Multimedia data have become predominant in several application domains. To cope with this huge amount of data, recommender systems come in handy by delivering valuable items to end-users. In particular, existing PC games stores could be supported by such systems. Even though the user can easily browse such platforms to find desired items, there is still room for improvements, i.e., enhancing available tagging systems, adopting new analysis techniques. With the aim of assisting users in the definition of the tags, we propose AURYGA, an AUtomated Recommender sYstem that exploits a well-founded stochastic model to predict GAme tags using the textual description of PC games stored on the Steam platform. First, we filter the initial set of tags by using the frequent itemset technique. Afterwards, we use a TF-IDF encoder on textual descriptions to feed the underpinning model to retrieve the list of tags. Due to the lack of a proper baseline, we evaluate AURYGA by means of a 10-fold cross validation on a real-world dataset. The initial results show that the proposed prototype represents a workable solution in the domain.
To properly manage information overload yielded by multimedia data [
To this end, we propose AURYGA, a novel recommender system based on a Complement Naive Bayesian (CNB) network to retrieve a set of tags given the textual description of the game. First, we empirically select 7 main game genres and extract frequent sub-tags patterns using the FP-growth mining algorithm. Then, the textual content is encoded by relying on a TF-IDF module. The produced vectors are eventually used by AURYGA to predict a final set of tags. As a preliminary evaluation, we used a Steam dataset composed of PC games with the corresponding tags. We conduct a 10-fold cross evaluation by computing diferent metrics, i.e., success rate, precision, recall, and F-measure. The initial results show that AURYGA is capable of recommending the correct tags most of the time. We make available the initial implementation as well as the dataset to replicate the experiments in the corresponding Github repository1.
In this section, we present an explanatory example in Figure 1 to highlight our contribution. The
left side of the picture shows Syberia, a point-and-click adventure game 2 and its user-defined
tags ranked according to their frequency, i.e., the number of users that define that tag for
the game. Although the top-3 tags are correct, i.e., Adventure, Point & Click and Puzzle the
game is labeled with ones that represent game’s features rather than the actual genre [
Based on this list, Steam suggests additional games to purchase as shown on the right side of the picture, i.e., More like this section. By carefully inspecting three of them, namely Red Dead Redemption II, Shadow of the Tomb Raider, and Half-Life Alyx, we discover that the recommended items belong to completely diferent genres i.e., western, purely action-adventure, and firstperson shooter3 respectively. Thus, inaccurate labeling can compromise the visibility of games as the platform groups them according to the user-based defined tags.
The proposed approach and its main components are shown in Fig.2. As the first step, we empirically select 7 main tags from the most famous genres 4 by eliciting the ones with higher 1https://github.com/Arkanoid01/Auryga 2https://fr.flossmanuals.net/creating-point-and-click-games-with-escoria/what-is-point-and-click-games/ 3https://www.pcmag.com/encyclopedia/term/first-person-shooter 4https://www.statista.com/statistics/189592/breakdown-of-us-video-game-sales-2009-by-genre/
Predicted tags Filtering
Filtered Train Data
CNB
Main Genre Steam Data
5
Untagged game
frequencies in the considered dataset, i.e., Action, Adventure, Rpg, Platform, Sport, Strategy, and
Simulation. On top of these main tags, the raw S t e a m d a t a 1 labeled with 370 unique game
tags is used by the F i l t e r i n g component 2 to retrieve the most frequent sub-patterns using
the FP-growth, a well-founded mining algorithm [
Afterwards, AURYGA employs 4 C o m p l e m e n t N a i v e B a y e s ( C N B ) network 6, an enhanced version of Multinomial Naive Bayes (MNB) that cope with unbalanced dataset [16]. Diferently from the MNB, CNB computes the probability that an item appears in the other classes and normalizes classification weights to properly manage word occurrence dependencies. When an u n t a g g e d g a m e arrives 5 , the textual description is encoded using the TF-IDF vectorizer 7. First, the underpinning model predicts the m a i n g e n r e 6 by relying on the original Steam data. Second, the CNB exploits the abovementioned filtered training sets to retrieve sub-tags when the main one is correctly predicted. The system eventually retrieves the p r e d i c t e d t a g s 7 that include the main genre and the most probable sub-tags. 3.1. Preliminary evaluation To assess the conceived approach, we conduct a 10-fold cross validation [17] on the mentioned Steam dataset composed of 27,334 games gathered in 2019 8. Each game is identified by a 5http://rasbt.github.io/mlxtend/api_subpackages/mlxtend.frequent_patterns/#fpgrowth 6https://scikit-learn.org/stable/modules/generated/sklearn.naive_bayes.ComplementNB.html#sklearn.naive_ bayes.ComplementNB 7https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.TfidfVectorizer.html 8https://www.kaggle.com/nikdavis/steam-store-games unique ID, the title, the description, and the list of tags. For each round, we used 90% of the descriptions as training and 10% as testing to compute Success rate, Precision, Recall, and FMeasure, considering the following definitions: a True positive (Tp) is a recommended tag that corresponds to the actual game tag. A false positive (FP) means that the suggested tag is not present in the actual list while a false negative (Fn) represents a tag that should be present in the result set but actually is not. Given a set of N recommended tags, the Success rate measures the rate at which the proposed approach can return at least c correct tags for c=1,2,3,4,5.
Success rate = =
+ (2) ∈ ( > 0) | |
× 100% =
+ (3) (1) − = 2 × × (4) +
Precision and Recall assess the rate of correct items over the entire set of retrieved items and the ratio of the right topics appearing in the retrieved set respectively. F-measure represents the harmonic mean of the two previous metrics. All the mentioned indexes are computed by considering the filtered set of tags, i.e., the ones obtained by the F i l t e r i n g component.
