Theses and Dissertations Collection

Many datasets have missing entries. Since downstream tasks often require full datasets with little

noise, accurately imputing the missing data is quite valuable. Autoencoders have proven themselves as

effective data imputers. However, while they exploit high order dependencies between the columns of

a dataset, autoencoders typically treat each row independently. This produces two problems. First,

imputation accuracy is suboptimal because not all of the data is used effectively. Second, downstream

classification tasks suffer since rows belonging to different classes get treated the same. Presented in this

thesis is CLAIM (CLustered Autoencoder IMputation), an algorithm that adapts existing autoencoder

networks in a way that directly addresses these issues. CLAIM first separates rows into clusters based

on similarity. Then, in the encoder, it applies different, loosely connected, learned linear transformations

to each cluster. Results show that this method improves accuracy with typical autoencoder imputation

strategies on large enough datasets. Also presented is a CLAIM-specific iterative clustering algorithm,

which allows CLAIM to improve initial cluster assignments as needed.