Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

ManifoldMind: Dynamic Hyperbolic Reasoning for Trustworthy Recommendations

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ManifoldMind claims that scoring multi-hop tag paths in hyperbolic space gives the best ranking, calibration, and diversity across four benchmarks.

desk verdict Innovative architecture, but Section 4.1's math is internally contradictory: the distance formula isn't a Poincaré geodesic, the curvature weighting is inverted, and the kernel's uncertainty sign is backwards—as written the model can't produce the reported results. read the letter →

arxiv 2507.02014 v1 pith:BIRNAZ7A submitted 2025-07-02 cs.IR cs.AIcs.LGstat.ML

Anoushka Harit , Zhongtian Sun , Suncica Hadzidedic This is my paper
classification cs.IRcs.AIcs.LGstat.ML
keywords hyperbolicembeddingsrecommendersystemsuncertaintycalibrationmulti-hopreasoningsemantictagsrecommendationdiversityinterpretabilityprobabilisticgeometry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ManifoldMind is a recommendation model whose entities are probabilistic spheres in hyperbolic space, each with its own curvature and uncertainty radius. The paper's central claim is that scoring multi-hop paths through shared semantic tags with a curvature-aware kernel produces better top-k ranking, calibration, diversity, and interpretability than six baselines on four public datasets. Concretely, it reports NDCG@10 of 0.473, exceeding the strongest baseline FineRec by 4.5%, with the lowest expected calibration error among all compared models. The motivation is trustworthiness: if the claim is right, recommenders can explain each suggestion as a chain of concepts and state how confident they are, which matters in sparse or high-stakes domains. The central mechanism is a maximum-product score over tag paths found by beam search.

What carries the argument

The central object is the probabilistic hyperbolic sphere $(\mu_e, r_e, \kappa_e)$ in the Poincaré ball, a space of constant negative curvature that naturally organizes hierarchical concepts. The argument runs through the curvature-aware kernel $K(e_i,e_j)=\exp\left(-d_{\kappa_{ij}}(\mu_i,\mu_j)^2/(r_i^2+r_j^2+\epsilon)\right)$, where $\kappa_{ij}$ is a harmonic-mean effective curvature and $d_{\kappa_{ij}}$ is the geodesic distance under that curvature. This kernel is the similarity primitive: it is multiplied along a semantic path $u \to c_1 \to \dots \to c_k \to i$, and the final recommendation score is the maximum over paths, found by beam search of width 5 and depth 3. Learnable curvature and radius per entity let geometry and uncertainty shape the reasoning, while log-space products and Riemannian optimization keep the model trainable.

What would settle it

Take two non-collinear points in the Poincaré ball and evaluate the Section 4.1 distance formula: if the argument $\|\mu_i-\mu_j\|/(1-\kappa_{ij}\langle\mu_i,\mu_j\rangle)$ ever equals or exceeds 1, the $\tanh^{-1}$ is undefined or imaginary, so the formula cannot be the model's geodesic distance. Separately, numerically differentiate the kernel $K(e_i,e_j)$ with respect to $r_i$; if $\partial K/\partial r_i >0$ anywhere, larger radii raise similarity, contradicting the paper's stated uncertainty mechanism.

Watch

Extended reading notes

Core claim

The paper sets out to show that recommendation can be treated as geometric, semantic reasoning rather than direct interaction fitting. Each user, item, and tag is embedded as a sphere $(\mu_e, r_e, \kappa_e)$ in the Poincaré ball, where $r_e$ encodes epistemic uncertainty and $\kappa_e<0$ is a learnable curvature. The score of a candidate item is the maximum, over tag paths $u \to c_1 \to \dots \to c_k \to i$, of the product of pairwise kernel values $K(e_j,e_{j+1}) = \exp(-d^2/(r_j^2+r_{j+1}^2+\epsilon))$, with a curvature-aware distance between sphere centers. On GoodBooks, MIND-small, Book-Crossing, and Avicenna-Syllogism, the paper reports NDCG@10 of 0.473, Recall@10 of 0.591, ECE of 0.103, and Diversity@10 of 0.294, each better than the compared baselines, together with 100% tag alignment and explanation coverage in its interpretability evaluation.

Load-bearing premise

The whole result stands on the Section 4.1 formulas doing what the text says: the distance must be a genuine hyperbolic distance and a larger uncertainty radius must lower the similarity kernel.

