Fig. 4.

Architectures for the two-stream active query suggestion model. (a) For model initialization, we train the supervised (E_s) and unsupervised (E_u) feature extractors using VAEs. (b) For two-stream clustering, we compare two design choices to combine E_u and E_s features in an either parallel (late-fusion) or hierarchical manner. The block C_i denotes the clustering algorithm. (c) For active clustering, we fine-tune E_u with triplet loss to encourage the learning of discriminative features.