Short-Term-to-Long-Term Memory Transfer for Knowledge Graphs under Partial Observability

Michael Cochez; Taewoon Kim; Vincent Fran\c{c}ois-Lavet

arxiv: 2605.22142 · v1 · pith:LBDCWGQCnew · submitted 2026-05-21 · 💻 cs.LG · cs.AI

Short-Term-to-Long-Term Memory Transfer for Knowledge Graphs under Partial Observability

Taewoon Kim , Vincent Fran\c{c}ois-Lavet , Michael Cochez This is my paper

Pith reviewed 2026-05-22 08:06 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords reinforcement learningknowledge graphspartial observabilitymemory transferneuro-symbolicshort-term memorylong-term memoryQ-learning

0 comments

The pith

Learned keep-or-drop decisions for each observed fact improve long-term knowledge graph memory use under partial observability.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how an agent decides which symbolic triples to move from a temporary short-term buffer into a persistent long-term knowledge graph memory when the environment is only partially observable. It treats the choice for every triple as a keep-or-drop action inside a value-based reinforcement learning setup. This matters because limited long-term capacity forces explicit decisions about what information to retain for future navigation and queries. The method employs per-item Q-learning with shared parameters and temporal-difference updates across matched items to cope with changing buffer sizes. Experiments on the RoomKG benchmark at capacity 128 show the learned policy beats both symbolic heuristics and neural history models, with a simple local variant performing best and decisions remaining interpretable.

Core claim

In a temporal knowledge-graph memory setting under partial observability, the agent learns a transfer policy that chooses for each observed triple whether to keep or drop it before long-term insertion. Using a per-item Q-learning design with shared parameters and temporal-difference updates over matched items, this learned policy outperforms symbolic and neural baselines on the RoomKG benchmark when long-term memory capacity is set to 128.

What carries the argument

Per-item Q-learning with shared parameters that assigns keep-or-drop values to individual triples and applies practical temporal-difference updates across consecutive steps to manage variable-sized short-term buffers.

If this is right

Learned transfer decisions keep navigation- and query-relevant facts while discarding lower-value candidate facts.
A lightweight local short-term-only variant of the policy performs best among the tested transfer-policy ablations.
Explicit memory decisions remain interpretable and support performance under strict memory constraints.
The approach beats both symbolic baselines that include temporal annotations and neural baselines that rely on LSTM or Transformer history.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The per-item design could be tested on other partially observable tasks that use symbolic structures to see whether variable-buffer handling remains effective.
Varying the long-term capacity beyond 128 might show at which point explicit transfer decisions become more or less valuable than fixed rules.
Integrating the transfer policy with different query types could reveal whether relevance signals generalize beyond navigation.

Load-bearing premise

That a per-item Q-learning design with shared parameters and a practical temporal-difference update over matched items across consecutive steps can effectively handle variable-sized short-term buffers in this setting.

What would settle it

If the learned transfer policy on the RoomKG benchmark at long-term memory capacity 128 fails to outperform the strongest symbolic baseline with temporal annotations or the best history-based LSTM or Transformer baseline, the claimed performance advantage would not hold.

Figures

Figures reproduced from arXiv: 2605.22142 by Michael Cochez, Taewoon Kim, Vincent Fran\c{c}ois-Lavet.

**Figure 2.** Figure 2: Keep-rate trajectory over held-out test steps (higher means more short-term items are transferred to long-term memory). The moving average highlights that transfer is adaptive rather than fixed-rate. alongside the queried object and the agent’s location. At step 26, the policy keeps (agent, at_location, studio) and (table, at_location, studio) while dropping four directional links from studio; this more di… view at source ↗

**Figure 3.** Figure 3: Bird’s-eye schematic of the hidden state at [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

**Figure 4.** Figure 4: Agent internal memory-state snapshots at two time points (steps 0 and 50) in a held-out [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

read the original abstract

Reinforcement learning under partial observability requires deciding what information to retain, yet most memory-based approaches do not explicitly model short-term-to-long-term transfer of symbolic observations. We study this transfer process in a temporal knowledge-graph memory setting and cast it as a neuro-symbolic value-based decision problem: for each observed triple, the agent chooses whether to keep or drop it before long-term insertion. To handle variable-sized short-term buffers, we use a per-item Q-learning design with shared parameters and a practical temporal-difference update over matched items across consecutive steps. On the RoomKG benchmark at long-term memory capacity 128, learned transfer decisions outperform symbolic and neural baselines, including symbolic baselines with temporal annotations and history-based LSTM/Transformer baselines. Across transfer-policy ablations, a lightweight local short-term-only variant performs best, and step-level behavior shows that the policy keeps navigation- and query-relevant facts while discarding lower-value candidate facts, supporting explicit and interpretable memory decisions under memory constraints.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a concrete per-item Q-learning design for deciding which triples to move from short-term buffer to long-term KG memory in partial-observability RL, with reported gains over baselines on RoomKG.

