Zoheb Borbora

Churn Prediction in MMORPGs Using Player Motivation Theories and an Ensemble Approach

In this paper, we investigate the problem of churn prediction in Massively multiplayer online role-playing games (MMORPGs) from a social science perspective and develop models incorporating theories of player motivation. The ability to predict player churn can be a valuable resource to game developers designing customer retention strategies. The results from our theory-driven model significantly outperform a diffusion-based churn prediction model on the same dataset. We describe the synthesis between a theory-driven approach and a data-driven approach to a problem and examine the trade-offs involved between the two approaches in terms of prediction accuracy, interpretability and model complexity. We observe that even though the theory-driven model is not as accurate as the data-driven one, the theory-driven model itself can be more interpretable to the domain experts and hence, more preferable over a complex data-driven model. We perform lift analysis of the two models and find that if a marketing effort is restricted in the number of customers it can contact, the theory-driven model would offer much better return-on-investment by identifying more customers among that restricted set who have the highest probability of churn. Finally, we use a clustering technique to partition the dataset and then build an ensemble on the partitioned dataset for better performance. Experiment results show that the ensemble performs notably better than the single classifier in terms of its recall value, which is a highly desirable property in the churn prediction problem.

Link Prediction Across Multiple Social Networks

The problem of link prediction has been studied extensively in literature. There are various versions of the link prediction problem \textit{e.g.,} link existence problem, link removal problem, predicting edge weights over time etc. In this paper we describe a new type of link prediction problem called the Inter-network link-prediction problem where the task is to predict links \textit{across} different networks. Thus given a set of nodes which participate in multiple networks the task is to determine if one can predict the edges that occur in one network by only using node attribute and edge information from other networks. We use insights from theories of evolution of social communication networks and the MTML framework to derive models which can be used to make link predictions across networks. For the experiments data from different \textit{types} of social networks from a Massively Multiplayer Online Role Playing Game (MMORPG) is used.

User Behavior Modelling Approach for Churn Prediction in Online Games

Massively Multiplayer Online Role-Playing Games (MMORPGs) are persistent virtual environments where millions of players interact in an online manner. Game logs capture player activities in great detail and user behavior modeling approaches can help to build accurate models of player behavior from these logs. We are interested in modeling player churn behavior and we use a lifecycle-based approach for this purpose. In a player lifecycle-based approach, we analyze the activity traits of churners in the weeks leading up to their point of leaving the game and compare it with the activity traits of a regular player. We identify several intuitive yet distinct behavioral profiles associated with churners and active players which can discriminate between the two populations. We use these insights to propose three semantic dimensions of engagement, enthusiasm and persistence to construct derived features. Using three session-related variables and the features derived from them, we are able to achieve good classification performance with the churn prediction models. Finally, we propose a distance-based classification scheme, which we call wClusterDist, which benefits from these distinct behavioral profiles of the two populations. Experimental results show that the proposed classification scheme is well-suited for this problem formulation and its performance is better than or comparable to other traditional classification schemes.

Bot Detection Based on Social Interactions in MMORPGs

The objective of this work is to detect the use of automated programs, known as game bots, based on social interactions in MMORPGs. Online games, especially MMORPGs, have become extremely popular among internet users in the recent years. Not only the popularity but also security threats such as the use of game bots and identity theft have grown manifold. As bot players can obtain unjustified assets without corresponding efforts, the gaming community does not allow players to use game bots. However, the task of identifying game bots is not an easy one because of the velocity and variety of their evolution in mimicking human behavior. Existing methods for detecting game bots have a few drawbacks like reducing immersion of players, low detection accuracy rate, and collision with other security programs. We propose a novel method for detecting game bots based on the fact that humans and game bots tend to form their social network in contrasting ways. In this work we focus particularly on the in game mentoring network from amongst several social networks. We construct a couple of new features based on eigenvector centrality to capture this intuition and establish their importance for detecting game bots. The results show a significant increase in the classification accuracy of various classifiers with the introduction of these features.

Exploration of Robust Features of Trust Across Multiple Social Networks

In this paper, we investigate the problem of trust formation in virtual world interaction networks. The problem is formulated as one of link prediction, intranet work and internet work, in social networks. We use two datasets to study the problem - SOEs Ever quest II MMO game dataset and IBMs Small Blue sentiments dataset. We explore features based on the nodes individual properties as well as based on the nodes location within the network. In addition, we take into account the nodes participation in other social networks within a specific prediction task. Different machine learning models built on the features are evaluated with the goal of finding a common set of features which are both robust and discriminating across the two datasets. Shortest Distance and Sum of Degree are found to be robust, discriminating features across the two datasets. Finally, based on experiment results and observations, we provide insights into the underlying online social processes. These insights can be extended to models for online social trust.

