FAQ for TransNets

1. Training TransNets: why is the loss computed between a dropped out vector and a non-dropped out vector

In Algorithm 1 for training TransNets, in Step 2 (line 15) the loss is computed between a dropped out version of the approximate latent representation ($\bar{z}_L$) and a non-dropped out version of the latent representation of the original review ($x_T$). This is because, $x_T$ is the ground truth (or final objective) for this particular section of training. As a sanity check, training TransNets with minimizing the loss with a dropped out version of $x_T$ gave a lower performance.

2. Test users/items that don't appear in train

In TransNets, these users/items will be represented using an empty string. In TransNets-Ext, the user/item embedding matrix maps them to a random vector.

3. user-item pair with rating but no review

Since the Target network requires a review, presently, TransNets is modeled such that only those user-item pairs that have a review are taken into account. Entries that do not have a review are filtered out. However, in a number of real-world scenarios, users rate an item but do not always enter a review. It would be interesting to extend the TransNets model to handle this.