read the letter

The main thing to know is that this work turns the keep-or-drop decision for each observed triple into a value-based learning problem and shows the resulting policy beats both symbolic rules and neural history models on the RoomKG benchmark at long-term capacity 128. The per-item Q-learning with shared parameters and a matched-item TD update is the part that lets them handle variable buffer sizes without forcing a fixed representation for the whole short-term set. That formulation is new relative to the baselines they cite, and the ablation results plus the step-level behavior analysis give some evidence that the policy actually keeps navigation- and query-relevant facts while dropping others. The lightweight local short-term variant coming out on top is also a useful data point. Those pieces are what the paper does cleanly. The matching step in the TD update is the soft spot worth checking. If items are matched only by triple identity, then a triple that stays in the buffer can have a different surrounding context at the next step, which could bias the target and make learning less stable under partial observability. The abstract does not spell out whether they augment the per-item state or just rely on symbolic equality, so that detail matters for whether the reported superiority holds up. Experimental specifics like error bars, exact baseline implementations, and variance across runs are also thin in the summary, though the central comparison appears to rest on external baselines rather than circular definitions. This is for researchers working on memory management in RL agents that use structured symbolic observations, especially knowledge graphs. A reader who needs a practical way to learn transfer decisions under capacity limits will find the design and the RoomKG results directly usable. The work is coherent enough on its own terms and grounded in a clear comparison to warrant peer review, even if it would benefit from more on the matching procedure and tests beyond the single benchmark. I would send it out for refereeing with those requests.

Referee Report

2 major / 2 minor

Summary. The paper proposes a neuro-symbolic value-based RL method for short-term-to-long-term memory transfer of symbolic triples in temporal knowledge graphs under partial observability. It casts retention decisions as per-item Q-learning with shared parameters and a TD update that matches items across consecutive variable-sized short-term buffers. On the RoomKG benchmark with long-term capacity 128, the learned policy outperforms symbolic baselines (including temporally annotated ones) and neural baselines (LSTM/Transformer with history), with ablations favoring a lightweight local short-term-only variant and qualitative evidence that the policy retains navigation- and query-relevant facts.

Significance. If the empirical superiority and the correctness of the per-item TD matching hold, the work offers a concrete, interpretable mechanism for explicit memory transfer in constrained symbolic RL settings, potentially bridging neuro-symbolic methods with standard memory-augmented architectures. The reported outperformance over both temporally annotated symbolic and history-based neural baselines at capacity 128 would be a useful data point for memory-management research under partial observability.

major comments (2)

[Method description (per-item Q-learning and TD update)] The central claim that learned transfer decisions outperform baselines at long-term capacity 128 rests on the per-item Q-learning TD update over matched items. The design implicitly assumes that items can be matched unambiguously (e.g., via triple equality) and that an item's value is sufficiently independent of the current buffer composition. Under partial observability and evolving buffer contents, a purely symbolic match may pair an item with a different surrounding context at t+1, biasing the TD target and potentially destabilizing the learned policy relative to full-state baselines such as LSTM/Transformer.
[Experimental evaluation on RoomKG] Experimental results lack reported error bars, exact baseline implementations, ablation statistics, and full experimental setup details. Without these, it is difficult to assess whether the reported superiority at capacity 128 is robust or sensitive to hyper-parameters and random seeds.

minor comments (2)

[Abstract] The abstract states that 'a lightweight local short-term-only variant performs best' across transfer-policy ablations; a table or figure explicitly comparing all variants with metrics would strengthen this claim.
[Method] Notation for the per-item Q-function and the matching procedure should be formalized with an equation or algorithm box to clarify how shared parameters are updated across variable buffer sizes.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential contribution. We address each major comment below with point-by-point responses and indicate revisions made to the manuscript.

read point-by-point responses

Referee: The central claim that learned transfer decisions outperform baselines at long-term capacity 128 rests on the per-item Q-learning TD update over matched items. The design implicitly assumes that items can be matched unambiguously (e.g., via triple equality) and that an item's value is sufficiently independent of the current buffer composition. Under partial observability and evolving buffer contents, a purely symbolic match may pair an item with a different surrounding context at t+1, biasing the TD target and potentially destabilizing the learned policy relative to full-state baselines such as LSTM/Transformer.