Editorial extensions

If this is right

  • If the reported gains hold, shifting from fixed-curvature embeddings to adaptive-curvature probabilistic spheres should improve ranking in tag-rich domains.
  • The low calibration error implies the model's confidence scores can be read as probabilities, which would allow downstream systems to threshold recommendations by trust.
  • Beam-searched tag paths give a concrete reason for each recommendation, so users can inspect and verify why an item was suggested.
  • The diversity gains suggest that hopping through intermediate tags counteracts over-personalization and redundant recommendation lists.
  • The reported training efficiency, 7.2 seconds per epoch with convergence in 34 epochs, indicates that interpretable multi-hop reasoning need not be computationally prohibitive.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the same curvature-aware kernel and path-scoring scheme could be transferred to knowledge-graph completion or multi-hop question answering, where entity-specific curvature and confidence radii are natural.
  • Beyond the paper: the authors' own Limitations section flags reliance on high-quality semantic tags and additional inference overhead, so the practical boundary of the method is tag availability and latency.
  • Beyond the paper: because diversity is achieved by hopping through intermediate tags, the method predicts that datasets with denser tag coverage will show larger diversity gains; this could be tested by varying the tag vocabulary size.
  • Beyond the paper: the paper does not report sensitivity to beam width or path depth, so ablating $k$ and $b$ would reveal whether the gains come from multi-hop transitivity or from the pairwise kernel geometry alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. ManifoldMind proposes a recommender that embeds users, items, and tags as probabilistic hyperbolic spheres (center, uncertainty radius, and learnable negative curvature). Pairwise similarity is a curvature-aware kernel, and final scores are computed by max-product over multi-hop tag paths found by beam search. The paper reports state-of-the-art NDCG, Recall, ECE, and diversity on four datasets, plus ablations and an interpretability study claiming 100% tag alignment and coverage. The contribution, if valid, would be a transparent, uncertainty-calibrated hyperbolic recommender with diverse reasoning paths.

Significance. The intended contribution is valuable: combining adaptive curvature, explicit uncertainty radii, and symbolically traceable multi-hop reasoning is a plausible direction for trustworthy recommendation. The paper is also commendable for evaluating on four public datasets, providing pseudocode, ablating the main components, and reporting efficiency and diversity metrics. However, the central mathematical definition in Section 4.1 is not a valid hyperbolic distance, and the uncertainty-derivative claim is unsupported. Because every path score and every reported result depends on this kernel, the paper as written cannot support its headline claims. The interpretability metric in Table 5 is also circular by construction. These are not presentation issues; they affect the validity of the core method and all experimental conclusions.