Socialization and trust formation: a mutual reinforcement? An exploratory analysis in an online virtual setting

Social interactions preceding and succeeding trust formation can be significant indicators of formation of trust in online social networks. In this research we analyze the social interaction trends that lead and follow formation of trust in these networks. This enables us to hypothesize novel theories responsible for explaining formation of trust in online social settings and provide key insights. We find that a certain level of socialization threshold needs to be met in order for trust to develop between two individuals. This threshold differs across persons and across networks. Once the trust relation has developed between a pair of characters connected by some social relation (also referred to as a character dyad), trust can be maintained with a lower rate of socialization. Our first set of experiments is the relationship prediction problem. We predict the emergence of a social relationship like grouping, mentoring and trading between two individuals over a period of time by looking at the past characteristics of the network. We find that features related to trust have very little impact on this prediction. In the final set of experiments, we predict the formation of trust between individuals by looking at the topographical and semantic social interaction features between them. We generate three semantic dimensions for this task which can be recomputed with an observed social variable (say grouping) to create a new semantic social variable. In this endeavor, we successfully show that, including features related to socialization, gives us an approximate increase of 4-9% accuracy for trust relationship predictions.

Robust features of trust in social networks

We identify robust features of trust in social networks; these are features which are discriminating yet uncorrelated and can potentially be used to predict trust formation between agents in other social networks. The features we investigate are based on an agent’s individual properties as well as those based on the agent’s location within the network. In addition, we analyze features which take into account the agent’s participation in other social interactions within the same network. Three datasets were used in our study—Sony Online Entertainment’s EverQuest II game dataset, a large email network with sentiments and the publicly available Epinions dataset. The first dataset captures activities from a complex persistent game environment characterized by several types of in-game social interactions, whereas the second dataset has anonymized information about people’s email and instant messaging communication. We formulate the problem as one of the link predictions, intranetwork and internetwork, in social networks. We first build machine learning models and then perform an ablation study to identify robust features of trust. Results indicate that shared skills and interests between two agents, their level of activity and level of expertise are the top three predictors of trust in a social network. Furthermore, if only network topology information were available, then an agent’s propensity to connect or communicate, the cosine similarity between two agents and shortest distance between them are found to be the top three predictors of trust. In our study, we have identified the generic characteristics of the networks used as well as the features investigated so that they can be used as guidelines for studying the problem of predicting trust formation in other social networks.

Guild play in MMOGs: rethinking common group dynamics models

Humans form groups and congregate into groups for a variety of reasons and in a variety of contexts e.g., corporations in offline space and guilds in Massively Multiplayer Online Games (MMOGs). In recent years a number of models of group formation have been proposed. One such model is Johnson et als [10] model of group evolution. The model is motivated by commonalities observed in evolution of street gangs in Los Angeles and guilds in an MMOG (World of Warcraft). In this paper we first apply their model to guilds in another MMOG (EQ2)1 and found results inconsistent from the models predictions, additionally we found support for the role of homophily in guild formation, which was ruled out in previous results, Alternatively, we explore alternative models for guild formation and evolution in MMOGs by modifying earlier models to account for the existence of previous relationships between people.

Automatic Detection of Compromised Accounts in MMORPGs

Account compromise can occur frequently in popular Massively Multiplayer Online Role-playing Games(MMORPGs) to gain easy profits and in extreme scenarios, it can lead to gold farming in the real world. Despite security concerns over compromised accounts in MMORPGs, few attempts have been made towards deeper exploration of the problem. Previous research has studied only the classification of cheating in online games. This paper describes a p-value based change point detection model for detecting compromised accounts in online games based on user behavior analysis. We evaluated the proposed model on real gaming datasets with ground truth. Experiment results demonstrate that our model is able to discover important features to detect compromised account.

Love all, trust a few: link prediction for trust and psycho-social factors in MMOs

Massively Multiplayer Online Games (MMOGs) where millions of people can interact with one another have been described as mirrors of human societies and offer excellent venues to analyze human behavior at both the psychological as well as the social level. Within the context of predictive analysis (link prediction as a classification task) in MMOGs, the connection between psycho-sociological theories of communication networks. A mapping of how various elements of trust and other social interactions (mentoring, adversarial relationship, trade) relate to prediction tasks is also established. Results from classification experiments indicate that social environments affect prediction tasks in cooperative vs. adversarial environments in MMOGs and the implications of these results for generalizability of link prediction algorithms is also analyzed.