposed system, we select 22 mobile games from 5 genres and recorded 94 touch gesture traces from 5 subjects. The evaluation results show that 1) touch gestures can serve robust signatures of gameplay as the key traits of the touch gestures from different subjects remain consistent; 2) our recommendation system can give reasonably accurate recommendations of similar games simply based on touch gestures.
Although our recommendation system is evaluated based on a small group (22) of games, we believe the potential of mobile game recommendation will be much more evident when the game pool is much larger, say, 10,000+. The collection of game signatures (i.e., the touch gestures) would be feasible by adopting the crowdsourcing strategy, that is, letting players upload their own touch gestures while playing a particular game. It is hoped that this work will motivate further research by showing that touch gestures are capable of describing gameplay and can serve a robust signature for mobile game recommendation.
The remainder of this paper is organized as follows. Section II discuss related works, and Section III describe the setup of the experiment and gesture trace we collected. Section IV and V illustrate how we analyze gestures and explain the recommendation algorithm. Next, we evaluate the proposed scheme in Section VI. Finally, Section VII draws our conclusion.

2  Related Work

Early implementations on personalized recommendation system in mobile environment usually focuses on the improvements of recommender user interface for mobile users [2]. The MovieLens Unplugged project [3] explores the usability of such recommender systems on mobile devices with limited display size and intermittent network connectivity to the backend data store. More recently, the idea of utilizing context in recommender systems has gained wider attention. The context means information which can be a feature of environment such as time and a location. For example, a context-aware restaurant recommendation system [4,[5] can suggest appropriate restaurants to users based on their current location [6] and time [4].
Traditionally, game recommendation is usually implemented based on a classification of game genres [7]. However, as pointed by Genvo and Wolf [8,[9], game genre classification is quite limited because 1) there are virtually unlimited gameplay so that classifying them into definite classes is rather hard and more or less subjective; and 2) technological evolution (such as multi-touch input devices, Wii, and Kinect) usually encourages new types of gameplay those cannot fit in existing categories; thus criterions for genre systematization need to be revised from time to time. Therefore, these seems no universally acceptable game classification system that can well categorize all the current and future computer games. For the above reason, Damien Djaouti [10] has proposed to classify video games by game rules, which are described by criterion bricks. For instance, one criterion brick would be whether the avatar is supposed to kill opponents, while another brick represents whether the avatar needs to avoid some stuff or events. All the games can then be characterized and categorized by the set of criterion bricks used in gameplay.

