NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:303
Title:CPM-Nets: Cross Partial Multi-View Networks

Reviewer 1

The authors picked a relevant problem and the approach they propose looks reasonable. I have two major concerns:(i) the methodology assumes that the latent representation encodes similarities between networks. This can behave very different from what the authors intended (see Donnat and Holmes in AOAS) since different measures of similarity on graph space may lead to radically different results when clustering; (ii) There is literature for modelling multiple networks via latent variables. It is not clear how the proposed method relates to established approaches (Gollini and Murphy, Nielsen and Witten). The results from the real data sets look promising.

Reviewer 2

This paper proposes a novel algorithm to utilize the data with missing views, and the algorithm shows the superior performance. Strength: 1) The proposed strategy of handling view-missing is novel and elegant compared with existing strategies. I agree that most existing methods [5][8][16] are not flexible especially for the data with large number of views and complex missing patterns. 2) The theoretical analysis of completeness and versatility is a good support for the success of the learned multi-view representation, which is inspiring for the field of multi-view learning. 3) The designed clustering-like loss is interesting and a good choice for small-sample-size data. Moreover, the conducted experiments are extensive, and the proposed model achieves better results consistently. Minor comments: (1) It will be better if the authors could provide analysis about the computational cost, and release their code, since there may be potential applications for the method. (2) In Figure 4, different clusters correspond to different digits, so can you label the clusters with corresponding digits?

Reviewer 3

The paper proposed a multi-view learning framework that is able to handle missing views. The improvement compared to state-of-the-art seem to be the preservation of versatility with the new formulation, and better numerical results shown in section 3. I increased my rating after reading the authors' feedback. The paper is clear and related algorithmic steps/analysis are well presented. I like the discussion on the consistence and complementarity while establishing the framework. The technical part (1)-(5) is intuitive and easy to follow, although the existence prove in prop 2.1 does not bear much guarentee on the versatility of (5), which requires a more in-depth study of its solution property. Together with the results in supplement, the evaluation part is comprehensive with alternatives compared and discussed. I would encourage the authors to share their implementation for reproducibility and future works on partial multi-view learning.