Authors: We appreciate this insightful observation on the per-item TD update. Matching is performed via exact symbolic triple equality, which is unambiguous given the discrete knowledge-graph representation and avoids ambiguity in item identity. The Q-function is intentionally factored per-item to enable scalable decisions; it conditions on local features of the triple and short-term buffer state rather than assuming full independence from global context. While context shifts can introduce approximation bias in the TD target, this is mitigated by the shared-parameter design and the fact that the policy still outperforms full-state neural baselines empirically. In the revision we have added a dedicated paragraph in Section 3.2 clarifying this approximation, its relation to standard Q-learning, and supporting analysis from the RoomKG runs showing that retained items align with query relevance irrespective of exact buffer composition. revision: partial
Referee: Experimental results lack reported error bars, exact baseline implementations, ablation statistics, and full experimental setup details. Without these, it is difficult to assess whether the reported superiority at capacity 128 is robust or sensitive to hyper-parameters and random seeds.

Authors: We agree that these details are essential for assessing robustness. The revised manuscript now reports mean performance with standard-error bars over five independent random seeds for all methods at capacity 128. We have added an appendix with exact baseline implementations (including hyper-parameters, network architectures for LSTM/Transformer, and the temporally annotated symbolic variants), full ablation tables with statistical significance tests (paired t-tests), and an expanded experimental-setup section listing all environment parameters, training schedules, and hardware used. These additions directly address the concerns and improve reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on external benchmark comparisons

full rationale

The paper presents a neuro-symbolic RL method for short-to-long-term memory transfer in knowledge graphs, using a per-item Q-learning design with shared parameters and TD updates over matched items. This is a modeling choice justified by the need to handle variable buffer sizes under partial observability, not a self-referential definition or fitted input renamed as prediction. Results are evaluated on the RoomKG benchmark against independent symbolic, neural, and history-based baselines (LSTM/Transformer), with ablations showing a local short-term variant performing best. No load-bearing self-citations, uniqueness theorems from prior author work, or ansatz smuggling are present in the derivation. The central performance claim at capacity 128 is supported by direct comparisons rather than reducing to the method's own inputs by construction. The design assumptions (e.g., item matching) are explicit and open to the noted skeptic concerns about context, but these are empirical limitations, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, invented entities, or detailed axioms are described. The approach implicitly relies on standard RL assumptions such as the validity of Q-learning for discrete keep/drop actions and the representativeness of the RoomKG benchmark.

axioms (1)

domain assumption Per-item Q-learning with shared parameters and temporal-difference updates over matched items can handle variable-sized short-term buffers.
This modeling choice is presented as enabling the neuro-symbolic transfer decisions.

pith-pipeline@v0.9.0 · 5707 in / 1298 out tokens · 44725 ms · 2026-05-22T08:06:53.004144+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

per-item Q-learning design with shared parameters and a practical temporal-difference update over matched items across consecutive steps
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

learned transfer decisions outperform symbolic and neural baselines... at long-term memory capacity 128

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 1 internal anchor

[1]

Nature , volume=

Human-level Control through Deep Reinforcement Learning , author=. Nature , volume=. 2015 , doi=

work page 2015
[2]

2017 , eprint=

Deep Recurrent Q-Learning for Partially Observable MDPs , author=. 2017 , eprint=

work page 2017
[3]

2022 , eprint=

Deep Transformer Q-Networks for Partially Observable Reinforcement Learning , author=. 2022 , eprint=

work page 2022
[4]

2020 , eprint=

The act of remembering: a study in partially observable reinforcement learning , author=. 2020 , eprint=

work page 2020
[5]

Long Short-Term Memory , journal=

Sepp Hochreiter and J\". Long Short-Term Memory , journal=. 1997 , doi=

work page 1997
[6]

Gomez and Lukasz Kaiser and Illia Polosukhin , title =

Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , year =

work page 2017
[7]

2014 , eprint=

Neural Turing Machines , author=. 2014 , eprint=

work page 2014
[8]

Nature , year=

Hybrid computing using a neural network with dynamic external memory , author=. Nature , year=

work page
[9]

International Conference on Learning Representations , year =

Neural Map: Structured Memory for Deep Reinforcement Learning , author =. International Conference on Learning Representations , year =

work page
[10]

Proceedings of the 34th International Conference on Machine Learning , year =

Neural Episodic Control , author =. Proceedings of the 34th International Conference on Machine Learning , year =

work page
[11]

