You are on page 1of 8

GAMESIFT: BUILDING A VIDEO GAMES SEARCH AND RECOMMENDATIONS TOOL BASED ON A DOMAIN ONTOLOGY

Nicolas Pajot
GameSift Enterprises 1-1265 St-Joseph E, Montral, QC, Canada, H2J 1L9

Claudine Bonneau
Universit du Qubec Montral Case postale 8888, succursale Centre-Ville, Montral, QC, Canada, H3C 3P8

ABSTRACT The currently available frameworks for describing and classifying video games rely mainly on game studies hierarchies of concepts focusing on specific research interests. These methods are not taking into account the multiplicity of internal and external video games components and their relationships, and can not be used to establish similarity or resemblance between games. The existing search and recommendation tools use either imprecise genre-based categories or gamers' purchase and preference behaviors statistics. This article describes the creation of a domain ontology and search and recommendation web-based prototype. We demonstrate that this application affords new ways for documenting and exploring video game knowledge and provides more accurate results than the existing tools. KEYWORDS Video games, ontology, tool, search, recommendation, comparison.

1. INTRODUCTION
By reviewing the already existing classification methods used in game studies and game industry (see following section), we found that we clearly lack a data-model allowing for the description of a multiplicity of meaningful concepts and attributes. Such domain ontology is required to organize video game knowledge, establish relationship between concepts and find similarity between games. Since our goal was to develop search and recommendations capabilities, we needed to find a way to classify a large amount of qualitative human knowledge about video games and translate it in a machine-processable format. The ability to document video game culture with precision and compare games is of obvious for web users and gamers. Considering the tremendous popularity of video games, there is a growing interest for tools powerful enough to allow for an efficient search through the large amount of knowledge available in this domain. There is also a commercial utility for games distributors and retailers in providing their customers with better search experience and recommendations based on their needs and preferences. Constant interaction with game development professionals as well as retailers in North America led us to believe that there is a need for such a system. For instance, a large development studio in Montreal highlighted the possible use of GameSift as a pre-production tool to accelerate research on game design mechanics. When we reviewed some of the most popular search and recommendation tools currently available on the Web, we found that they are either based on imprecise genre-based categories ("industry typology") or gamers' purchase and preference behaviors statistics ("social recommendation tools"). The problem is twofold. First, the existing classifications methods cannot be used to build an efficient search and recommendations tool. Second, the existing games recommendations tools are not based on meaningful ontology. Therefore, there was a need to create new domain ontology, and then use it to build a search and recommendations tool.

2. RELATED WORK
In the following sections, we first review currently existing classification methods and identify their strengths and weaknesses. Then, we explore some of the most popular search and recommendations tools currently available on the Internet and explain why they often return imprecise or unreliable results.

2.1 Classification methods


The general problem we found in the existing classification methods is that they tend to focus on specific objectives, either to allow scholars to study particular research questions or to provide game creators with design methods. For example, they analyze games in terms of their use of space and time, as semiotic systems, as a narrative form or in terms of sets of features in a taxonomic space. Lindley's Game Classification Space (2004) categorizes games by: game, narrative, simulation, action or virtuality. As others already pointed out (Moody and al., 2007), these kinds of classifications do not help define how users interact with the system, where the settings of the game occurs, or classify any real differences between games besides those particular characteristics. Aarseth and al. (2003) have created a model for game classification and comparison, the "multi-dimensional typology of games". The main strength of their model is that it classifies games in an analytical way, thus solves the problem with the use of arbitrary genres. However, its primary function is to serve as a framework for exploring specific research questions in the field of game studies and provide scholars and game designers with a terminology allowing them to communicate with precision. The "Interactive Digital Entertainment Unification Framework" created by Moody and al. (2007) has basically the same focus, since it is a typology incorporating elements from other taxonomies with the clear objective of guiding future academic research and design. Therefore, these models do not cover areas that can be of interest to the gamers, such as aesthetic or themes. Also, the latter is aimed at entertainment and "hedonic" devices, leaving aside the growing number of serious and educational games, as well as professional simulations. Like Zagal and al. noticed, there is a need to develop an ontological approach: "An ontology is different than a game taxonomy in that, rather than organizing games by their characteristics or elements, it is the elements themselves that are organized." (Zagal and al., 2005). They have proposed an approach to describe the "things essential to the Gameness of games". However, their game ontology has a limited depth (150 elements) and focuses on physical, structural and functional aspects of the gameplay within the game design space. For instance, it doesn't document game's theme, cultural references and aesthetics, which are important to allow users to perform search queries based on multiple interests. Furthermore, we think that an ontological approach should be used not only to describe intrinsic game components, but also to organize external elements which are useful to identify other aspects related to the video games culture, such as control devices, audience, awards, press, fan-generated content, etc.

