Page 110 - AIH-1-3

Artificial Intelligence in Health ISM: A new multi-view space-learning model

Figure 7. Signature 915 data: treemap of integrated sources model loadings of the view-mapping matrix

error associated with a rank is not as critical if it exceeds chosen rank, in line with its parent methods, NMF and
the number of known classes. Compared to a 10-rank ISM NTF.
model, a 12-rank model also finds 10 classes and gives a
slightly higher purity index (6.24 vs. 5.81), despite a larger 3.3.2. About changing the sparsity coefficient
relative error (0.60 vs. 0.52) (Table 2, bottom part). The final We have already mentioned that the initial degree of
part of this section discusses this point further. sparsity of H returned by NMF is a critical part of ISM,
For the Signature 915 dataset, where the chosen ISM as zero-loading attributes are anchors that maintain
rank is 16, the relative error does not change significantly consistency between view components during the
for neighboring embedding dimensions: 0.33 for a embedding process. However, it is extremely difficult to
15-embedding and 0.34 for a 17-embedding (Table 3, predict how sparse an NMF representation will be, as this
upper part). Choosing an embedding dimension equal to depends on the dataset under analysis. To ensure that a
33
the rank is more consistent with the ISM workflow, where sufficient number of anchors will guide the embedding,
the embedding and latent spaces are united during the only significant loadings are retained, while other loadings
straightening process. Therefore, we chose an embedding are set to 0. ISM uses the inverse of the HHI to identify
dimension of 16. In terms of purity, a 17-rank ISM model significant loadings, but an additional sparsity parameter is
gives results that are slightly superior to the 16-rank ISM provided to allow this index to be relaxed. This parameter
model (Table 3, bottom part). is set to 0.8 by default. In this section, we examine the effect
Overall, these results confirm that ISM provides of changing this parameter in the UCI Digits and Signature
relatively stable estimates in the neighborhood of the 915 experiments (Tables 4 and 5, respectively).

Volume 1 Issue 3 (2024) 104 doi: 10.36922/aih.3427