Experimental Results

EQ2: Effectiveness of Jointly Capturing Temporal and Atemporal Dependencies between Entities

This figure shows the results where the x-axis represents the value of w, and the y-axis does the accuracy. The horizontal line indicates the accuracy of THOR(aT).

The results show that the additional usage of factual information at adjacent past and future timestamps is significantly effective in addressing the lack of information in a TKG.

EQ3: Effectiveness of Capturing Structural Dependency between Relations

We analyze the effectiveness of RelGCN in detail. We claimed that the relations included in the facts expressed by the same entity do not always be relevant each other. Based on this claim, RelGCN uses the value of w as a threshold that determines the neighborhood of each relation.

To validate this design choice, we made a variant, denoted as THOR(Tw=8&aT&R ̃)_SSL, which naively constructs directed edges between relations if two relations are included in incoming and outgoing facts expressed by the same entity (i.e., without considering a threshold w).

The below table shows THOR(Tw=8&aT&Rw=8)_SSL dramatically outperforms THOR(Tw=8&aT&R ̃)_SSL. Specifically, THOR(Tw=8&aT&Rw=8)_SSL achieves up to 17.70%, 25.84%, 13.95%, and 6.56% higher MRR, Hits@1, Hits@3, and Hits@10, respectively, than THOR(Tw=8&aT&R ̃)_SSL. The results indicate that we need to carefully examine and pick the pairs of co-occurring relations in a TKG.