major comments (4)
  1. [§4.1] The displayed formula for dκij is not the geodesic distance of the Poincaré ball with curvature κij. For κ<0 the canonical distance has the form d(x,y)=2/√(-κ) tanh^{-1}(√(-κ)‖(-x)⊕_κ y‖), where ⊕_κ is the Möbius addition; the paper instead writes 2√(-κij) tanh^{-1}(‖μi−μj‖/(1−κij⟨μi,μj⟩)). The printed expression can be ill-defined: for μi=(0.8,0), μj=(−0.8,0), and κij=−1, the tanh^{-1} argument is 1.6/0.36≈4.44, whose inverse hyperbolic tangent is undefined in the real numbers. Since every path score in §4.3 and every entry in Algorithm 1 is built from this quantity, the model as written is not a well-defined hyperbolic recommender, and Tables 2–6 cannot be reproduced from the manuscript.
  2. [§4.1] The bullet claims ∂K/∂r_i<0 and that K approximates a von Mises-Fisher distribution. The derivative claim does not follow from the displayed kernel K=exp(−d²/(r_i²+r_j²+ε)). If d is treated as independent of r_i, then ∂K/∂r_i = K·2d²r_i/(r_i²+r_j²+ε)² > 0 for r_i>0, the opposite of the claim. If d depends on r_i through κij, the derivative acquires the extra term −(1/S)∂(d²)/∂r_i with S=r_i²+r_j²+ε, and its sign is not determined by anything shown in the paper. The implementation is not given, so the reader cannot tell whether the uncertainty mechanism is inverted or undisclosed. The von Mises-Fisher statement is also unsupported, since exp(−d²/(r_i²+r_j²+ε)) is not a standard vMF density.
  3. [§6.4] Table 5's 100% tag alignment and 100% coverage for ManifoldMind are circular. By construction, the explanation path is a sequence of tags from T (Section 4.3 and Algorithm 1), and the training objective optimizes scores over such tag paths. Saying that these paths "contain tags" therefore does not measure semantic alignment with user intent, and comparing against baselines that generate no tag paths (coverage 0) is uninformative. This does not invalidate the ranking experiments by itself, but it invalidates the interpretability and trustworthiness claims made in the abstract and Section 8.
  4. [§6.1 and §5.3] Table 2 reports only averages over five seeds, with no standard deviations, confidence intervals, or significance tests, and no per-dataset breakdown for NDCG, Recall, or ECE. The central claim that ManifoldMind "achieves the best overall performance across all core metrics" and the specific +4.5% NDCG@10 gain over FineRec are therefore not statistically supported; differences of this magnitude can easily arise from seed noise. Please report per-dataset results with variance and a significance test, and state how each baseline was tuned. Without this, the performance claim is not verifiable.
minor comments (5)
  1. [§1] The introduction contains a duplicated sentence: "We evaluate the framework across four open-source datasets ..." appears twice in consecutive sentences. Please remove the duplicate.
  2. [§4.3 and §5.4] The claimed beam-search complexity O(b^k) is inconsistent with the pseudocode in Figure 3, which expands each path by appending tags from the full tag set T; that expansion costs O(b·k·|T|) unless additional pruning is described. Please clarify the actual complexity and the expansion procedure.
  3. [Algorithm 1] Algorithm 1 says "for each semantic path P" while Section 4.3 says beam search with width b=5. Exhaustively iterating over all tag paths is exponential and contradicts the stated complexity; please make the algorithm's search procedure match the described beam search.
  4. [§7] The Limitations section says "though 4 demonstrates resilience to partial data removal," but no experiment on partial data removal appears in the paper. Either add the experiment or remove the reference.
  5. [§5.3] Table 3's column header "A V" should be "AV" for Avicenna. Also, T-ILS@10 results are given per dataset, but Table 2's corresponding diversity numbers are not; please align the reporting to make cross-table comparisons possible.

Circularity Check

1 steps flagged · score 6.0 of 10

The 100% tag-alignment and explanation-coverage results are guaranteed by the definition of an explanation as a tag path; ranking and calibration are held-out and remain non-circular.

  1. self definitional [Section 4.3 (Eq. 1) and Algorithm 1; Section 6.4, Table 5]
    "We define a reasoning path P = {u, c1, ..., ck, v} over k tags cj ∈ T. Its score is: s(u, v) = max_P Π K(ei, ei+1). ... [Algorithm 1:] Save P as top-ranked path. ... ManifoldMind offers symbolic, multi-hop explanations with strong alignment and coverage. Table 5 shows 100% tag alignment and explanation coverage, with the highest confidence among all models."

    Tag alignment and coverage are not defined in the paper, but any such metric must check whether the returned explanation contains or bridges the tag vocabulary. Algorithm 1 constructs every returned explanation as a semantic path whose intermediate nodes are exactly tags from T (P = {u,c1,...,ck,i}, cj ∈ T) and always returns a top-ranked path. Consequently every prediction is guaranteed to have a tag-aligned, covered explanation; the reported 100.0/100.0 restates the output format rather than measuring interpretability against an independent standard. The 'Confidence' value 0.893 is likewise the model's own kernel path score, not an external human or behavioural confidence measure, so comparing it across models with different score scales is not an empirical validation.

full rationale

The core ranking and calibration claims are evaluated on held-out test interactions with standard metrics (NDCG@10, Recall@10, ECE), so they are not circular: the reported scores are not equal to any fitted parameter or construction input. The ablation study also compares model variants on the same held-out protocol. The self-citations in Related Work are background references and are not load-bearing for the central derivation. The Section 4.1 inconsistencies (geodesic formula not equal to the standard Poincaré distance, and the stated derivative sign of the kernel contradicting the printed equation) are serious correctness and reproducibility concerns, but they are not circularity: an incorrect formula is not the same as a prediction that reduces to its input by construction. The genuine circularity is confined to the interpretability claims in Section 6.4: because every explanation is, by definition, a path through the tag set, 'tag alignment' and 'coverage' are satisfied automatically. This makes one family of headline claims (explanation quality) circular by construction, while the ranking and calibration claims retain independent content. Score 6 reflects this partial, construction-level circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a large number of learned parameters (centers, radii, curvatures for every user, item, and tag) and on several unvalidated modeling assumptions. The most fragile assumptions are the correctness of the distance formula and the monotonicity of the kernel with respect to radii; both are contradicted by the paper's own equations. The model also assumes the availability of semantically meaningful tags and that product-kernel path scores correspond to logical reasoning. No new entities are postulated, so the invented-entities ledger is empty.