2.2 Search and recommendations tools


Most game search sites or functionalities are keyword-based: they return results grammatically matching keywords entered and not necessarily corresponding to the actual meaning of the user's query. For example, when a user is looking for "puzzles games" in the IGN web site search engine, the results will only display games having the word "puzzle" in the actual game title, leaving aside all other puzzles games that might have been of interest to the user. Other functionalities found on the web classify games by reference to one or more genres. GameSpot features a recommendation capability labeled "Games you may like". When a user enters a game title in the search engine, the site returns related game titles. The classification method used is based on genres, which are translated into 36 categories tagged to each game, for example: first-person shooters, platformers, realtime strategy, etc. Like Elverdam and Aarseth pointed it, this type of classification is often contradictory or made up for marketing purposes: "Typically, genres describe completely orthogonal aspects of games, like mood and aesthetics (dark fantasy, time (real time), or focus (strategy). The creation of new genres and the interpretation of an individual genre is free, subject only to the intentions one might have by using them" (Aarseth and al., 2003). Therefore, search engines based solely on genres often return inaccurate results to

users looking for non-traditional games overlapping many genres (or not fitting in any commercial genre at all.) Another problem we see is that there is no way for the user to understand on which criteria the recommendations are based, or how is a genre defined. We think that the recommendations must be displayed in a transparent way, in order for the user to know exactly what are the similarities and differences between games. Such genre-based classification also lack the precision needed in the case of a search query based on another point of entry than the genre or game title. What if a user would like to get results based on a combination of specifics game elements, such as goals, rules or player role? As for social filtering tools, they record what other users buy or prefer, and then use the collected data to derive a list of recommended items based on similar purchases/preferences. The recommendations are generated haphazardly as a side-effect of consuming behaviors, and have nothing to do with the actual properties of the products recommended. For example, Amazon is an eCommerce web site featuring recommendation capabilities displaying products to their customers by comparing their profiles with other users' purchases and preferences. If a user searches for the game Rock Band, Amazon will recommend other titles bought by people who also bought Rock Band, in the section labeled "Customers Who Bought This Item Also Bought... The games suggested are often best-selling titles, which might have nothing in common with Rock Band (for example, Gears of War 2 or Call of Duty). While such systems might work well for other type of media (ex: Last.fm for music) or for purely commercial purposes (ex: Amazon), it is nearly impossible for a user to discover non-commercial or obscures games titles matching his interests, nor for a game designer to find a game matching a specific aesthetic. Like in the genre-based tools, the query input field is limited to a title or a high-level genre. Having those problems in mind, our objective was to make sure that our comparison capabilities would be based on a more descriptive model, thus avoiding recommendations based on marketing-oriented genres or users' purchases. To do so, we first worked on multi-levels domain ontology and then used it to build our search and recommendations online prototype, which is described in the next sections.