Journal of Artificial Intelligence Research , volume =

On Overfitting and Asymptotic Bias in Batch Reinforcement Learning with Partial Observability , author =. Journal of Artificial Intelligence Research , volume =. 2019 , doi =

work page 2019
[12]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

A Machine with Short-Term, Episodic, and Semantic Memory Systems , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2023 , doi=

work page 2023
[13]

2025 , howpublished=

Partially Observable Reinforcement Learning with Memory Traces , author=. 2025 , howpublished=

work page 2025
[14]

The Semantic Web:

Modeling Relational Data with Graph Convolutional Networks , author =. The Semantic Web:

work page
[15]

International Conference on Learning Representations , year=

Deep Reinforcement Learning with Relational Inductive Biases , author=. International Conference on Learning Representations , year=

work page
[16]

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

Message Passing for Hyper-Relational Knowledge Graphs , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

work page 2020
[17]

2018 , eprint=

Towards Symbolic Reinforcement Learning with Common Sense , author=. 2018 , eprint=

work page 2018
[18]

Proceedings of the 38th International Conference on Machine Learning , pages =

Discovering symbolic policies with deep reinforcement learning , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

work page 2021
[19]

Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,

Detect, Understand, Act: A Neuro-Symbolic Hierarchical Reinforcement Learning Framework (Extended Abstract) , author =. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,. 2022 , month =

work page 2022
[20]

Sample-Efficient Neurosymbolic Deep Reinforcement Learning

Sample-Efficient Neurosymbolic Deep Reinforcement Learning , author=. 2026 , eprint=. doi:10.48550/arXiv.2601.02850 , url=

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.02850 2026
[21]

2026 , eprint=

Neuro-symbolic Action Masking for Deep Reinforcement Learning , author=. 2026 , eprint=. doi:10.48550/arXiv.2602.10598 , url=

work page doi:10.48550/arxiv.2602.10598 2026
[22]

2018 , eprint=

Action Branching Architectures for Deep Reinforcement Learning , author=. 2018 , eprint=

work page 2018
[23]

2026 , eprint=

Temporal Knowledge-Graph Memory in a Partially Observable Environment , author=. 2026 , eprint=

work page 2026
[24]

2026 , type=

Gregg Kellogg and Olaf Hartig and Pierre-Antoine Champin and Andy Seaborne , title=. 2026 , type=

work page 2026
[25]

2026 , type=

Gregg Kellogg and Andy Seaborne and Dominik Tomaszuk , title=. 2026 , type=

work page 2026
[26]

2026 , type=

Olaf Hartig and Andy Seaborne and Ruben Taelman and Gregory Williams and Thomas Pellissier Tanon , title=. 2026 , type=

work page 2026
[27]

Proceedings of the Tenth International Conference on International Conference on Machine Learning , pages =

Tan, Ming , title =. Proceedings of the Tenth International Conference on International Conference on Machine Learning , pages =. 1993 , isbn =

work page 1993
[28]

and Rula, Anisa and Schmelzeisen, Lukas and Sequeda, Juan and Staab, Steffen and Zimmermann, Antoine , year=

Hogan, Aidan and Blomqvist, Eva and Cochez, Michael and D’amato, Claudia and Melo, Gerard De and Gutierrez, Claudio and Kirrane, Sabrina and Gayo, José Emilio Labra and Navigli, Roberto and Neumaier, Sebastian and Ngomo, Axel-Cyrille Ngonga and Polleres, Axel and Rashid, Sabbir M. and Rula, Anisa and Schmelzeisen, Lukas and Sequeda, Juan and Staab, Steffe...

work page doi:10.1145/3447772
[29]

International Conference on Learning Representations , year =

Semi-Supervised Classification with Graph Convolutional Networks , author =. International Conference on Learning Representations , year =

work page

[1] [1]

Nature , volume=

Human-level Control through Deep Reinforcement Learning , author=. Nature , volume=. 2015 , doi=

work page 2015

[2] [2]

2017 , eprint=

Deep Recurrent Q-Learning for Partially Observable MDPs , author=. 2017 , eprint=

work page 2017

[3] [3]

2022 , eprint=

Deep Transformer Q-Networks for Partially Observable Reinforcement Learning , author=. 2022 , eprint=

work page 2022

[4] [4]

2020 , eprint=

The act of remembering: a study in partially observable reinforcement learning , author=. 2020 , eprint=

work page 2020

[5] [5]

Long Short-Term Memory , journal=

Sepp Hochreiter and J\". Long Short-Term Memory , journal=. 1997 , doi=

work page 1997

[6] [6]

Gomez and Lukasz Kaiser and Illia Polosukhin , title =

Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. Advances in Neural Information Processing Systems 30 (NeurIPS 2017) , year =

work page 2017

[7] [7]