free parameters (6)
  • Entity center embeddings μ_i = learned in D^20
    Initialized via exponential map at origin and updated with Riemannian Adam; no analytic solution, learned from data.
  • Entity uncertainty radii r_i = learned via softplus
    Learnable raw radius passed through softplus; the paper's claim that larger radii down-weight similarity is contradicted by the kernel equation.
  • Entity curvatures κ_i = uniform(-5,-0.01) init, learned
    Learnable per-entity curvature, regularized by R_curv with an unreported δ; the text claims smaller radii dominate the combined curvature, but the harmonic mean formula is dominated by larger radii.
  • Margin γ = not reported
    Used in the pairwise margin loss L_rank; value not specified in the paper.
  • Regularization weight λ and offset δ = not reported
    Used in R_curv; values not given.
  • Beam width b, path length k, dimension d = b=5, k=3, d=20
    Hand-chosen hyperparameters; no sensitivity analysis is provided.
assumptions (5)
  • ad hoc to paper Poincaré ball distance formula in Section 4.1 is the correct geodesic distance for the effective curvature κij
    The paper uses this formula, but it does not match the standard Poincaré distance in higher dimensions and can produce invalid artanh inputs.
  • ad hoc to paper Larger radii r_i imply lower similarity
    Used to justify uncertainty modeling, but the printed kernel gives the opposite monotonicity.
  • domain assumption Product of kernel similarities along tag paths represents transitive semantic reasoning
    The paper assumes that multiplying pairwise kernel values along a tag chain yields a meaningful confidence score; no derivation or external evidence is provided.
  • domain assumption High-quality semantic tags are available for all entities and are aligned with user-item interactions
    Acknowledged in the Limitations section as a dependency; the tag extraction pipeline is only described as 'lightweight rules'.
  • ad hoc to paper The kernel approximates a von Mises-Fisher distribution
    Stated in Section 4.1 without derivation or a precise link to directional statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ManifoldMind: Dynamic Hyperbolic Reasoning for Trustworthy Recommendations." pith.science (2026). https://pith.science/paper/BIRNAZ7A

@misc{pith2026250702014,
  author       = {Pith},
  title        = {Pith review of: ManifoldMind: Dynamic Hyperbolic Reasoning for Trustworthy Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIRNAZ7A}},
  note         = {Machine review of arXiv:2507.02014}
}
read the original abstract

We introduce ManifoldMind, a probabilistic geometric recommender system for exploratory reasoning over semantic hierarchies in hyperbolic space. Unlike prior methods with fixed curvature and rigid embeddings, ManifoldMind represents users, items, and tags as adaptive-curvature probabilistic spheres, enabling personalised uncertainty modeling and geometry-aware semantic exploration. A curvature-aware semantic kernel supports soft, multi-hop inference, allowing the model to explore diverse conceptual paths instead of overfitting to shallow or direct interactions. Experiments on four public benchmarks show superior NDCG, calibration, and diversity compared to strong baselines. ManifoldMind produces explicit reasoning traces, enabling transparent, trustworthy, and exploration-driven recommendations in sparse or abstract domains.

Figures

Figures reproduced from arXiv: 2507.02014 by the authors.