3  Data Collection

Table 1: Selected mobile games in our study
Genre Shorthand Game title # Downloads†
Temple Temple Run 50M
Pitfall Pitfall! 1M
Subway Subway Surf 100M
Gran Angry Gran Run 10M
Fred Running Fred 5M
Trigger Dead Trigger 10M
CS CS Portable 10M
Strike Enemy Strike 1M
Gun Gun of Glory 0.5M
Cytus Cytus 1M
Tap Tap Tap 4 10M
Dance Finger Dance Lite 5M
Hero Music Hero 10M
iFighter iFighter 1945 1M
Raiden Raiden Fighter 2013 5M
Laser SuperLaser 5M
AirAttack AirAttack HD Lite 5M
Jewels Jewels Dash 1M
Diamond Diamond Blast 5M
Blink Blink!Blink!Blink! 5M
Marble Marble Blast 3 1M
† The number of downloads for each app is according to Google Play (as of Sep 2013).
Table 2: A sample segment of raw touch gesture logs from Sony Xperia Tablet Z
event id time gesture id x-axis y-axis event
11 36.10 42691 110.00 721.00 press
12 36.20 -1 330.00 661.00 press
13 36.30 -1 467.00 332.00 press
14 36.40 -1 984.00 227.00 release
15 37.10 42692 221.00 429.00 press
16 37.20 -1 224.00 456.00 release
17 37.40 42693 237.00 434.00 press
18 37.40 -1 237.00 434.00 release
Table 3: A summary of collected traces
Subject Age Sex # Tr. l l
Subj1 27M 2195 (4.5) min 4029 (42.6)
Subj2 27M 21103 (4.9) min 3899 (37.8)
Subj3 22F 1886 (4.8) min 4647 (54.0)
Subj4 19M 1574 (4.9) min 5268 (71.4)
Subj5 28F 1930 (1.6) min 1136 (38.4)
Overall 94 388 (4.1) min 18979 (48.8)
We select 22 mobile games from 5 popular genres as the study targets in this work. These games can all be freely downloaded from Google Play (as of Sep 2013). We summarize them in Table I.
As discussed in Section I, we consider that the touch gestures performed by gamers while playing a mobile game can be seen as a signature of the gameplay of the particular game; thus, here we describe how we collect the touch gestures during game play. Because the low-level touch gestures are not available via Android system calls, we set up an experiment testbed comprising a Sony Xperia Tablet Z (10" multi-touch LCD) and a personal computer with Microsoft Windows installed. The Tablet Z is connected to the PC via a USB cable so that we can use the Android Debug Bridge (adb) to instruct the system-level operations on Tablet Z from the Windows PC. adb supports a shell getevent command that can be used to dump the raw touch events occurred on the Tablet Z. Table II lists a sample segment of raw touch gesture logs captured from Tablet Z. Later in Section  IV-B we will describe how we extract features from these gesture logs.
We recruited 5 subjects and made them play the selected mobile games on Tablet Z in our laboratory. Each subject was asked to play selected games with a random order for nearly two hours (except the fifth subject who quitted earlier due to personal matters) and given US$16 as a reward. Table III presents a summary of the subjects and the collected traces. Our subjects include 3 males and 2 females with ages within the range of 19 and 28. We have collected a total of 94 traces (i.e., 94 game sessions) for 388 minutes (6.4 hours) of touch gesture logs. On average, a player performs 48.8 gestures per minute when they were playing the selected mobile games. Disparity exists between the average gesture rate among the subjects. For example, Subj2 performs 37 gestures per minute and Subj4 performs 71 gestures per minute, which is a significant difference. We believe such discrepancy is due to two reasons: 1) Individual differences in gameplay strategies; and 2) because not every subject played all the 22 games, the average gesture rate may be subject to the set of games that were played by a subject in our experiment. For example, endless running games require much less actions (i.e., gestures) for a decent gameplay than first-person shooter games, that would require players to hit the screen continually in sub-seconds.

4  Touch Gesture Analysis

In this section, we analyze the touch gestures we collected for the selected mobile games. We begin with an overview of aggregate gestures for each game and investigate the similarity and dissimilarity between the games. Secondly, we describe how we extract the features from the touch gestures. Lastly, we visualize the touch gestures using a multi-dimensional analysis and examine the distribution of the gesture traces.

4.1  Gesture Overview

Table 4: The screenshots and aggregate gesture trajectories for 15 sample games
l   Temple er1.pngt_run_1.png   Subway er2.pngt_run_3.png   Fred er3.pngt_run_5.png
l   Trigger fps1.pngt_fps_1.png   CS fps2.pngt_fps_2.png   Strike fps4.pngt_fps_4.png
l   Cytus mug1.pngt_mug_1.png   Dance mug2.pngt_mug_3.png   Hero mug3.pngt_mug_4.png
l   iFighter ss1.pngt_bhell_1.png   Raiden ss4.pngt_bhell_3.png   AirAttack ss3.pngt_bhell_5.png
l   Jewels tm1.pngt_square_1.png   Diamond tm2.pngt_square_2.png   Blink tm3.pngt_square_3.png
To observe the similarity and dissimilarity of touch gestures across games, we plot the aggregate gestures for the selected 15 mobile games in Table IV. In each of the cells of the table, a screenshot of a mobile game and an aggregate view of the touch gestures performed by our subjects are presented. On the graph, a tap gesture is plotted in a blue dot and a slide gesture is represented in a purple line segment. We define a gesture as a tap gesture if the locations associated with the press and release actions (c.f., Table II) of a gesture are within 0.5 cm5; otherwise it is considered a slide gesture.
From Table IV, we can make the following observations: 1) Games in different genres can elicit very different patterns of touch gestures. For example, players tend to make vertical and horizontal slides in endless running games, while players make a significant amount of taps at a few particular locations on the touch screen and make very few slides in rhythm games. 2) Games in the same genres tend to elicit similar patterns in touch gestures. It can be seen that the touch gestures are so similar for the three scrolling shooter games (by sliding all over the touch screen toward any possible direction) and for the tile matching games (by tapping on certain grid locations and sliding between adjacent grids). The same observations apply to the other three genres. From these observations, we are now more confident that touch gestures can well serve the signature of the gameplay for mobile games and can be used to discover mobile games with similar gameplay.

