conclusions.txt

In this study, we have applied exploratory analysis and machine learning techniques to online video game reviews culled from the
Giant Bomb and GameSpot websites. To gather the data we used the Giant Bomb API and wrote a custom web scraper to extract review information
directly from GameSpot HTML source. In total we analyzed more than 40000 video game reviews, from both professional critics and anonymous 
users.

We gained valuable insights about these particular data sets and the video game industry as a whole via our initial exploratory analysis of the reviews
and associated metadata. In one highlight, we found that user review scores correlated well with the scores given by professional critics. This 
is an encouraging sign in the context of our broader goal of generalizing the analysis we've present to arbitrary to video game related text found on the 
internet. We also found that, even without recourse to sophisticated machine learning techniques, the data gathered were sufficient to identify
"low-hanging fruit" for future game developers -- that is, we found game genres which are highly sought after by users, yet have an underabundance of titles
and poor user reviews. This is an example of a way in which data science can provide actionable insights to the gaming industry.

In our final analysis, we applied a variety of machine learning techniques to the problem of analyzing review sentiment. We used Naive Bayes classification
to predict whether a review conveyed positive or negative sentiment, and found $\sim$76% accuracy after optimizing our classifier. We further 
applied random forest classification to the same problem, finding $\sim$73% accuracy. These findings are very comparable to those of the Bayesian Tomatoes
movie review analysis, which found Naive Bayes classification accuracies of $\sim$74-77%. We further extended beyond the Bayesian Tomatoes study by building 
regression models to predict the numerical score corresponding to a review based on its text. We achieve an RMSE of $<1.3$, with reviews rated on a 1-10 scale.

Lastly, we applied our optimized Naive Bayes classifier, trained on GameSpot reviews, to predict the sentiment of reviews from the Giant Bomb data set. Remarkably,
the classifier still achieved $>$75% accuracy when applied to a completely different movie review data set than that used to train the model. Thus, this study
represents a good first step in the goal of applying machine learning and data science techniques to arbitrary video game related online content in order
to gain insights about consumer opinion and future development endeavors.