NLP FOR MOBILE APP ANALYSIS
WHY
-
Zalora has many Apple & Android app store reviews, both positive and negative across the 6 app stores in different markets. Currently, there is no process to react to these customer feedback
-
This Python tool enables automation & quick identification of key issues for technical & business teams to take action – this allows for fast reaction to customers & effective long-term app performance monitoring
HOW
-
Used latent dirichlet allocation, which is a three-level hierarchical Bayesian model
-
Performed hyperparametric tuning across 2 parameters: the number of topics and the learning decay
-
Gridsearched across 6 number of topics and 3 learning decays, and select the best model based on the highest log likelihood score, which is a statistical measure of goodness of fit
-
Note the model’s perplexity, which a lower perplexity is better
-
To check for goodness of fit, ensure that most reviews are a good match for ideally only 1 topic