4.2  Feature Extraction and Visualization

Table 5: The features characterizing touch gestures
Feature Meaning # Dimensions
    rate Rate of actions, taps, and slides 3
    burst Gesture burstiness 4
    interval interval time between successive gestures 4
    position Press and release position 8
    movement horizontal and vertical movement 4
    multitouch Ratio of simultaneous gestures 2
    t-duration Tap duration 5
    t-position Press and release position 4
    s-duration Slide duration 5
    displacement Displacement 5
    distance Movement distance 5
    speed Movement speed 5
    angle Angle between the x-axis and movement vector 6
We now describe how we characterize touch gestures using a set of numeric features. We begin by introducing the concept of segment. While each touch gesture trace is associated with a game session (normally 5 minutes), we do not consider the touch gestures throughout a game session can represent the gameplay of the game because there might be some periods the player is at the title screen, choosing stage, or adjusting options. The touch gestures made during these periods would be recorded by our logger but are not dependent on the gameplay. Thus, we partition each gesture trace into a number of 10-second segments and discard the segments with a gesture rate lower than a certain threshold6. In other words, we include only the "active" segments with a higher gesture rate in further analysis as those segments would be more relevant to the gameplay behavior.
We extract features for touch gestures in a segment level. Three categories of features are defined: General, Tap, and Slide. The General category defines the common properties for gestures, such as the rate of gestures, the interval time between successive gestures, and the ratio of multi-touch gestures within a segment. The Tap category defines additional properties for "tap" gestures, including the duration and positions of tap gestures within a segment. The Slide category defines the displacement, distance, speed, and other properties of "slide" gestures within a segment.
Figure 1: The distribution of traces in our database.
After extracting the 69 features for each segment, we then compute the feature vector associated with each gesture trace by averaging the features of all the segments from the gesture trace. In order to visualize the similarity of gesture traces based on the feature vectors, we compute the Bray-Curtis dissimilarity [11] between every pair of gesture traces and then map the traces onto a three-dimensional coordinate system based on their dissimilarity matrix using the isometric feature mapping method [12]. We next invoke the non-metric multidimensional scaling (MDS) to map the 3D coordinates to 2D coordinates and plot the resulting diagram in Figure 1. The symbols in different colors denote games in different genres. From the graph, we can see that most traces from the same genre are clustered together except a few cases. A closer look reveals that this is due to the different control methods adopted by the games even thought the gameplay is very similar. For example, Running Fred is an endless running game with gameplay similar to other endless running games such as Temple Run and Subway Surf. A typical game of this kind usually uses slide gestures to control the actions of the avatar, such as jump, glide, and direction change. However, Running Fred chooses to use tap gestures to control the avatar actions instead. This explains why Running Fred looks dissimilar from other endless running games but more similar to tile matching games, which elicits mostly tap gestures, according to Table III. A rhythm game Cytus has a similar situation with Running Fred. While typical rhythm games regulate players to tap on a number of fixed locations on screen, in Cytus, tap locations are not fixed but can be everywhere on the screen; plus, slide gestures are occasionally required in the gameplay. These difference in control methods make Cytus more dissimilar from other rhythm games and more similar to tile matching and first-person shooter games.
Despite the above two special cases, touch gesture traces we collected are mostly concentrated together on the two-dimensional MDS plot. We consider that this indicates that touch gestures are robust, player-independent signatures for gameplay and can be further employed for discovering similar games based on this informative property.

5  The Recommendation Algorithm

In this section, we describe the algorithm in our recommendation system which utilizes touch gestures as a signature for gameplay. Given a target game, the algorithm outputs a set of games with similar gameplay.
The algorithm takes a set of touch gesture traces, GX, from game X as the input and identifies the game Y with a gameplay most similar to X. For each touch gesture trace, we first derive the 69-dimension feature vector of the trace (ref. Section  IV-B). We then compile a rank list of touch gesture traces from all the candidate games by computing the cosine distances between the vector with the vectors of those touch gesture traces. For brevity, we shall call the touch gesture traces from all the candidate games as candidate touch gesture traces. Subsequently, each trace from GX has an associated rank list of candidate touch gesture traces. We summarize the rank lists by using the Borda count method [13], which is a consensus-based election method. Given a rank list with n candidates, Borda count method gives each candidate a certain number of points corresponding to n+1−rank(candidate). By accumulating the points given by Borda count method, we obtain a score for each of the candidate touch gesture traces. Next, to obtain the score for every candidate game, we average the scores of the candidate touch gesture traces which are associated with the same candidate game. The candidate game Y which has the highest average score is considered the game with the most similar gameplay with the game X and therefore recommended to the users.