3. PROJECT OVERVIEW
GameSift is a part-human/part-machine operated search and recommendations engine for video games. Its mission is to provide meaningful answers to the following question: "what am I going to play now?" as well as remedy most of other search paradigms shortcomings described previously. Its inputs are either a game title, the expression of a "desire" to find something or a particular, precise element of the video game culture; its outputs are game titles recommendations based on proximity scores. The mission statement implicitly encompasses two slightly different aspects, i.e. find titles that precisely meet the search requirements (search), but also suggest titles that are substantially different from but still somewhat linked to the search requirements (recommendation). The latter allows users to discover new games with similar attributes. In order to achieve these goals, we created a pool of formal knowledge about what "a game is" (the ontology) and, using this knowledge, we documented the content of individual games (the attributes). The implementation of this paradigm required a custom language used to "encode" the ontological concepts, an "engine" that calculates game proximity, a set of tools to input and validate data and a web interface to provide input and output functionalities. GameSift's content is edited, updated and validated through human operation while data processing and search results generation is delegated to machines. The ontology construction is a dialectic process largely based on empirical experience (building and playing games, evaluating relevance of GameSift's search results) and theory (study of other videogame ontologies and taxonomies). At this point, the data entry was performed internally. However, once the prototype reaches a beta state, we expect the community to play an integral role in documenting and validating GameSift's ontology. Similar experiences have shown that users are willing to dedicate efforts in the tasks of ontology construction and annotation and are able to produce high-quality conceptual choices, especially if presented with proper "rewards" for their contributions, like playing a game (Siorpaes and Hepp, 2008) or achieve a special user status giving them more rights and permissions on the system. Since the classification could be used by game designers in their research phases, they could also act as experts in documenting the games they have created. Therefore, the tools used to input

and validate data were built in an open ended way, allowing users to easily modify or add concepts, without compromising the integrity of the ontology as a whole.

3.1 GameSifts ontology


GameSift's ontology is a semi-structured network of meaningful concepts, orthogonal with one another, that are used to describe a video game or any particular aspect of the videogame culture. At the largest level it can be seen as four major "families of concepts", hierarchically organized and ranging from the more micro, game specific concepts (a game's intrinsic characteristics) to the more macro, general, cultural elements (game industry, culture, press, etc.). Concepts are organized as tree structure: each leaf (a concept) has exactly one parent and may have one or more children. Branches represent families of concepts (gameplay, themes, player roles, hardware, etc.). This structure lacks a way to explicitly specify the level of proximity (and eventually the nature of the relationship) between two concepts, may they be siblings or located in a different branch of the tree. In order to circumvent this issue, we expanded the ontology into a semi-structured network by introducing the idea of "proximity" or "friend-ness" between concepts, a measure of "how close" two concepts are. Following the proximity links when comparing a game with another allows for a greater variety of game titles (albeit of less "relevance") to be listed as results. For example, we could link the gameplay mechanic Driving a Car to other gameplay mechanics (piloting a plane and riding an animal, respectively 30% and 20% similar to driving a car) as well as to concepts located in other branches of the tree: the theme toys and gadgets > cars (50% proximity) or the genre Formula 1 driving game (100% proximity). One can easily understand that piloting a plane and driving a car are related behaviors. Therefore the proximity value is largely based on common sense and can be refined later by blind evaluators as it is often the case in the field of artificial intelligence. An example of such project is ConceptNet 3 at MIT Media Lab. In another example of linking, the "First Person Shooting" genre could be linked to the Grand Theft Auto Clone genre because both feature a FPS-like experience, although GTA also features other gameplay mechanics. We are currently expanding this "linkage mechanism" by allowing a "meaning" to be attached to the relationship between concepts through the use of meta-concepts that semantically describe the nature of the relationship. Linkage concepts include similar to, explains, belongs to, works for, etc. In terms of implementation, each concept is encoded as a unique short word (a mnemonic), to which is attached a description and a type: it can be a binary digit (presence/absence) or a 5-level quantifier (0 - 4 inclusively). At the current time, the ontology uses more that 1,300 individual concepts.

3.2 Search and recommendations capabilities implementation


In the following sections, we describe how GameSift's ontology is implemented as a base for the search and recommendations prototype.

3.2.1 Proximity engine


The proximity engine is an offline application that incrementally compares each game with any other in the database, calculates the difference between them and outputs results to a cache database that stores all relationships between games for fast retrieval. The comparison process itself is based on the concepts that are shared between the two games, those that are missing in either one or the other, the possible replacement candidates (friend concepts) for each missing concept and the relative weight of each concept.

