Detecting Manipulated Wine Ratings with Autoencoders and Supervised Machine Learning Techniques

Titelangaben

Baumann, Michaela ; Baumann, Michael Heinrich:
Detecting Manipulated Wine Ratings with Autoencoders and Supervised Machine Learning Techniques.
In: International Journal on Advances in Internet Technology. Bd. 15 (2022) Heft 3&4 . - S. 64-77.
ISSN 1942-2652

Volltext

Link zum Volltext (externe URL):

Weitere URLs

Abstract

In this study, we analyze the ability of different machine learning methods to detect manipulated wine ratings. We consider autoencoders, regression models (neural networks, support vector machines, random forests) and classification models (support vector machines, random forests) and two different kinds of manipulation strategies. We find that autoencoders perform best on unmanipulated test data, i.e., their reconstruction error is smaller than the supervised models’ prediction error. However, on the manipulated test data, the supervised models outperform autoencoders. This is interesting since autoencoders are generally used for outlier detection. When comparing only the supervised methods, we find that, basically, both support vector machines and random forests perform and detect better than regression neural networks. Additionally, the optimization and training times for these two model types are smaller. In order to consider a relatively large grid of hyperparameters especially for the neural networks, we introduce a hyperparameter tuning method called sequential accumulative selection. To sum up, when trying to detect manipulations, different methods have usually both advantages and disadvantages.