Multiple Treatment Effect Estimation using Deep Generative Model with Task Embedding

GraphITE: Estimating Individual Effects of Graph-structured Treatments

“…(v) Counterfactual regression (CFR) (Shalit, Johansson, and Sontag, 2017), which is a state-of-the-art deep neural network model based on balanced representations between treatment and control instances; we used the MMD as its IPM. Following previous studies (Yoon, Jordon, and van der Schaar, 2018;Saini et al, 2019), we extend the CFR (Shalit, Johansson, and Sontag, 2017) to the multiple-treatment setting; we regard the most frequent treatment as the control treatment. We also tested several variants of GraphITE: (vi) a variant with no bias mitigation that does not have the HSIC regularization term and Fig.…”

“…There have been several approaches designed for multiple treatments (Schwab, Linhardt, and Karlen, 2018;Wager and Athey, 2018;Chipman et al, 2010); however, most of them are limited to a relatively small number of treatments, making it difficult to consider more than a few dozen treatments. Saini et al (2019) whose motivation was somewhat similar to ours, considered combinatorial treatments; however, their focus was on a large number of combinations made from a small number of treatments, whereas we focus on many single treatments with the help of information on the treatments.…”

“…However, in some scenarios, the number of treatments can be considerably larger. For example, when modeling the effects of drugs on target cells, the number of candidate drugs (i.e., treatments) can be large, and the number of observations per drug can be small due to the high cost of clinical trials (Ganter et al, 2005;Hoelder, Clarke, and Workman, 2012); a similar situation can also occur in advertisement (Saini et al, 2019). Such data scarcity exacerbates the above problems.…”

“…It is built upon the recent significant advances in learning representations from graphs using graph neural networks (GNNs) (Kipf and Welling, 2016;Gilmer et al, 2017). Bias mitigation with a large (possibly infinite) number of treatments is another issue because most existing frameworks (Shalit, Johansson, and Sontag, 2017;Schwab, Linhardt, and Karlen, 2018;Saini et al, 2019) are not designed for such cases. To reduce the treatment selection bias depending on the individual target, GraphITE finds representations of the target and treatment that are as independent of each other as possible.…”

Deep causal learning for robotic intelligence

“…Task embedding-based causal effect variational autoencoder (TECE-VAE) scales CEVAE with task embedding to estimate the individual treatment effect using observational data for the applications that have multiple treatments (Saini et al, 2019). Additionally, TECE-VAE adopts the encoder-decoder architecture.…”