3.2.2 Web-based search prototype


The web-based search prototype provides three types of search: -Concept search: the most basic, precise search. It searches for presence of a concept or a group of concepts in games (example: Drive (terrestrial) (gameplay), PlayStation 2 (platform), poverty (theme)). -Quick searches based on particular "meaning groups". It searches for groups of weighted concepts that define certain meanings, like moods, personal objectives or desires. The prototype implements five different types of quick searches:

-Emotions (I feel relaxed, sadistic, weird, industrious, social); -Personal objectives (I want to develop my brain, to impress my friends, to interact with people); -Quick Play (a free strategy game, a realistic flight simulator, a game without violence); -Tech Marvels (Top graphic rendering, amazing physics, never-heard before sound, fantastic AI); -Community built searches (search groups created by the community, including greek mythology, save the girl, current social issues, Japan); It is easy to imagine more quick searches based on various interests (geography, aesthetic values, history, narrative content and characters, etc.) -Game proximity: the most detailed comparison. It searches for games that are closer to a given title by using the concept-to-concept comparison mechanism: In the current implementation, comparison algorithms largely focuses on gameplay mechanics, but also factors in thematic elements (other elements (technology, aesthetics, etc.) are currently ignored). This search could also allow multiple titles entry: in this case, the engine would first create a hypothetical game corresponding to the union of the input titles and then perform search. Results would be games that most closely match this game. The web implementation is based on ASP.Net 2.0 and was written in C#. Other technologies used are DHTML and AJAX. Search parameters are presented in the leftmost pane, while most of the screen is dedicated to the results. A top bar acts as a breadcrumb widget and reminds the user of the parameters of the current search. Results are either presented as a list (game title, an embedded video feed of the gameplay, a quick description, a box shot and a proximity score) or as a grid (game title, a box shot, and proximity score).

Figure 1. Two different layouts are available to display results: the left screenshot is showing results presented as a list, while the right screenshot is showing results presented as a grid

Contextual menus appear when hovering the mouse over the proximity score or clicking on Play Now. The former displays a synthetic view of the content of the game (see Figure 2) while the latter displays locations where to play, buy, rent or download the game. Finally, each game is documented in an unbiased (ie: no opinion, strong focus on facts), encyclopedic way.

4. PRELIMINARY RESULTS
The current prototype still has limited scalability and hasn't been made available to the general public yet. However, a preliminary evaluation was conducted based on testing performed internally by the production team. Three well known sites were selected to compare our results to: a hobbyist game database (MobyGames), a general public review site (GameSpot) and an e-commerce site (Amazon.com). These sites cover a large spectrum of target audiences, user experiences and search/recommendation functionalities (see Table 1). It is important to note that unfortunately, in most cases, direct comparison between GameSift and the aforementioned sites is impossible because the search request is difficult (if not impossible) to replicate: the sites lack the necessary search points of entry (most sites only allowing a title input and a small set of criteria, usually limited to a handful of concepts, such as platforms, publishers and genre, when, GameSifts search exposes hundreds of search criteria). We couldnt either rely on the tag (or group) systems these sites provide. Indeed, tags only provide a subset of the search functionality needed to unequivocally pinpoint certain game components (namely, the ability to attach some level of meaning to a game title). Tag systems implemented on these sites featured a limited number of tags (less than what would be required for

meaningful searches). Moreover the lack of formal tag organization (tags are not orthogonal with one another and are most often loosely defined), and the absence of linkage mechanisms (no way to link poverty with social concern for instance) prevented any meaningful game to game comparison.
Table 2. Selected web sites for comparison Web Site GameSpot Search / Recommendation Features Word search (title and other text content) Search Tag System Genre, Platform Recommendation "Similar Games" List Word search (title and other text content) Themes Search Genre, Platform, Release Year, Rating Systems, Companies (publishers, developers) Tag System (named "Groups") Recommendation No explicit "Similar Games" List Word search Platform, ESRB rating, "Brand" (mostly Search publishers", Customer Review, Price, Seller, Availability Tag System Recommendation "Recommended Games" List URL www.gamespot.com