6  Performance Evaluation

To evaluate the performance of the proposed recommendation system, in this Section, we first investigate the consistency when touch gestures as a key trait from different subjects, and evaluate the recommendation accuracy of similar games.

6.1  Consistency Check

Table 6: Consistency Check with Position Checked
Genre Game Average position Relative position†
Temple 2.75 91%
Pitfall 1.33 98%
Subway 1.25 98%
Gran 1.00 100%
Fred 1.00 100%
Trigger 3.33 88%
CS 1.20 99%
Strike 2.00 95%
Gun 6.50 73%
Cytus 9.50 59%
Tap 3.80 86%
Dance 2.40 93%
Hero 1.40 98%
iFighter 1.20 99%
Galaxy 2.20 94%
Raiden 4.25 84%
Laser 1.00 100%
AirAttack 2.20 94%
Jewels 1.75 96%
Diamond 1.40 98%
Blink 7.25 70%
Marble 9.00 61%
l- 3.1 90.1%
† The percentage of average ranks.
In this section, we want to support the robustness that gesture traces can serve signatures as the key traits of the touch gestures from different subject. We evaluate the average similarity ranks of traces which are the signatures associated with the same game. The result is summarized by games in Table VI. The third column presents the average of similarity ranking position from any other traces associated with the same game. In most games, the relative ranks are above 90%, which indicate that the gesture traces can serve sturdy signatures of gameplay in these games. However, we note that there are four games below 80% in the relative ranks, namely, Cytus, Gun, Blink, and Marble. To give a close look, we find out players in these games may display different ways of control from subject to subject. For instance, the gesture traces of Cytus, which we have discussed in Section  IV-B, are much far away in Figure 1. As another instance, the game Gun also has 2 dissimilar gesture traces. We find out the reason is that Gun supports two control methods in that players can attack the enemies by tapping directly to them or sliding the sight to shoot. In sum, most cases support the robustness of touch gestures as signatures except a few exceptions.

6.2  Games Recommendation Accuracy

Figure 2: Recommended games by target games
With a different point of view from above, we evaluate the accuracy of the proposed algorithm. We defined the accuracy is the proportions of the recommendation games which are in the same genre with the target game. Figure 2 is a diagram to show the first three of recommendation list for games. In the figure, most points stay in the diagonal, which means the first three games recommended by our system are as same genres as the target games in most cases. Table VII, which has the specific list and defined accuracy, also shows similar results. However, according to the table we can find out there are still some recommendation lists with unsatisfactory accuracy, which may be caused by the following reasons:
  1. As the recommendation system is based on touch gestures, the difference of control is the basis of low-accuracy even if the gameplay looks similar. Fred is the explicit case we discuss in Section  IV-B, which is dissimilar among other endless running games because of the different ways of control. This also explains the reason why the system recommended Fred with all tile matching games, which elicit lots of tap actions.
  2. Games with more versatile control methods may cause lower recommendation accuracy. Cytus and Gun we listed in Section  VI-A are suitable instances that players can display very different gameplay even if they are playing the same game.
  3. The lacks of games in certain genres also degrade the accuracy. In the rhythm games, there are only three rhythm games if we do not take Cytus into consideration. Therefore, the recommendation system have no choice but recommend the 3rd game which is not in the rhythm genre. Same reason also happened in Tile Matching games. For specifically, players use tap action to clear tiles in Jewels and Diamond. Blink, on the other side, tap twice at a time to clear tiles. Marble, unlike all above games, use tap action to shoot tiles. This explains why the recommendation performance is lower in rhythm and tile matching genres.