Figure 1
Figure 1. Overview of ManifoldMind. Users, tags, and items are embedded as adaptive spheres (µ, r, κ) in D n . Multi-hop semantic paths are scored using a curvature-aware kernel. enable structured semantic reasoning, we define a multi-hop path Pu→i = {u, c1, . . . , ck, i} through intermediate tags cj ∈ T . The final confidence score is: s(u, i) = max Pu→i Y (ej ,ej+1)∈P K(ej , ej+1). This formulation enables interpretable, t… view at source ↗
Figure 2
Figure 2. Method overview: ManifoldMind embeds users, items, and semantic tags as probabilistic hyperbolic spheres. A curvature￾aware kernel evaluates pairwise similarity over multi-hop semantic paths, and final confidence scores reflect both path quality and uncertainty. 4.1. Probabilistic Hyperbolic Embeddings Given two entities ei , ej , we first compute an effective cur￾vature: κij = ri + rj ri κi + rj κj , This harmonic … view at source ↗
Figure 3
Figure 3. Training pseudocode for ManifoldMind using beam￾based semantic path reasoning. and empirical suitability for capturing hierarchical seman￾tics (Mathieu et al., 2019). 6. Results We evaluate ManifoldMind on four real-world datasets,GoodBooks, MIND, Avicenna, and Book￾Crossing, using standard metrics for classification (NDCG @ 10, Recall @ 10), calibration (ECE), diversity (Diversity @ 10, T-ILS) and interpretability.… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RicciFlowRec: A Geometric Root Cause Recommender Using Ricci Curvature on Financial Graphs

    cs.LG 2025-08 reject novelty 4.0 of 10

    A curvature and flow based recommender that attributes financial shocks to source nodes and re-ranks stocks by structural risk reports gains on S&P 500 data, but its attribution test is partly self-referential and sev...

Reference graph

Works this paper leans on

30 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Talebpour, A

    Aghahadi, Z. and Talebpour, A. Avicenna: a challenge dataset for natural language generation toward commonsense syllogistic reasoning. Journal of Applied Non-Classical Logics, 32 0 (1): 0 55--71, 2022

    work page 2022

  3. [3]

    and Tay, F

    Cao, L.-J. and Tay, F. E. H. Support vector machine with adaptive parameters in financial time series forecasting. IEEE Transactions on neural networks, 14 0 (6): 0 1506--1518, 2003

    work page 2003

  4. [4]

    Hyperbolic graph convolutional neural networks

    Chami, I., Ying, Z., R \'e , C., and Leskovec, J. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

    work page 2019

  5. [5]

    Neural reasoning for sure through constructing explainable models, 2025

    Dong, T., Jamnik, M., and Lio, P. Neural reasoning for sure through constructing explainable models, 2025. arXiv:2501.12345

    arXiv 2025

  6. [6]

    Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q. On calibration of modern neural networks. In International conference on machine learning, pp.\ 1321--1330. PMLR, 2017

    work page 2017

  7. [7]

    and Sun, Z

    Harit, A. and Sun, Z. Causal spherical hypergraph networks for modelling social uncertainty. arXiv preprint arXiv:2506.17840, 2025

    arXiv 2025

  8. [8]

    Harit, A., Sun, Z., Yu, J., and Moubayed, N. A. Breaking down financial news impact: A novel ai approach with geometric hypergraphs. arXiv preprint arXiv:2409.00438, 2024

    arXiv 2024

Show all 30 references
  1. [9]

    Neural collaborative filtering

    He, X., Liao, L., Zhang, H., Nie, L., Hu, X., and Chua, T.-S. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web, pp.\ 173--182, 2017

    2017

  2. [10]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    He, X., Deng, K., Wang, X., Li, Y., Zhang, Y., and Wang, M. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pp.\ 639--648, 2020

    2020

  3. [11]

    Geoopt: Riemannian optimization in pytorch

    Kochurov, M., Karimov, R., and Kozlukov, S. Geoopt: Riemannian optimization in pytorch. arXiv preprint arXiv:2005.02819, 2020

    2005 arXiv

  4. [12]

    K., and Huo, H

    Li, A., Yang, B., Hussain, F. K., and Huo, H. Hsr: Hyperbolic social recommender. Information Sciences, 585: 0 275--288, 2022

    2022

  5. [13]

    Hyperbolic graph neural network for temporal knowledge graph completion

    Li, Y., Zhang, X., Cui, Y., and Ma, S. Hyperbolic graph neural network for temporal knowledge graph completion. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pp.\ 8474--8486, 2024

    2024

  6. [14]

    Csrec: Rethinking sequential recommendation from a causal perspective

    Liu, X., Yuan, J., Zhou, Y., Li, J., Huang, F., and Ai, W. Csrec: Rethinking sequential recommendation from a causal perspective. arXiv preprint arXiv:2409.05872, 2024

    2024 arXiv

  7. [15]

