My latest Machine Learning blog post Confidence Splitting Criterions Can Improve Precision And Recall in Random Forest Classifiers is out on the Airbnb Data blog:
The Trust and Safety Team maintains a number of models for predicting and detecting fraudulent online and offline behaviour. A common challenge we face is attaining high confidence in the identification of fraudulent actions. Both in terms of classifying a fraudulent action as a fraudulent action (recall) and not classifying a good action as a fraudulent action (precision).
A classification model we often use is a Random Forest Classifier (RFC). However, by adjusting the logic of this algorithm slightly, so that we look for high confidence regions of classification, we can significantly improve the recall and precision of the classifier’s predictions. To do this we introduce a new splitting criterion (explained below) and show experimentally that it can enable more accurate fraud detection.
Have a read and let me know what you think!
While almost all members of the Airbnb community interact in good faith, there is an ever shrinking group of bad actors that seek to take advantage of the platform for profit. This problem is not unique to Airbnb: social networks battle with attempts to spam or phish users for their details; ecommerce sites try to prevent the use of stolen credit cards. The Trust and Safety team at Airbnb works tirelessly to remove bad actors from the Airbnb community and to help make the platform a safer and trustworthy place to experience belonging.
Missing Values In A Random Forest
We can train machine learning models to identify new bad actors (for more details see the previous blog post Architecting a Machine Learning System for Risk). One particular family of models we use is Random Forest Classifiers (RFCs). A RFC is a collection of trees, each independently grown using labeled and complete input training data. By complete we explicitly mean that there are no missing values i.e.
NaN values. But in practice the data often can have (many) missing values. In particular, very predictive features do not always have values available so they must be imputed before a random forest can be trained.