Figure 3 shows the results obtained in terms of success rate. As we can see, AURYGA is capable of suggesting at least the first tag with good results, i.e., it reaches 98% with N=8. As expected, increasing the number of correct tags c impacts negatively on the performances. For instance, the success rate dramatically decreases if we consider c=5, i.e., the system recommends exactly five correct tags only in the 0,3% of the cases.
100 80 ) % t(ea 60 r s s e c c Su 40 20 0 93:2 45 13:4 N3
C=1 C=2 C=3 C=4 C=5 96
98 57:8 28:3 12
0:3
N5 Recommended Set 66:3 40:1 N8 23:7 12:6
To further study the AURYGA’s performances, we compute additional metrics, i.e., precision, recall, and F-measure showed in Table 1. Overall, increasing the number of recommended items leads to low results on average i.e., the maximum F-measure value is reached with N=3. In particular, the precision decrease from 50.01% to 31.2% by increasing the number of retrieved items. Contrariwise, augmenting N yields to better performance in terms of recall, i.e., it achieves 67.2% with N=8. This is quite expected since the system retrieves more tags that should be included. It is worth noting that we aim to recommend 100 diferent unique tags, considering both the main and sub-tags. It is our strong belief that reducing the number of tags in the training will lead to better accuracy since this strategy has been proved successful in existing studies [18].
The first taxonomy of games was proposed in [ 19] by manually inspecting their features e.g., genre, the number of players, description. A curated discussion over games’ genres has been conducted in [20] by considering several aspects i.e., family resemblances, prototype theory, faceted classification, and appeal factors. Horppu et al. [21] make use of the SVM model to classify 2,443 games on the IoS App store by using title and description. Similarly, Amiriparian et al. [22] propose an audio-based videogame genre categorization using linear SVM and CNN models on a dataset composed of 1,566 multimedia content including in-game video, audio, and visual elements to categorize games into six categories. A multi-modal deep neural network [23] trained with the cover image and the textual information has been employed to classify 50,000 PC games stored on IGDB.com website by predicting 21 unique tags.
Videogame online stores manage hundred of PC games usually labeled with a list of tags. Even though such a tagging system is successfully employed by several platforms, increasing the accuracy of the defined categories is still an open challenge. To fill the gap, we propose AURYGA, a recommender system that automatically retrieves a set of game tags by analyzing the textual description of the content. The system combines a well-founded frequent itemset algorithm and a stochastic model to recommend the main genre and its most frequent sub-tags. The preliminary evaluation shows encouraging results in terms of predicted tags even though some metrics achieve low values.
For future works, we can feed the underpinning algorithm with additional textual data, i.e., game title, developer, reviews. Furthermore, we can enhance the preprocessing phase by tuning the filtering component. Last but not least, we are going to conduct a controlled user study in which the participants evaluate the system qualitatively. 1 1 4 5 / 1 0 1 4 0 5 2 . 1 0 1 4 1 4 1 . [16] J. D. M. Rennie, L. Shih, J. Teevan, D. R. Karger, Tackling the poor assumptions of naive bayes text classifiers, in: Proceedings of the Twentieth International Conference on International Conference on Machine Learning, ICML’03, AAAI Press, 2003, p. 616–623. [17] R. Kohavi, A study of cross-validation and bootstrap for accuracy estimation and model selection, in: Proceedings of the 14th International Joint Conference on Artificial Intelligence - Volume 2, IJCAI’95, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1995, pp. 1137–1143. URL: http://dl.acm.org/citation.cfm?id=1643031.1643047. [18] W. Siblini, P. Kuntz, F. Meyer, A review on dimensionality reduction for multi-label classification, IEEE Transactions on Knowledge and Data Engineering 33 (2021) 839–857. doi:1 0 . 1 1 0 9 / T K D E . 2 0 1 9 . 2 9 4 0 0 1 4 . [19] K. J. H.G., The gaming landscape: a taxonomy for classifying games and simulations, in: DiGRA ༿ - Proceedings of the 2003 DiGRA International Conference: Level Up, 2003.
URL: http://www.digra.org/wp-content/uploads/digital-library/05163.55012.pdf. [20] R. I. Clarke, J. H. Lee, N. Clark, Why Video Game Genres Fail: A Classificatory Analysis, Games and Culture 12 (2017) 445–465. URL: http://journals.sagepub.com/doi/10.1177/ 1555412015591900. doi:1 0 . 1 1 7 7 / 1 5 5 5 4 1 2 0 1 5 5 9 1 9 0 0 . [21] I. Horppu, A. Nikander, E. Buyukcan, J. Mäkiniemi, A. Sorkhei, F. Ayala-Gómez, Automatic classification of games using support vector machine, CoRR abs/2105.05674 (2021). URL: https://arxiv.org/abs/2105.05674. a r X i v : 2 1 0 5 . 0 5 6 7 4 . [22] S. Amiriparian, N. Cummins, M. Gerczuk, S. Pugachevskiy, S. Ottl, B. Schuller, “are you playing a shooter again?!” deep representation learning for audio-based video game genre recognition, IEEE Transactions on Games 12 (2020) 145–154. doi:1 0 . 1 1 0 9 / T G . 2 0 1 9 . 2 8 9 4 5 3 2 . [23] Y. Jiang, L. Zheng, Deep learning for video game genre classification, arXiv:2011.12143 [cs] (2020). URL: http://arxiv.org/abs/2011.12143, 00001 arXiv: 2011.12143.