The Importance of Suppressing Complete Reconstruction in Autoencoders for Unsupervised Outlier Detection

Yafei Shen, Ling Yang

Abstract: Autoencoders are widely used in outlier detection due to their superiority in handling high-dimensional and nonlinear datasets. The reconstruction of any dataset by the autoencoder can be considered as a complex regression process. In regression analysis, outliers can usually be divided into high leverage points and influential points. Although the autoencoder has shown good results for the identification of influential points, there are still some problems when detect high leverage points. Through theoretical derivation, we found that most outliers are detected in the direction corresponding to the worst-recovered principal component, but in the direction of the well-recovered principal components, the anomalies are often ignored. We propose a new loss function which solve the above deficiencies in outlier detection. The core idea of our scheme is that in order to better detect high leverage points, we should suppress the complete reconstruction of the dataset to convert high leverage points into influential points, and it is also necessary to ensure that the differences between the eigenvalues of the covariance matrix of the original dataset and their corresponding reconstructed results in the direction of each principal component are equal. Besides, we explain the rationality of our scheme through rigorous theoretical derivation. Finally, our experiments on multiple datasets confirm that our scheme significantly improves the accuracy of outlier detection.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)