Discussion and Conclusion - GetRecced/IR670_Spring2018 GitHub Wiki

Discussion

• Baseline approach is a good approach to start with, as it gives good recommendations in short time.
• In case of HFT, beyond a certain value, increasing the number of topics gives negligible improvements in RMSE.

Takeaways

• Using review topics in conjunction with ratings gives out better recommendations than traditional methods.
• It is difficult to make accurate predictions by relying solely on review text and biases (LDA).
• HFT is dataset dependent - some categories of items have reviews that better express the subjective tastes of the user/properties of the item
• App reviews are short and to-the-point, and mostly positive.

Future Work

• Word2Vec or GloVe can be used as part of hidden topic extraction from reviews
• The HFT model may be adapted to work with implicit data instead of ratings, along with just reviews.