Probabilistic Case-based Reasoning for Open-World Knowledge Graph Completion

Abstract

A case-based reasoning (CBR) system solves a new problem by retrieving 'cases' that are similar to the given problem. If such a system can achieve high accuracy, it is appealing owing to its simplicity, interpretability, and scalability. In this paper, we demonstrate that such a system is achievable for reasoning in knowledgebases (KBs). Our approach predicts attributes for an entity by gathering reasoning paths from similar entities in the KB. Our probabilistic model estimates the likelihood that a path is effective at answering a query about the given entity. The parameters of our model can be efficiently computed using simple path statistics and require no iterative optimization. Our model is non-parametric, growing dynamically as new entities and relations are added to the KB. On several benchmark datasets our approach significantly outperforms other rule learning approaches and performs comparably to state-of-the-art embedding-based approaches. Furthermore, we demonstrate the effectiveness of our model in an "open-world" setting where new entities arrive in an online fashion, significantly outperforming state-ofthe-art approaches and nearly matching the best offline method. 1

Optimizations for Computing Relatedness in Biomedical Heterogeneous Information Networks: SemNet 2.0

“…For example, TransE treats each entity as a point in Euclidean space and assumes that relations can be effectively modeled as translations between entity embeddings, i.e., s + r ≈ t for source node s, target node t, and relation r. A wide variety of other models operate on some variant of this assumption, substituting translation by element-wise scaling [55], rotation in complex space [56], rotation in Quaternary space [57], or rotation and reflection in hyperbolic space [58]. An smaller, alternative family of knowledge base completion literature focuses instead on inferring missing relations by aggregating information either explicitly [59] or implicitly [60,61] encoded in the (meta)paths between them. This approach is more desirable for biomedical KBs due to the fact that relevant nodes and paths can be extracted from the graph to provide an understandable explanation of the predictions.…”

Low Resource Quadratic Forms for Knowledge Graph Embeddings

“…Case-based reasoning for knowledge graph completion was introduced in (Das et al, 2020). The authors show that a k-nearest neighbor (KNN) based approach for can be both efficient and scalable for this application.…”

Multi-hop knowledge base question answering by case retrieval

“…CBR is a paradigm that reusing solutions of other similar problems to derive a solution for the given problem. Existing CBR-based KBQA methods compute dense representation of given question and use it to retrieve similar k-nearest neighbor (KNN) questions from a training set, leveraging their solutions to help answer given question (Das et al, 2021 [13] , 2022 [14] ). A case can be defined as a problem along with its solution in a CBR system.…”

“…A disadvantage of their method is that it relies on the availability of logical form annotations, which is very labor-intensive and expensive. Das et al [14] define a case as a question with its specific KB subgraph. They use a depth-first search (DFS) to collect KB paths that connect the topic entity to its answer entities as the specific subgraph for each training question, and then traverse KB followed the paths in each subgraph of KNN questions to form the subgraph of given question.…”