2014 , eprint=

Neural Turing Machines , author=. 2014 , eprint=

work page 2014

[8] [8]

Nature , year=

Hybrid computing using a neural network with dynamic external memory , author=. Nature , year=

work page

[9] [9]

International Conference on Learning Representations , year =

Neural Map: Structured Memory for Deep Reinforcement Learning , author =. International Conference on Learning Representations , year =

work page

[10] [10]

Proceedings of the 34th International Conference on Machine Learning , year =

Neural Episodic Control , author =. Proceedings of the 34th International Conference on Machine Learning , year =

work page

[11] [11]

Journal of Artificial Intelligence Research , volume =

On Overfitting and Asymptotic Bias in Batch Reinforcement Learning with Partial Observability , author =. Journal of Artificial Intelligence Research , volume =. 2019 , doi =

work page 2019

[12] [12]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

A Machine with Short-Term, Episodic, and Semantic Memory Systems , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2023 , doi=

work page 2023

[13] [13]

2025 , howpublished=

Partially Observable Reinforcement Learning with Memory Traces , author=. 2025 , howpublished=

work page 2025

[14] [14]

The Semantic Web:

Modeling Relational Data with Graph Convolutional Networks , author =. The Semantic Web:

work page

[15] [15]

International Conference on Learning Representations , year=

Deep Reinforcement Learning with Relational Inductive Biases , author=. International Conference on Learning Representations , year=

work page

[16] [16]

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

Message Passing for Hyper-Relational Knowledge Graphs , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

work page 2020

[17] [17]

2018 , eprint=

Towards Symbolic Reinforcement Learning with Common Sense , author=. 2018 , eprint=

work page 2018

[18] [18]

Proceedings of the 38th International Conference on Machine Learning , pages =

Discovering symbolic policies with deep reinforcement learning , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

work page 2021

[19] [19]

Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,

Detect, Understand, Act: A Neuro-Symbolic Hierarchical Reinforcement Learning Framework (Extended Abstract) , author =. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,. 2022 , month =

work page 2022

[20] [20]

Sample-Efficient Neurosymbolic Deep Reinforcement Learning

Sample-Efficient Neurosymbolic Deep Reinforcement Learning , author=. 2026 , eprint=. doi:10.48550/arXiv.2601.02850 , url=

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.02850 2026

[21] [21]

2026 , eprint=

Neuro-symbolic Action Masking for Deep Reinforcement Learning , author=. 2026 , eprint=. doi:10.48550/arXiv.2602.10598 , url=

work page doi:10.48550/arxiv.2602.10598 2026

[22] [22]

2018 , eprint=

Action Branching Architectures for Deep Reinforcement Learning , author=. 2018 , eprint=

work page 2018

[23] [23]

2026 , eprint=

Temporal Knowledge-Graph Memory in a Partially Observable Environment , author=. 2026 , eprint=

work page 2026

[24] [24]

2026 , type=

Gregg Kellogg and Olaf Hartig and Pierre-Antoine Champin and Andy Seaborne , title=. 2026 , type=

work page 2026

[25] [25]

2026 , type=

Gregg Kellogg and Andy Seaborne and Dominik Tomaszuk , title=. 2026 , type=

work page 2026

[26] [26]

2026 , type=

Olaf Hartig and Andy Seaborne and Ruben Taelman and Gregory Williams and Thomas Pellissier Tanon , title=. 2026 , type=

work page 2026

[27] [27]

Proceedings of the Tenth International Conference on International Conference on Machine Learning , pages =

Tan, Ming , title =. Proceedings of the Tenth International Conference on International Conference on Machine Learning , pages =. 1993 , isbn =

work page 1993

[28] [28]

and Rula, Anisa and Schmelzeisen, Lukas and Sequeda, Juan and Staab, Steffen and Zimmermann, Antoine , year=

Hogan, Aidan and Blomqvist, Eva and Cochez, Michael and D’amato, Claudia and Melo, Gerard De and Gutierrez, Claudio and Kirrane, Sabrina and Gayo, José Emilio Labra and Navigli, Roberto and Neumaier, Sebastian and Ngomo, Axel-Cyrille Ngonga and Polleres, Axel and Rashid, Sabbir M. and Rula, Anisa and Schmelzeisen, Lukas and Sequeda, Juan and Staab, Steffe...

work page doi:10.1145/3447772

[29] [29]

International Conference on Learning Representations , year =

Semi-Supervised Classification with Graph Convolutional Networks , author =. International Conference on Learning Representations , year =

work page