MobyGames

www.mobygames.com

Amazon

www.amazon.com

4.1 Concept search


This section of the prototype allows a user to search for videogames that feature a specific concept or group of concepts. In the example below, we search for a game that is based around driving a car, that runs on Sonys Playstation 2 and that features the theme of Poverty. The user selects concepts in a tree view (that duplicates the tree structure of the ontology and exposes concepts) and clicks on the Sift! button. This request returns two results: Grand Theft Auto: San Andreas and Mafia: the City of Lost Heaven. Those two games are the only ones in GameSifts database that feature all requested concepts: core open space driving gameplay mechanics, poverty theme and the Playstation 2 platform. Although being the most straightforward, simple search functionality our engine provides, it proved impossible to replicate on any of the three selected comparison web sites. None of the sites allowed us to select the necessary criteria and only provided partial answers to the search request. Platform and core gameplay mechanic were easy to select, but social concerns are seldom documented. In fact, none of the tagging systems had a poverty tag.

4.2 Quick search


The end-user can search for games that fit a particular player's mood (ex: sadistic, relaxed), highlight certain elements (ex: social issues) or fit certain arbitrary criteria (ex: I want to impress my friends). The search groups are created by editors or by the community and are selected by the user in a similar way to concept search. In the example below, we search for a game that would match a mood, which was chosen for the purpose of this test to be sadistic.
Table 2. Results for Sadistic Mood Search Title Manhunt Ninja Gaiden 2 Truck Dismount God of War 2 God of War: Chains of Olympus Mortal Kombat 4 Condemned 2: Bloodshot Relevance Score 100% 74% 60% 39% 39% 36% 34%

Results show that games ranking high (over 50%) all exhibit sadistic gameplays, themes or artistic treatment in one way or another: Manhunt is a game in which the protagonists performs in a snuff movie, killing random targets in the most gruesome ways; Ninja Gaiden 2 is a fighting game with strong gore elements (dismemberment, high quantity of blood); Truck Dismount is a turn-based, physics driven game in which you have to inflict as much damage as possible to a crash test dummy sitting in or on a rolling truck. Games ranking fairly high (over 30%) exhibited themes closely related to sadism (blood and gore, extreme violence, etc.). Unfortunately, none of the three reference web sites provided quick searches, rendering mood searches (and other group searches) impossible to replicate.

4.3 Game proximity search


The end-user can search for games that are similar to a given title (proximity can be defined as proximity of gameplay mechanics, theme, goals, etc.). This is the recommendation part of the GameSift engine. The game Grand Theft Auto: San Andreas was chosen for the purpose of this test not only because it is rather well known (allowing us to compare results with the three other web sites) but also because it features an interesting mix of core gameplays (open space roaming / driving / riding, shooting, role playing, dancing, etc.) along with a strong theme (gangsterism).
Table 3. Results for Grand Theft Auto: San Andreas Title Grand Theft Auto: Vice City Stories Grand Theft Auto 4 Saints Row Mafia: The City of Lost Heaven True Crime: Streets of New York True Crime: Streets of LA Grand Theft Auto 2 Grand Theft Auto 1 Crackdown Unreal Tournament 3 Crysis Halo 3 Wild Metal Proximity Score 91% 89% 84% 82% 78% 77% 68% 68% 58% 47% 35% 28% 11% Significative differences (ie: no RPG elements, etc.) but same overall gameplay and themes Comments

Minor differences (ie: no "dancing" mechanic, etc.)

Major differences in gameplay and themes

Games ranking over 80% are almost exact clones of Grand Theft Auto: San Andreas; obvious sequels/prequels show up in the first rows of the results table. The minor score differences between these games can be explained by the absence or presence of minor, secondary gameplay mechanics (ex: dancing in Grand Theft Auto 4). It is worth noting that Grand Theft Auto III does not show up in the results because it is not documented in GameSifts database. Games ranking below 80% and over 60% are usually pretty similar to the target game but lack some more important aspects: in our case, no role playing elements and a stronger focus on driving or shooting explain the score differences. True Crime, Grand Theft Auto 1 & 2 and Crackdown can all be considered less elaborate versions of Grand Theft Auto: San Andreas. Games ranking below 60% most often prominently feature only one aspect of the gameplay mix found in Grand Theft Auto. Unreal Tournament 3 allows the user to drive vehicles, but the core gameplay is clearly more of the First Person Shooter type. The same is true of Crysis and Halo 3. We ran the same search query on each of the three reference web sites. GameSpots recommended games for Grand Theft Auto: San Andreas returned two Grand Theft Auto games at the top spots (Vice City and III). No other Grand Theft Auto-like game was featured in the list. Instead, stealth action games like Hitman, Metal Gear Solid and Splinter Cell were presented. These games feature different gameplay mechanics than the ones found in the Grand Theft Auto series: no open space driving / shooting and no role playing elements, and a strong focus on infiltration and killing targets without being seen. Lastly, only Hitman exploits a gangster/mafia theme; Splinter Cell and Metal Gear Solids themes range from military / special forces to science fiction.

