Recent technological advances enable to collect a variety of knowledge and heterogeneous data from multiple domain. As various types of data including prior knowledge and multimodality are generated, numerous methods to integrate such dataset have been developed to extract complementary knowledge from multiple domain. However, integrating prior knowledge and multimodal data is challenging in four aspects: small sample size problem (P1), sequential data processing (P2), irregularity of heterogenous data (P3), and model interpretability (P4). In this thesis, we suggest two sample exploitation methods for incorporating multimodal data resolving four aspects of knowledge and data integration issue. In the first study, we especially focus on small sample size problem (P1) for multimodal data integration in the field of bioinformatics where available sample size is extremely small. The suggested model is intrinsically able to integrate irregular multimodal data (P3) while recognizing subtype-sensitive genes (P4). Subsequently, we expand our study to time series data with multimodality (P2, P3) using sample exploitation approach (P1) while model interpretability (P4) is kept. Across two studies sample exploitations are performed via kernel-reweighting and separate learning phase, respectively. The suggested methods are validated using 4 experiments. For the first study, L1-regularized kernel-reweighting regression model is used for inferring subtype-specific patterns between gene expression and DNA methylation. Subsequent experiments include simulation study, predicting Alzheimer’s disease progression of patients in mild cognitive impairment, and analyzing genomic variation affecting AD progression.