    J., Tomioka, R., and Teh, Y

    Mathieu, E., Le Lan, C., Maddison, C. J., Tomioka, R., and Teh, Y. W. Continuous hierarchical representations with poincar \'e variational auto-encoders. Advances in neural information processing systems, 32, 2019

    2019

  8. [16]

    and Kiela, D

    Nickel, M. and Kiela, D. Poincar \'e embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017

    2017

  9. [17]

    Neural matrix factorization++ based recommendation system

    Ong, K., Ng, K.-W., and Haw, S.-C. Neural matrix factorization++ based recommendation system. F1000Research, 10: 0 1079, 2021

    2021

  10. [18]

    Bpr: Bayesian personalized ranking from implicit feedback

    Rendle, S., Freudenthaler, C., Gantner, Z., and Schmidt-Thieme, L. Bpr: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618, 2012

    2012 arXiv

  11. [19]

    I., Yu, J., Shi, L., and Al Moubayed, N

    Sun, Z., Harit, A., Cristea, A. I., Yu, J., Shi, L., and Al Moubayed, N. Contrastive learning with heterogeneous graph attention networks on short text classification. In 2022 International Joint Conference on Neural Networks (IJCNN), pp.\ 1--6. IEEE, 2022

    2022

  12. [20]

    I., Wang, J., and Lio, P

    Sun, Z., Harit, A., Cristea, A. I., Wang, J., and Lio, P. Money: Ensemble learning for stock price movement prediction via a convolutional network with adversarial hypergraph model. AI Open, 4: 0 165--174, 2023 a

    2023

  13. [21]

    I., Wang, J., and Lio, P

    Sun, Z., Harit, A., Cristea, A. I., Wang, J., and Lio, P. A rewiring contrastive patch performermixer framework for graph representation learning. In 2023 IEEE International Conference on Big Data (BigData), pp.\ 5930--5939. IEEE Computer Society, 2023 b

    2023

  14. [22]

    Actionable interpretability via causal hypergraphs: Unravelling batch size effects in deep learning

    Sun, Z., Harit, A., and Lio, P. Actionable interpretability via causal hypergraphs: Unravelling batch size effects in deep learning. arXiv preprint arXiv:2506.17826, 2025 a

    2025 arXiv

  15. [23]

    Advanced hypergraph mining for web applications using sphere neural networks

    Sun, Z., Harit, A., Yu, J., Wang, J., and Li \`o , P. Advanced hypergraph mining for web applications using sphere neural networks. In Companion Proceedings of the ACM on Web Conference 2025, pp.\ 1316--1320, 2025 b

    2025

  16. [24]

    Z., Zhang, Y., and Yao, L

    Wang, S., Chen, X., Sheng, Q. Z., Zhang, Y., and Yao, L. Causal disentangled variational auto-encoder for preference understanding in recommendation. In Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval, pp.\ 1874--...

    2023

  17. [25]

    Neural graph collaborative filtering

    Wang, X., He, X., Wang, M., Feng, F., and Chua, T.-S. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, pp.\ 165--174, 2019

    2019

  18. [26]

    Mind: A large-scale dataset for news recommendation

    Wu, F., Qiao, Y., Chen, J.-H., Wu, C., Qi, T., Lian, J., Liu, D., Xie, X., Gao, J., Wu, W., et al. Mind: A large-scale dataset for news recommendation. In Proceedings of the 58th annual meeting of the association for computational linguistics, pp.\ 3597--3606, 2020

    2020

  19. [27]

    Self-supervised graph learning for recommendation

    Wu, J., Wang, X., Feng, F., He, X., Chen, L., Lian, J., and Xie, X. Self-supervised graph learning for recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pp.\ 726--735, 2021

    2021

  20. [28]

    Goodbooks-10k: a new dataset for book recommendations

    Zajac, Z. Goodbooks-10k: a new dataset for book recommendations. FastML, 2017

    2017

  21. [29]

    Finerec: Exploring fine-grained sequential recommendation

    Zhang, X., Xu, B., Wu, Y., Zhong, Y., Lin, H., and Ma, F. Finerec: Exploring fine-grained sequential recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\ 1599--1608, 2024

    2024

  22. [30]

    M., Konstan, J

    Ziegler, C.-N., McNee, S. M., Konstan, J. A., and Lausen, G. Improving recommendation lists through topic diversification. In Proceedings of the 14th international conference on World Wide Web, pp.\ 22--32, 2005

    2005

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.