https://docs.google.com/presentation/d/1s0V5Rxpj1zC6UNL7ffNd-ScdfUDiF7aXAnSptVCEI8A/edit#slide=id.g1307426a92_0_36 Machine learning model to predict the "helpfulness" rating of Amazon product reviews.
For the first iteration additional features seem to add only slight incrememntal inreases in predictive power if any at all
The UniqueIDs feature did not seem to produce any value for the model and may even have decreased the models overall accuracy;
Struggeling to find an effective method to include counts of multiple reviews by individual users as a feature I ended up stripping out users who did not write more than one review, effectively cutting the amount of data I had to work with in half.
Adding the additional features which I included in the original data set resulted in a similar bump in accuracy in the smaller data set, but there is not enough evidence to suggest the decrease in accuracy seen in the repeat review writers trial would be remedied by including the rest of the data. We'll see.
##Update: Running a count vectorizer on the Summary text of each review increased the models accuracy significantly.
Based on this discovery further methods were run on the Summary and other text based elements of the data; Namely an NLTK Stemmer which was rolled into the CountVectorizer. After this expansion the highest accuracy thus far was attained with nearly a 50% increase in true positive rate.
The Multinomial Naive Bayes classifier was chosen as the final model because while the overall accuracy was not the highest of the models run, it had a significantly higher true positive rate.(turns out if you guess that virtually none of the reviews are helpful you will have a relatively high true negative rate. 😆)