7  Conclusion

In this paper, we propose a mobile game recommendation system based on players' touch gestures while playing a game. We consider that the touch gestures as a signature of a particular game and recommend similar games based on the signature. Furthermore, the evaluation results show that 1) touch gestures can serve robust signatures of gameplay as the key traits of the touch gestures from different subjects remain consistent; 2) our recommendation system can give reasonably accurate recommendations of similar games simply based on touch gestures. We also hope that this work will motivate further research by showing that touch gestures are capable of describing gameplay and can serve as a robust signature for mobile game recommendation.
Table 7: Recommendation List
Target game 1st 2nd 3rd Accuracy
Temple Subway Pitfall Gran 100%
Pitfall Subway Temple Gran 100%
Subway Temple Pitfall Gran 100%
Gran Pitfall Subway Temple 100%
Fred Diamond Jewels Marble 0%
Trigger CS Strike Gun 100%
CS Strike Trigger Gun 100%
Strike CS Trigger Gun 100%
Gun CS Strike Raiden 67%
Cytus Trigger CS Strike 0%
Tap Dance Guitar CS 67%
Dance Guitar Diamond Tap 67%
Guitar Tap Dance CS 67%
iFighter Galaxy AirAttack Laser 100%
Galaxy iFighter Raiden AirAttack 100%
Raiden iFighter Laser AirAttack 100%
Laser iFighter AirAttack Raiden 100%
AirAttack iFighter Laser Galaxy 100%
Jewels Diamond Fred Marble 67%
Diamond Jewels Fred Marble 67%
Blink Diamond Jewels Strike 67%
Marble Diamond Jewels Fred 67%
Overall - - - 79%


[1] C. Davidsson and S. Moritz, "Utilizing implicit feedback and context to recommend mobile applications from first use," in Proceedings of the 2011 Workshop on Context-awareness in Retrieval and Recommendation, ser. CaRR '11, 2011, pp. 19-22.
[2] M.-H. Park, J.-H. Hong, and S.-B. Cho, "Location-based recommendation system using bayesian user’s preference model in mobile devices," in Ubiquitous Intelligence and Computing.    Springer, 2007, pp. 1130-1139.
[3] B. N. Miller, I. Albert, S. K. Lam, J. A. Konstan, and J. Riedl, "Movielens unplugged: experiences with an occasionally connected recommender system," in Proceedings of the 8th international conference on Intelligent user interfaces, ser. IUI '03, 2003, pp. 263-266.
[4] G. Tewari, J. Youll, and P. Maes, "Personalized location-based brokering using an agent-based intermediary architecture," Decis. Support Syst., vol. 34, no. 2, pp. 127-137, Jan. 2003.
[5] H.-W. Tung and V.-W. Soo, "A personalized restaurant recommender agent for mobile e-service," in IEEE International Conference on on e-Technology, e-Commerce and e-Service, 2004, pp. 259-262.
[6] I. Lee, J. Kim, and J. Kim, "Use contexts for the mobile internet: A longitudinal study monitoring actual use of mobile internet services," International Journal of Human-Computer Interaction, vol. 18, no. 3, pp. 269-292, 2005.
[7] T. H. Apperley, "Genre and game studies: toward a critical approach to video game genres," Simul. Gaming, vol. 37, no. 1, pp. 6-23, Mar. 2006.
[8] S. Genvo, Le game design de jeux vidéo: Approches de l'expression vidéoludique.    l'Harmattan, 2006.
[9] M. J. Wolf, "Genre and the video game," The medium of the video game, pp. 113-134, 2002.
[10] D. Djaouti, J. Alvarez, J.-P. Jessel, G. Methel, and P. Molinier, "A gameplay definition through videogame classification," Int. J. Comput. Games Technol., vol. 2008, pp. 4:1-4:7, Jan. 2008.
[11] J. R. Bray and J. T. Curtis, "An ordination of the upland forest communities of southern wisconsin," Ecological monographs, vol. 27, no. 4, pp. 325-349, 1957.
[12] J. B. Tenenbaum, V. d. Silva, and J. C. Langford, "A global geometric framework for nonlinear dimensionality reduction," Science, vol. 290, no. 5500, pp. 2319-2323, 2000.
[13] M. van Erp and L. Schomaker, "Variants of the borda count method for combining ranked classifier hypotheses," in Proceedings of the Seventh International Workshop on Frontiers in Handwriting Recognition, 2000, pp. 443-452.


1. This work was supported in part by the National Science Council under the grant NSC100-2628-E-001-002-MY3.
4. Gameplay is the specific way in which players interact with a game.
5. This definition of taps is purposely designed to take account of position inaccuracies due to touch screen.
6. We choose the 25% quantile of the gesture rates of all the segments in this work.

Sheng-Wei Chen (also known as Kuan-Ta Chen) 
Last Update September 19, 2017