DUAL DENOISING LOGICAL REASONING FOR INDUCTIVE KNOWLEDGE GRAPH COMPLETION

This path sampling section of DDLR is a blatant plagiarism of the work “A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs” [1] available on arxiv since June 2022, I’ll therefore refer to the original 2022 work as “the A*Net paper” or “A*Net”, and to this ICLR submission as “DDLR”.

This submission (DDLR) copies almost verbatim several sections of the A*Net paper including images and equations without any reference or citation of the original work:

• Section 1: Figure 1 in DDLR is a re-colored version of Figure 2(a) in A*Net.
• Section 3 (Preliminary): The Knowledge Graph Completion paragraph is a copy of the same paragraph in the A*Net with some words replaced by synonyms, eg,
  • [A*Net] “Each fact is a triplet $(x,r,y) \in \mathcal{V} \times \mathcal{R} \times \mathcal{V}$, which indicates a relation $r$ from entity $x$ to entity $y$.”
  • [DDLR] “Each fact is a triplet $(x,r,y) \in \mathcal{V} \times \mathcal{R} \times \mathcal{V}$, signifying the presence of relation $r$ from entity $x$ to entity $y$.”
• Section 3 (Preliminary): The Knowledge Graph Completion paragraph
  • [A*Net] The task of knowledge graph reasoning aims at answering queries like $(u,q,?)$ or $(?,q,u)$. … we assume the query is $(u,q,?)$, since $(?,q,u)$ equals to $(u,q^{-1},?)$ with $q^{-1}$ being the inverse of $q$.
  • [DDLR] Knowledge graph completion is a task focused on predicting answers for queries such as $(u,q,?)$ or equivalently $(u,q^{-1},?)$, where $q^{-1}$ represents the inverse relation of $q$. Here the authors make a mistake because $(u,q^{-1},?)$ is not equivalent to $(u,q,?)$ but instead is equivalent to the head prediction task to $(?,q,u)$.
• Section 3 (Preliminary) -> Path-based Approaches: almost verbatim copy of the text, equation 1, and the same notation from the A*Net paragraph Path-based Methods.
• Section 4.1 (Path Sampling) -> Important Paths for Reasoning: the whole paragraph is a verbatim copy of Section 3.1 (Important Paths for Reasoning) from the A*Net paper with non-sensical word replacements, eg
  • [A*Net] Given a query relation and a pair of entities, only some of the paths between the entities are important for answering the query.
  • [DDLR] Given a query and a lot of entities, only some of the paths between the entities are important for answering the query.
• Section 4.1 (Path Sampling) -> Path Scoring includes copy-paste excerpts from A*Net’s paragraphs Iterative Computation of Important Paths, Reasoning with A* Algorithm, and Neural Priority function. Particularly,
  • Equations 3,4,5,7 (DDLR) are the same as Equations 6,8,9,10 from A*Net with the same notation and descriptions
  • The priority function $s_{uq}(x)$ is named as “learnable path scoring function” in DDLR instead of “neural priority function” in A*Net

Even if we omit the borrowed A*Net content from the list of contributions, the paper boils down to pre-computing co-occurences of all pairs of relations and tuning the threshold manually for each dataset - this is not a substantial algorithmic contribution per se.

Overall, I believe the paper should be rejected as it clearly violates the ICLR Code of Ethics (that authors are supposed to accept upon the submission) and escalated to the AC / SAC / PC level for further investigation.

[1] https://arxiv.org/abs/2206.04798

1. The paper is not organized clearly, which is not friendly for understanding. Specifically, there are lack the detailed discussion about related work [1].
2. This paper should be sent to the ICLR ethical board review committee on the subject of scientific misconduct and there are some ethics problems in this paper according to the ANet paper [1]. [1] ANet: A Scalable Path-based Reasoning Approach for Knowledge Graphs

The only part of the model that is a significant difference between this work and A*Net is Section 4.2.

I don't see any excuses for not comparing this work with A*Net, especially since this paper includes a more recent arxived baseline GraPE.

A few sections of the paper are nearly identical or highly similar to the A*Net paper [1]. However, the A*Net paper is not cited anywhere. This raises ethical concerns, and puts the novel contributions of this work into question. The authors need to clearly distinguish their contributions compared to A*Net and other prior works through proper citation and explanation. As presented, the incremental novelty is unclear since the core ideas appear derived from A*Net without attribution. This is a major weakness.

[1] A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs