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

Anoushka Harit; Zhongtian Sun

REVIEW 5 major objections 5 minor 23 references

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

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read RicciFlowRec uses Ricci curvature gradients on dynamic financial graphs to locate shock origins and re-rank assets away from unstable regions, reporting better stability and attribution fidelity than five baselines on S&P 500 stress tests.

desk verdict A novel geometric recommender idea, but the root-cause evaluation is circular and the reported metrics are under-defined; worth referee time but not acceptance as is. read the letter →

arxiv 2508.09334 v1 pith:M3I2MTNE submitted 2025-08-12 cs.LG cs.AIcs.IR

Zhongtian Sun , Anoushka Harit This is my paper

classification cs.LGcs.AIcs.IR

keywords Riccicurvatureflowrootcauseanalysisfinancialgraphrisk-awarerecommendationstructuralrisksentimentS&P500

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

This paper proposes a recommendation framework that models the S&P 500 as a dynamic graph of stocks, macroeconomic indicators, and news entities, with edges built from return correlations, semantic similarity, and sector knowledge. Its central claim is that Ricci curvature gradients reveal where stress originates and how it propagates, so simulated Ricci flow can trace root-cause paths and flag structurally unstable zones. Those geometric signals are added to return forecasts as a risk penalty, producing rankings that are more stable under synthetic shocks. On 2018-2023 S&P 500 data with FinBERT-derived news sentiment, the method reports higher ranking quality, lower ranking volatility under stress, and higher root-cause attribution fidelity than five baselines, with ablations showing that both the flow and the risk penalty contribute. The authors present this as ongoing work, and the stress test is synthetic.

What carries the argument

Discrete Ricci curvature and Ricci flow on a time-indexed heterogeneous graph. The curvature measures whether an edge is a structural bottleneck (negative curvature signals fragility), and the flow rule $d w_{uv}/dt = -\kappa(u,v) w_{uv}$ penalizes fragile connections over a rolling window. The curvature shift $\Delta\kappa$ between time steps is the signal used to identify unstable nodes, trace root-cause paths by backward search, and define the structural risk exposure $\rho(a_i)$ that adjusts the ranking score.

What would settle it

Inject a shock through a variable held out of graph construction, for example a latent fundamental change (earnings surprise or liquidity shock) that only affects returns after a delay, so rolling correlations and sentiment embeddings are unchanged at injection time. If curvature backtracking still localizes the true source, the attribution is signal-driven; if RCA fidelity collapses to baseline levels, the reported 0.78 measures edge-feature perturbation, not causal root-cause discovery.

Watch

Extended reading notes

Core claim

The paper's central claim is that curvature gradients reveal causal substructures in financial graphs, and that penalizing exposure to curvature-identified unstable regions improves ranking robustness. Concretely, RicciFlowRec computes edge-level Ricci curvature, evolves edge weights under Ricci flow, and defines a structural exposure for each asset as the sum of absolute curvature shifts over its incident edges. The final score is a weighted combination of predicted return and this exposure, and the top-K assets are returned with a path-traced explanation subgraph. The authors report that this outperforms all five baselines on S&P 500 data across ranking quality, ranking stability, and root

Load-bearing premise

The stress test treats a synthetic shock—a jump in realized volatility and negative sentiment—as ground truth for root-cause identification, but these are the same kinds of features used to build the graph edges, so high RCA fidelity may just mean the perturbation moved the curvature it was designed to move.

Editorial extensions

If this is right

If curvature shifts trace shock propagation, then re-ranking assets by structural exposure should reduce top-10 ranking volatility under market stress relative to returns-only or sentiment-only rankers.
The reported RCA fidelity of 0.78 versus 0.65 for the strongest attribution-capable baseline implies that curvature-flow backtracking locates perturbed nodes better than self-supervised or counterfactual graph methods in this setup.
Ablation results show both components matter: removing Ricci flow drops attribution fidelity from 0.780 to 0.620, and removing the RCA penalty drops it from 0.780 to 0.440.
Because inference updates are reported to complete in under 200 ms, real-time stress attribution and re-ranking on a 450-stock universe is computationally plausible.

Reading between the lines

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

A stronger attribution test would inject shocks through variables excluded from graph construction, such as a latent earnings surprise that changes returns only after a delay, separating geometric signal from the feature-mover effect of the current perturbation protocol.
Comparing Ollivier-Ricci and Forman-Ricci on the same stress protocol would show whether the attribution claim depends on the choice of discrete curvature, since the write-up defines the former and implements the latter.
The mechanism is not finance-specific: the same flow-and-penalty recipe could be tested on supply-chain, epidemic, or social-influence networks where localized shocks propagate along measurable edges.
If curvature shifts remain informative under non-synthetic historical shocks, the RCA paths could double as event explanations, something the current synthetic-shock evaluation does not yet establish.

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.

Desk Editor's Note

read the letter

RicciFlowRec is a clear idea with a circular empirical core: its headline root-cause attribution result is mostly an artifact of how the shocks are injected, and the reporting is too loose to back the ranking claims.

What's actually new: the integration of discrete Ricci curvature and Ricci flow into a recommendation scoring function, with backtracked 'RCA paths' as explanations, doesn't appear in the cited literature. The paper is upfront about being ongoing work, and the pipeline is easy to follow.

The soft spots are real and load-bearing. The RCA fidelity metric is a self-consistency check, not causal evidence. The injection protocol (§4.4) increases realized volatility and negative sentiment, which are the same signals used to build the correlation and sentiment edges (§4.3). Perturbing a node therefore moves the curvature of its incident edges by construction, and a backward BFS along high-|Δκ| edges will naturally land on that node. A fidelity of 0.78 tells us the curvature update is internally consistent, not that it detects an independently defined root cause. The paper needs a perturbation that doesn't touch the edge-construction features, or a ground truth from a generative model, to make the claim.

The rest of the evaluation is under-specified. Table 1 is reported on 2021-2022, the same validation window used for grid search, so the numbers include selection bias. NDCG@10 ground truth and the base forecaster r̂ behind eq. (1) are never defined. AUC appears in Tables 4 and 5 without saying what is being classified. The paper also contradicts itself on curvature: §3.2 defines Ollivier-Ricci via Wasserstein distance, but §4.7 and Algorithm 1 use Forman-Ricci. Since the numbers depend on the curvature definition, that's a reproducibility blocker. No code or data are supplied.

The lower volatility under perturbation is a plausible directional result, but the same tuning-window issue means it shouldn't be taken at face value. The scoring function also changes between eq. (1) and §3.4 (alpha vs lambda), which adds to the confusion.

Who this is for: readers interested in geometric methods for financial recommenders or interpretable risk scoring will find the framework thought-provoking. But this is an idea paper, not a validated method. I'd send it to peer review because the idea deserves referee time, but with the current evidence I'd expect a reject or a heavy revision. A clean holdout split, a non-circular attribution benchmark, consistent curvature, and released code would change that.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes RicciFlowRec, a framework for financial asset recommendation with root-cause attribution. It constructs a daily heterogeneous graph over S&P 500 stocks, macro indicators, and news entities, with edges from 30-day rolling return correlations, FinBERT semantic similarity, and sector/knowledge links; computes discrete Ricci curvature; simulates Ricci flow; identifies unstable nodes from curvature shifts; and scores assets by a weighted combination of a predicted return r̂ and a structural risk term. Experiments over 2018–2023 with synthetic volatility/sentiment injections are reported on NDCG@10, Top-10 Volatility, and RCA Fidelity, with RicciFlowRec outperforming five baselines (e.g., NDCG@10 0.512 vs 0.488 for SS-GNN). Ablations and hyperparameter sensitivity tables for α and θ are included. The abstract explicitly labels the results preliminary and the work ongoing.

Significance. The idea of using Ollivier/Forman–Ricci curvature as an interpretable, geometry-based risk signal in financial recommendation is potentially interesting; if a clean causal-validation study existed, curvature-gradient tracing could be a useful addition to the interpretability toolbox. The paper is transparent about many implementation parameters and reports runtime figures, which is helpful for reproducibility. However, the evidence as presented does not support the central causal attribution claim: the RCA fidelity test is confounded by construction, the main ranking comparison uses the validation period, and the base forecaster and NDCG relevance definition are unspecified. The contribution is therefore currently a prototype with an illustrative evaluation, not a validated method.

major comments (5)

[§4.4, §4.3, §3.3, Table 1] RCA Fidelity is not an independent test of root-cause attribution. The stress protocol in §4.4 injects shocks by 'increasing their realised volatility and negative sentiment scores,' while the graph in §4.3 is built from 30-day rolling return correlations, FinBERT semantic similarity, and knowledge links. The curvature shift Δκ in §3.3 is computed on these edges, so perturbing a node's own volatility and sentiment moves the incident-edge weights by construction. The backward search in §4.8 then retraces the perturbation through the same signal that created it. A fidelity of 0.78 (Table 1) therefore measures internal consistency of the curvature update, not the ability to localize an independently defined cause. A credible test would inject shocks through variables not used to construct edges (e.g., an idiosyncratic earnings shock orthogonal to sentiment and returns) or validate on histor
[§4.5, Table 1] The main numerical comparison is performed on the validation set used for hyperparameter selection. §4.5 states that all methods use train/val/test splits 'with 2021–2022 as validation' and that hyperparameters are tuned via grid search using NDCG@10 on the validation set. Table 1 is explicitly labelled '2021–2022.' Reporting performance on the same period used for model selection makes the NDCG and volatility advantages in-sample results; they do not establish out-of-sample ranking robustness. The authors should report the held-out test-period metrics and, ideally, multiple seeds with error bars.
[§3.4, §4.8, Eq. (1), §4.5] The scoring function is not fully specified. Eq. (1) defines s(ai)=α·r̂(ai)−(1−α)·Risk(ai), while §3.4 defines s(ai)=r̂_i−λ·ρ(ai), and Algorithm 1 uses the Eq. (1) form with α=0.7. More importantly, the base forecaster r̂ that produces predicted returns is never described: no architecture, training procedure, feature set, or hyperparameters are given ('e.g., LSTM, GAT' in §3.4 is not an implementation). Since every NDCG@10 number depends on r̂, the reader cannot reproduce or interpret the ranking results. The relevance/ground-truth definition for NDCG (what counts as a relevant asset in the test period) is also missing from §4.6.
[§3.2, §4.7, Algorithm 1] The curvature definition is inconsistent. §3.2 and §4.3 define Ollivier–Ricci curvature with Wasserstein-1 distance, while §4.7 lists 'Ricci curvature type: Forman' and Algorithm 1 calls compute_forman_ricci; the cited GraphRicciCurvature package implements Ollivier–Ricci. These are different quantities, and the sign/threshold behavior in §4.7/§4.8 (θ=-0.05 with |Δκ|>θ, but Algorithm 1 uses avg_curv_change(v)<θ) compounds the ambiguity. The exact theoretical quantity must be fixed for the RCA claims to be checkable.
[Table 1, §4.6] RCA Fidelity for the baselines is not defined. Table 1 reports RCA Fidelity of 0.58, 0.62, and 0.65 for CausalRec, FinGNN, and SS-GNN, but §4.6 defines RCA Fidelity as 'percentage of perturbed nodes successfully identified via curvature backtracking.' No curvature backtracking is described for these baselines, and no alternative attribution mechanism is specified. Without a stated baseline RCA procedure, the comparison in Table 1 is not interpretable. Also, Tables 4 and 5 report 'AUC' even though AUC is not defined in §4.6 and no ROC-type task is described.

minor comments (5)

[§4.3, §4.7] The top-k edge retention in §4.3 is used to threshold edges, but the value of k is not listed in the hyperparameter block in §4.7. Please state it.
[§5.2, Table 2] The qualitative attribution examples in Table 2 are anecdotes: no quantitative precision/recall is given, and the 'RCA source' is the method's own output. These examples illustrate behavior but do not validate correctness.
[§4.4, §5.1] No standard deviations, seeds, or significance tests are reported. All numerical claims are point estimates from what appears to be a single run, which is insufficient for comparing ranking methods.
[§4.3] Calling sector-level knowledge links 'Causal structure' is misleading; these are relational/semantic links, not identified causal relations. The terminology should be qualified throughout.
[§4.1, §4.2] The filtering from 500 to 450 stocks is not justified; if the filter removes stocks with missing data, survivorship-bias risks should be discussed. The tweet-to-ticker linking rule ('$APPLE', company name, or industry keyword) is also prone to false positives and should be evaluated.

Circularity Check

1 steps flagged · score 6.0 of 10

RCA fidelity is a self-consistency check: the synthetic shock is injected through the same price/sentiment streams that define graph edges, so curvature backtracking rediscovers the perturbed node by construction.

self definitional [§3.3 Simulating Ricci Flow, §4.3 Graph Construction, §4.4 Stress Testing Protocol, §4.8 RicciFlowRec Algorithm]
"Edges(u,v)∈E_t are constructed using three signals: (1) Rolling correlations: Pearson correlation of log returns over a 30-day window; (2) Semantic proximity: Cosine similarity between FinBERT embeddings of co-mentioned news ... To evaluate model robustness, we inject synthetic volatility shocks into randomly selected nodes by increasing their realised volatility and negative sentiment scores ... Δκ(u,v)=κ_{t+Δt}(u,v)−κ_t(u,v) ... Edges with large |Δκ| signal temporal instability. We extract RCA paths by tracing high-magnitude curvature shifts backward through G_t."

The perturbed-node ground truth (RCA source) is created by increasing the same realized-volatility and FinBERT-sentiment features that define edge weights in §4.3. Discrete Ricci curvature and Δκ are functions of those edge weights, so a node whose own features are changed necessarily has large |Δκ| on incident edges. The backward BFS in §4.8 then follows |Δκ|>θ edges back to the perturbed node. Thus RCA Fidelity 0.78 only confirms that the algorithm can detect the feature change it injected; it does not test whether curvature identifies an independently defined root cause. The abstract's claim that 'curvature gradients reveal causal substructures' is therefore self-referential by design.

full rationale

The paper's central contribution is root cause attribution via Ricci curvature, and the only quantitative evidence for it is RCA Fidelity under synthetic shocks. That evaluation is circular: §4.4 injects a shock by increasing realized volatility and negative sentiment, which are exactly the features used in §4.3 to build correlation and semantic edges; κ and Δκ are computed on those edges; and §4.8 recovers the perturbed node by following large |Δκ|. The fidelity number therefore measures internal consistency of the curvature computation, not the ability to localize an externally defined cause. The NDCG@10 and Top-10 Volatility results are less affected and do provide a legitimate comparison against baselines, but they support ranking robustness, not the causal substructure conclusion. No exogenous shock or independent ground-truth causal model is provided. Self-citations to prior work by the same authors appear but are not the load-bearing argument here. Because the causal claim reduces to a self-consistency check while secondary ranking results retain independent content, a score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The paper's contribution rests on a chain of modeling choices: graph edges built from correlations and sentiment, curvature as a stress signal, Ricci flow as a propagation model, and an injection-based ground truth for root causes. Each link is asserted rather than independently tested. The unspecified base forecaster and edge-retention threshold mean several numbers that enter the scoring function are unaccounted for.

free parameters (6)

Return-stability weight alpha = 0.7
Grid-searched on validation NDCG@10; directly defines the score trade-off in eq (1) and Algorithm 1; all reported results depend on it.
Curvature-change threshold theta = -0.05
Determines which edges/nodes count as unstable for RCA path extraction; sensitivity tested in Table 5; plausible on the Ollivier curvature scale but inconsistent with the Forman setting in §4.7.
Risk sensitivity lambda (§3.4) = unspecified
Appears only in the s(ai)=r̂-lambda·rho form of §3.4, which is inconsistent with eq (1) and Algorithm 1; its value is never given though it controls all rankings in that formulation.
Edge retention top-k = unspecified
§4.3 says edges are 'thresholded to retain only the top-k significant relations' but k is never stated; graph sparsity determines all curvature values.
RCA stop-criterion epsilon = unspecified
Algorithm 1 stops path search when 'curvature decay falls below epsilon'; epsilon is never given and affects RCA paths and fidelity.
Base forecaster hyperparameters = unspecified
r̂ is produced by an unnamed base forecaster (§3.4); its architecture and parameters are never stated though they determine the NDCG results.

assumptions (6)

domain assumption Negative Ollivier-Ricci or Forman curvature indicates structural fragility and root-cause potential in financial networks
§3.2 interprets negative curvature as 'structural divergence or bottleneck' and root cause of systemic stress; this bridge from graph geometry to financial causality is asserted, not derived or validated against real shock episodes.
domain assumption Ricci-flow dynamics d w/dt = -kappa·w model how real market shocks propagate through the network
§3.3: the evolution equation is a geometric metaphor; no evidence that financial stress follows curvature-driven dynamics rather than funding-liquidity, order-flow, or other channels.
domain assumption Edges built from 30-day rolling return correlations, FinBERT semantic similarity, and sector links capture causal interdependencies
§3.1 and §4.3: Pearson correlation is explicitly non-causal; the causal reading of these edges is asserted without support.
domain assumption Synthetic volatility and sentiment shocks on random nodes define the true root cause
§4.4: ground truth in the RCA test is defined by the injection; the protocol is not validated against naturally occurring shock episodes with independently known causes. The only qualitative episode (2022-03-08 NVIDIA) is illustrative.
standard math Wasserstein-1 distance and Ollivier-Ricci definitions used in §3.2 are the standard ones
Definitions from [11,12] are standard; no new mathematics is claimed.
ad hoc to paper A trained base forecaster r̂ exists and generates the return predictions used in scoring
§3.4 posits 'a base forecaster (e.g., LSTM, GAT)' but the paper never specifies which one, how it is trained, or how it interacts with the penalty; the NDCG results depend on it.

invented entities (1)

Curvature-gradient RCA path (root-cause attribution path)
purpose: A backward BFS path through edges with |Δκ|>θ, attached to each recommended asset as a causal explanation of ranking adjustment (§3.3, §3.5, Algorithm 1).
The path is only validated against synthetically injected shocks that alter the same features used to construct the graph, so the 'root cause' it names is defined by the injection protocol; no independent causal ground truth (for example, documented historical episodes) is provided. It functions as an in-method artifact rather than an entity with external evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RicciFlowRec: A Geometric Root Cause Recommender Using Ricci Curvature on Financial Graphs." pith.science (2026). https://pith.science/paper/M3I2MTNE

@misc{pith2026250809334,
  author       = {Pith},
  title        = {Pith review of: RicciFlowRec: A Geometric Root Cause Recommender Using Ricci Curvature on Financial Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3I2MTNE}},
  note         = {Machine review of arXiv:2508.09334}
}

read the original abstract

We propose RicciFlowRec, a geometric recommendation framework that performs root cause attribution via Ricci curvature and flow on dynamic financial graphs. By modelling evolving interactions among stocks, macroeconomic indicators, and news, we quantify local stress using discrete Ricci curvature and trace shock propagation via Ricci flow. Curvature gradients reveal causal substructures, informing a structural risk-aware ranking function. Preliminary results on S\&P~500 data with FinBERT-based sentiment show improved robustness and interpretability under synthetic perturbations. This ongoing work supports curvature-based attribution and early-stage risk-aware ranking, with plans for portfolio optimization and return forecasting. To our knowledge, RicciFlowRec is the first recommender to apply geometric flow-based reasoning in financial decision support.

Figures

Figures reproduced from arXiv: 2508.09334 by the authors.

**Figure 1.** Dynamic financial graph G𝑡 with Ricci curvature 𝜅(𝑒) shown on edges (red: unstable, blue: stable). Nodes represent assets, macro indicators, and news entities. Unstable zones inform 𝜌 (𝑎𝑖 , G𝑡) for scoring and Top-𝐾 recommendations. Challenges. (1) Structural risk emerges from long-range dependencies that are not captured by local neighbourhoods. (2) Most systems fail to attribute volatility to root causes in the … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages

[1]

Dogu Araci. 2019. Finbert: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063 (2019)

arXiv 2019
[2]

Stephen Bonner and Flavian Vasile. 2018. Causal embeddings for recommendation. In Proceedings of the 12th ACM conference on recommender systems . 104–112

work page 2018
[3]

Kai Chen, Yi Zhou, and Fangyan Dai. 2015. A LSTM-based method for stock returns prediction: A case study of China stock market. In2015 IEEE international conference on big data (big data) . IEEE, 2823–2824

work page 2015
[4]

Thomas Fischer and Christopher Krauss. 2018. Deep learning with long short- term memory networks for financial market predictions. European journal of operational research 270, 2 (2018), 654–669

work page 2018
[5]

Anoushka Harit and Zhongtian Sun. 2025. Causal Spherical Hypergraph Net- works for Modelling Social Uncertainty. arXiv preprint arXiv:2506.17840 (2025)

arXiv 2025
[6]

Anoushka Harit, Zhongtian Sun, and Suncica Hadzidedic. 2025. ManifoldMind: Dynamic Hyperbolic Reasoning for Trustworthy Recommendations. arXiv preprint arXiv:2507.02014 (2025)

work page Pith review arXiv 2025
[7]

Anoushka Harit, Zhongtian Sun, Jongmin Yu, and Noura Al Moubayed. 2024. Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset. In NeurIPS 2024 Workshop on Behavioral Machine Learning

work page 2024
[8]

Anoushka Harit, Zhongtian Sun, Jongmin Yu, and Noura Al Moubayed. 2024. Breaking Down Financial News Impact: A Novel AI Approach with Geometric Hypergraphs. arXiv preprint arXiv:2409.00438 (2024)

work page Pith review arXiv 2024

Show all 23 references

[9]

Tingsong Jiang and Andy Zeng. 2023. Financial sentiment analysis using FinBERT with application in predicting stock movement. arXiv preprint arXiv:2306.02136 (2023)

2023 arXiv
[10]

Youngbin Lee, Yejin Kim, Javier Sanz-Cruzado, Richard Mccreadie, and Yongjae Lee. 2024. Stock Recommendations for Individual Investors: A Temporal Graph Network Approach with Mean-Variance Efficient Sampling. In Proceedings of the 5th ACM International Conference on AI in Fina...

2024
[11]

Chien-Chun Ni, Yu-Yao Lin, Jie Gao, Xianfeng David Gu, and Emil Saucan. 2015. Ricci curvature of the internet topology. In 2015 IEEE conference on computer communications (INFOCOM). IEEE, 2758–2766

2015
[12]

Chien-Chun Ni, Yu-Yao Lin, Feng Luo, and Jie Gao. 2019. Community detection on networks with Ricci flow. Scientific reports 9, 1 (2019), 9984

2019
[13]

Helder Rojas and David Dias. 2021. Stress testing network reconstruction via graphical causal model. Applied Stochastic Models in Business and Industry 37, 1 (2021), 74–83

2021
[14]

Zhongtian Sun, Anoushka Harit, Alexandra Cristea, Christl A Donnelly, and Pietro Liò. 2025. GLANCE: Graph Logic Attention Network with Cluster En- hancement for Heterophilous Graph Representation Learning. arXiv preprint arXiv:2507.18521 (2025)

2025
[15]

Zhongtian Sun, Anoushka Harit, Alexandra I Cristea, Jingyun Wang, and Pietro Lio. 2023. Money: Ensemble learning for stock price movement prediction via a convolutional network with adversarial hypergraph model. AI Open 4 (2023), 165–174

2023
[16]

Zhongtian Sun, Anoushka Harit, Alexandra I Cristea, Jingyun Wang, and Pietro Lio. 2023. A Rewiring Contrastive Patch PerformerMixer Framework for Graph Representation Learning. In 2023 IEEE International Conference on Big Data (Big- Data). IEEE Computer Society, 5930–5939

2023
[17]

Zhongtian Sun, Anoushka Harit, and Pietro Lio. 2025. Actionable Interpretability via Causal Hypergraphs: Unravelling Batch Size Effects in Deep Learning. arXiv preprint arXiv:2506.17826 (2025)

2025 arXiv
[18]

Zhongtian Sun, Anoushka Harit, Jongmin Yu, Jingyun Wang, and Pietro Liò. 2025. Advanced hypergraph mining for web applications using sphere neural networks. In Companion Proceedings of the ACM on Web Conference 2025 . 1316–1320

2025
[19]

Zhongtian Sun, Jingyun Wang, Ahmed Alamri, and Alexandra Cristea. 2025. SPAR-GNN: Knowledge Tracing with Behavioural Patterns and Selective LLM Feedback. In International Conference on Artificial Intelligence in Education . Springer, 328–335

2025
[20]

Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2017. Graph attention networks. stat 1050, 20 (2017), 10–48550

2017
[21]

Kexin Zhang, Qingsong Wen, Chaoli Zhang, Rongyao Cai, Ming Jin, Yong Liu, James Y Zhang, Yuxuan Liang, Guansong Pang, Dongjin Song, et al. 2024. Self- supervised learning for time series analysis: Taxonomy, progress, and prospects. IEEE transactions on pattern analysis and mac...

2024
[22]

Yongfeng Zhang, Xu Chen, et al. 2020. Explainable recommendation: A survey and new perspectives. Foundations and Trends ® in Information Retrieval 14, 1 (2020), 1–101

2020
[23]

Yaochen Zhu, Jing Yi, Jiayi Xie, and Zhenzhong Chen. 2024. Deep causal reasoning for recommendations. ACM Transactions on Intelligent Systems and Technology 15, 4 (2024), 1–25

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Dogu Araci. 2019. Finbert: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063 (2019)

arXiv 2019

[2] [2]

Stephen Bonner and Flavian Vasile. 2018. Causal embeddings for recommendation. In Proceedings of the 12th ACM conference on recommender systems . 104–112

work page 2018

[3] [3]

Kai Chen, Yi Zhou, and Fangyan Dai. 2015. A LSTM-based method for stock returns prediction: A case study of China stock market. In2015 IEEE international conference on big data (big data) . IEEE, 2823–2824

work page 2015

[4] [4]

Thomas Fischer and Christopher Krauss. 2018. Deep learning with long short- term memory networks for financial market predictions. European journal of operational research 270, 2 (2018), 654–669

work page 2018

[5] [5]

Anoushka Harit and Zhongtian Sun. 2025. Causal Spherical Hypergraph Net- works for Modelling Social Uncertainty. arXiv preprint arXiv:2506.17840 (2025)

arXiv 2025

[6] [6]

Anoushka Harit, Zhongtian Sun, and Suncica Hadzidedic. 2025. ManifoldMind: Dynamic Hyperbolic Reasoning for Trustworthy Recommendations. arXiv preprint arXiv:2507.02014 (2025)

work page Pith review arXiv 2025

[7] [7]

Anoushka Harit, Zhongtian Sun, Jongmin Yu, and Noura Al Moubayed. 2024. Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset. In NeurIPS 2024 Workshop on Behavioral Machine Learning

work page 2024

[8] [8]

Anoushka Harit, Zhongtian Sun, Jongmin Yu, and Noura Al Moubayed. 2024. Breaking Down Financial News Impact: A Novel AI Approach with Geometric Hypergraphs. arXiv preprint arXiv:2409.00438 (2024)

work page Pith review arXiv 2024