Questions and Background Information

Question: Can we predict which video games will be highly rated?

Our target variable is critic score. The data was collected from VGChartz. The VGChartz Network is a collection of five video game websites: VGChartz, gamrFeed, gamrReview, gamrTV, and gamrConnect. VGChartz sits at the center of the network and is a video game sales tracking website, providing weekly sales figures of console software and hardware by region. Related projects include: Video Game Analysis and Analyzing gaming platform popularity using video game sales

We are asking this question in order to determine if there are specific variables that can be used to predict how well a video game will be received by critics. If we find certain predictor variables, we can pass this information to video game creators/publishers so they can develop games that will receive good ratings. Ratings are an important factor in video game scales. Consumers want to buy good quality games, and one of the easiest ways to determine which games are worth buying is by checking reviews.

Exploratory Data Analysis

The variables we believe to be important are platform, genre, and rating because certain platforms, genres, or ratings are likely to have higher ratings. This is important to video game developers so that they know which platform, genre, and rating to prioritize so that their video games will be rated higher. We think sales will be our key predictor variable.

Data Distribution

Bar Charts

Scatterplots

Prevalence

## [1] 0.3122211

Methods

In order to create a model to answer our ML question, we are employing a kNN model We explored using clustering and/or cross-validation, but the kNN model, alone, was strong enough so that we feel we don’t need anything beyond it.

KNN Model: k=9

## Confusion Matrix and Statistics
## 
##           Actual
## Prediction   0   1
##          0 448  34
##          1   7 200
##                                          
##                Accuracy : 0.9405         
##                  95% CI : (0.9201, 0.957)
##     No Information Rate : 0.6604         
##     P-Value [Acc > NIR] : < 2.2e-16      
##                                          
##                   Kappa : 0.8635         
##                                          
##  Mcnemar's Test P-Value : 4.896e-05      
##                                          
##             Sensitivity : 0.8547         
##             Specificity : 0.9846         
##          Pos Pred Value : 0.9662         
##          Neg Pred Value : 0.9295         
##              Prevalence : 0.3396         
##          Detection Rate : 0.2903         
##    Detection Prevalence : 0.3004         
##       Balanced Accuracy : 0.9197         
##                                          
##        'Positive' Class : 1              
## 

Selecting the Correct “k”

KNN Model: k=3

Evaluating the Final Model

## Confusion Matrix and Statistics
## 
##           Actual
## Prediction   0   1
##          0 446  22
##          1   9 212
##                                           
##                Accuracy : 0.955           
##                  95% CI : (0.9367, 0.9692)
##     No Information Rate : 0.6604          
##     P-Value [Acc > NIR] : < 2e-16         
##                                           
##                   Kappa : 0.8983          
##                                           
##  Mcnemar's Test P-Value : 0.03114         
##                                           
##             Sensitivity : 0.9060          
##             Specificity : 0.9802          
##          Pos Pred Value : 0.9593          
##          Neg Pred Value : 0.9530          
##              Prevalence : 0.3396          
##          Detection Rate : 0.3077          
##    Detection Prevalence : 0.3208          
##       Balanced Accuracy : 0.9431          
##                                           
##        'Positive' Class : 1               
## 

ROC Curve

AUC performance

## [[1]]
## [1] 0.9831831

We have 95% accuracy and our sensitivity is around 0.90 which means about 90% of high critic score games are correctly labeled and specificity around 0.98 which means about 2% of games labeled as low-critic score are false positives. We also have a really good AUC at 0.98 which shows that our model is good at correctly predicting high critic score games and low critic score games. Overall, it seems that the model is performing very well.

Fairness Assessment

We do not need to conduct an FA because we have no protected classes in our data.

Conclusions

Our results show that video games rated T and E, PlayStation games, and action games tend to have higher ratings. Despite limitations of not having enough data or numeric variables, our model works well with 95% accuracy and 90% of high critic score games correctly predicted. Our AUC is also high at .98, meaning our model is correctly predicting the classifications of video game critic scores.

Future Work

One of the limitations we faced during this project is just not having enough data. After we went through, cleaned our data, and eliminated the rows in the Critic_Score column with N/A values, we were left with half of the initial dataset, which poses obvious issues surrounding any conclusions we can make with our model. If we were able to run this model with complete data, the analysis would be more complete and we could more accurately answer our question. Nonetheless, we still had close to 7000 rows of data at our disposal, so our conclusions most likely remain robust. Adding other numeric variables to our dataset would also allow us to better understand and predict using clustering and thus we could have more visuals that depict our decisions when creating a video game We are using a predictor variable (sales) that is very good at predicting, but if we wanted to understand how to achieve a high critic rating before the game is released, this could pose a limitation to our model.