1301962d8b7bd03fffaa27119aa7fc2b-Supplemental.pdf

For a totalN nodes, we have an N-dimensional Hawkes process. The edges between nodes are inferred through the dimensionwiseendogenous matrix. Firstly, we can see that there isn't much variations for the latent features' proportions, which is calculated asP Figure 4: From top to bottom: NP-PP, NP-HP, NP-HP-C. The values do not always have a sharp decrease in thebeginningperiod. In Eq. (1), the sigmoid functionσ()and the Gaussian process function{fk()}k make it difficult toform convenient format toenable mean-field variationalinference. Thus, we use the technique of [2] to overcome these issues.