Amazon search also returns Grand Theft Auto III and Vice City as top results. The two other recommendations are God of War and God of War 2, which are, again, substantially different from Grand Theft Auto: San Andreas. They do not feature roaming / driving / shooting and emphasize melee combat with blade weapons. They both feature very strong ancient Greece themes but none of the gangster / mafia themes found in Grand Theft Auto games. MobyGames does not return any similar title. Instead, it displays a list of groups (similar to tags) that, although interesting in their own right, did not help much in finding overall similar games to Grand Theft Auto: San Andreas. The groups were a mish-mash of themes (black protagonists, ganster protagonists), gameplay elements (Simon says games), technology (Graphics Engine: RenderWare) and commercial data (Playstation 2 greatest hits releases). These findings tend to show that the recommendation engines of these three sites are crippled with false positive results (God of War, Splinter Cell) and false negative results (no trace of True Crime, Mafia, Saints Row or even the other GTA games in the series). Since the database is currently limited in scope, we deliberately chose to work on certain game types. Examples presented in this paper clearly illustrate of the kind of results one can expect from using GameSifts engine.

5. CONCLUSION
In this article, we described the GameSift's ontology, currently consisting of more than 1,300 individual concepts. We have presented a web-based prototype allowing users to perform searches using various points of entry in order to get recommendations based on their specific interests. As the refinement of the web interface and comparison algorithms will go on, it seems fair to assume that our prototype will soon reach a beta stage that could be released to the public. Our initial tests provide preliminary evidence that the ontology can not only serves as the basis of GameSift's search and recommendations capabilities, but could also be made available as a web service (API). There are some issues we would like to address in future version of the search engine. First, we will build a more formal unit test bench to ensure better validation. We also hope to get better results by working on expressing the "complexity" factor of a game and get better search input by parsing search text and linking words and concepts. We will also work on improving the user experience by running usability testing on the user interface and refining the pages layout and navigation.

ACKNOWLEDGEMENT
The authors would like to thank the following people who have contributed to this project over the last year and its many incarnations: Carolyn J. Tice, Lee Tipton and Gokhan Calislar. The work presented has been funded by private investments. We also thank the anonymous evaluators for their helpful comments.

REFERENCES
Aarseth, E., Smedstad, S., Sunnana, L., 2003, A multi-dimensional typology of games. Level Up: Digital Games Research Conference Proceedings. Utrecht, The Netherlands, pp. 30-45. Elverdam, C., Aarseth, E., 2007. Game Classification and Game Design: Construction Through Critical Analysis. In Games and Culture, Vol. 2, No. 1, pp. 3-22. Lindley, C., 2004, Trans-reality gaming. Proceedings of the Second Annual International Workshop in Computer Games Design and Technology. Liverpool, UK, pp. 1-10. Moody, G., Wells, T., Lowry, P. B, 2007, The Interactive Digital Entertainment (IDE) Unification Framework: Creating a Taxonomy of IDE and Lifestyle Computing. 40th Annual Hawaii International Conference on System Sciences. Hawaii, pp. 160a. Siorpaes, K., Hepp, M., 2008, OntoGame: Weaving the Semantic Web by Online Games. The Semantic Web: Research and Applications. Springer Berlin / Heidelberg, pp. 751-766 Zagal, J. P., Mateas, M., Fernandez-Vara, C., Hochhalter, B., Lichti, N., 2005, Towards an Ontological Language for Game Analysis. Proceedings of DiGRA 2005 Conference: Changing Views - Worlds in Play. Vancouver, Canada.

You might also like