Patent Search Using Triplet Networks Based Fine-Tuned SciBERT

In this paper, we propose a novel method for the prior-art search task. We fine-tune SciBERT transformer model using Triplet Network approach, allowing us to represent each patent with a fixed-size vector. This also enables us to conduct efficient vector similarity computations to rank patents in query time. In our experiments, we show that our proposed method outperforms baseline methods.

Multilevel Memetic Hypergraph Partitioning with Greedy Recombination

The Hypergraph Partitioning (HGP) problem is a well-studied problem that finds applications in a variety of domains. The literature on the HGP problem has heavily focused on developing fast heuristic approaches. In several application domains, such as the VLSI design and database migration planning, the quality of the solution is more of a concern than the running time of the algorithm. KaHyPar-E is the first multilevel memetic algorithm designed for the HGP problem and it returns better quality solutions, compared to the heuristic algorithms, if sufficient computation time is given. In this work, we introduce novel problem-specific recombination and mutation operators, and develop a new multilevel memetic algorithm by combining KaHyPar-E with these operators. The performance of our algorithm is compared with the state-of-the-art HGP algorithms on $150$ real-life instances taken from the benchmark datasets used in the literature. In the experiments, which would take $39,000$ hours in a single-core computer, each algorithm is given $2, 4$, and $8$ hours to compute a solution for each instance. Our algorithm outperforms all others and finds the best solutions in $112$, $115$, and $125$ instances in $2, 4$, and $8$ hours, respectively.