4. Conclusion
The aim of this project was to build a detector able to predict the likelihood for a player to finish a game or not, and we proposed a simple logistic regression model with accaptable performance. The goal of this study was not to establish a highly sophisticated state-of-the-art machine learning model with best prediction performance. Instead, it rather serves as a proof-of-concept for leveraging and identifying potential features for predicting the success of a player's game. Under the assumption that unsuccesful games demotivate a player to stay engaged, we here propose a first building block to understanding why player are not able to finish a game in the first place. As large datasets play a crucial role in data analysis, since they provide a greater sample size and allow for more robust and accurate statistical analyses, having an active and motivated player base can therefore boost research purposes.
In our study we focused on the Wikispeedia game. We observed that in total there are about 75% of the players that finish all the time or that finish partially. This means that about 25% percent of all the players never finished, which is a huge loss of data. Results showed that they also have played less games. It seems that they got more quickly demotivated and therefore stopped playing the game. In addition, of the remaining 75%, we found some players that did not return after having had an unsuccesful game. This suggests that on top of the 25%, there are more players that got demotivated to continue playing.
In our analysis to identify factors driving this behavior, we found 6 potential predictors: concept of starting page - concept of target space - number of unfinished games in the past - most recent streak of unfinished games - time difference between first and last game and the largest consecutive finished streak per player. Among these features, we found that the latter has the largest magnitude and thus strongly correlates with having a finished game for a player.
This information can be leveraged for further attempts to keep players engaged. We therefore propose a set-up where a detector (as provided in this study) is used as a tool to keep track of the likelihood of the player to finish a game. For example when a player has had 3 unfinished games in a row already, the Wikispeedia engine could provide the player with a 'easy' game to help the player out. This 'easy' game could be derived from the features that would predict a finished game. This could be an easy concept as target page. A limitation of the current study is that the features heaviliy rely on the history of the player only. We only considered starting page and target page of the current game to influence the prediction. This makes it difficult for the detector to dynamically adapt its prediction during the game as the players travels in Wikipedia space. This also limits the oppertunity to provide adequate hints during the game. We briefly proposed, by tracking the progress score, a potential predictor that could fill this gap. However, the results were not too convincing. More thorough analysis (using more data potentially) should help in further deciphering possible in game features. Not only should this provide information for the Wikispeedia engine on how to intervene, it also helps in finding the right moment on when to intervene. The found path length of 4-5, where the progress score plateaus, could be a promising moment in time to do so.
Overall we provided an intuitive model to pave the way for further analysis in order to accommodate an activate player community. The resulting findings are not only relevant for Wikispeedia, but also for other games that could benefit from motivated players in order to boost the generation of relevant data.