Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Next-Generation Conflict Forecasting: Unleashing Predictive Patterns through Spatiotemporal Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read HydraNet, trained only on past conflict fatalities, achieves higher average precision than VIEWS 2020 for all three violence types.

desk verdict HydraNet is a genuinely interesting architecture with a plausible AP edge over VIEWS on conflict history alone, but the abstract overclaims and the out-of-sample wording must be fixed before I'd trust the headline numbers. read the letter →

arxiv 2506.14817 v1 pith:UZTZG4PM submitted 2025-06-08 stat.OT cs.AIcs.CYcs.LGstat.AP

Simon P. von der Maase This is my paper
classification stat.OTcs.AIcs.CYcs.LGstat.AP
keywords conflictforecastingspatiotemporaldeeplearningLSTMU-NetMonteCarlodropoutaverageprecisionVIEWSbenchmarkUCDPGEDrare-eventprediction
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

HydraNet is a deep neural network that forecasts three types of political violence - state-based, non-state, and one-sided - at monthly subnational grid level, up to 36 months ahead, using only the past fatality counts of those same three types as inputs. The paper's central claim is that this single model, with no hand-built features, achieves higher Average Precision than the best VIEWS 2020 ensemble on all three violence types in an out-of-sample test: 0.304 vs 0.272 for state-based, 0.134 vs 0.047 for non-state, and 0.162 vs 0.138 for one-sided violence. It also produces approximate predictive posterior distributions through Monte Carlo Dropout, so each forecast comes with uncertainty samples rather than a single point. The significance, if true, is that manual feature engineering may be capturing only a fraction of the predictive signal in conflict history, and that a unified spatiotemporal architecture can extract the rest automatically.

What carries the argument

The load-bearing component is HydraNet itself: a convolutional encoder-decoder U-Net reorganized inside an LSTM recurrence, with six decoders sharing one encoder. Convolutions learn spatial patterns, U-Net skip connections preserve cell-level precision, LSTM hidden and cell states carry short- and long-term temporal memory, and Monte Carlo Dropout active at test time yields approximate posterior samples. The model is trained on 32x32 spatial patches sampled by a curriculum that starts in conflict-dense regions, using focal loss for the three binary classification tasks, shrinkage loss for the three regression tasks, and an uncertainty-weighted multi-task combination. At forecast time the network switches from observed inputs to its own autoregressive predictions, freezing the cell state as a regularizer, and the mean of 128 dropout samples serves as the point estimate for evaluation.

What would settle it

Recompute the VIEWS 2020 ensemble's average precision on the same 36-month hold-out sample (January 2016 through December 2018, Middle East excluded) using the paper's data snapshot and evaluation code; if the reproduced VIEWS numbers do not match the published values, or if HydraNet's scores change under identical preprocessing, the headline comparison fails.

Watch

Extended reading notes

Core claim

The paper's discovery is that a Monte Carlo Dropout LSTM U-Net trained exclusively on three channels of logged conflict fatalities can out-predict a feature-rich ensemble on the primary evaluation metric. On the VIEWS 2020 evaluation scheme - Africa and the Middle East, January 1990 through December 2019, last 36 months held out, Middle East excluded - HydraNet posts Average Precision of 0.304 versus 0.272 for state-based violence, 0.134 versus 0.047 for non-state violence, and 0.162 versus 0.138 for one-sided violence, while also producing continuous magnitude estimates and 128-sample predictive distributions for each of its six targets. The author interprets this as evidence that past conflict patterns carry far more high-variance predictive signal than manually constructed lags and decay features express, and that learning those patterns directly is sufficient to reach the best reported precision.

Load-bearing premise

The load-bearing premise is that the published VIEWS 2020 scores were computed on exactly the same data version, grid definition, test months, and metric implementation as HydraNet; if that comparability fails, the average-precision advantage could be an artifact.

Editorial extensions

If this is right

  • A model using only three input channels of historical fatalities can beat a feature-rich ensemble on the field's primary precision metric, so manual spatiotemporal feature engineering is not necessary to reach top performance.
  • Jointly predicting probability and magnitude across violence types with shared representations improves rare-event detection, especially for non-state violence, where average precision rises from 0.047 to 0.134.
  • Monte Carlo Dropout gives every forecast a 128-sample approximate posterior, enabling scenario-based uncertainty quantification in 36-month early-warning outputs.
  • The same architecture transfers to other spatiotemporal forecasting tasks, and adding new data channels or auxiliary targets requires no architectural redesign.
  • Forecast precision degrades with horizon, and the appendix shows state-based average precision falling from about 0.6 to below 0.25 by month 36, so long-range forecasts carry weaker signal.

Reading between the lines

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

  • If the strong non-state gain (0.047 to 0.134) is real, it suggests manual features may have actively suppressed rare-event signal; training HydraNet with the VIEWS feature set added as extra channels could isolate the source of the gain.
  • The paper evaluates only the posterior mean; scoring the full Monte Carlo distributions with proper scoring rules, such as the continuous ranked probability score, would test whether the predictive uncertainty is calibrated.
  • The most decisive extension is a shared-code rebenchmark where HydraNet and VIEWS are trained and evaluated inside one pipeline on identical data versions and test months, since the paper currently relies on published VIEWS numbers.
  • The architecture's transferability claim could be tested by applying the same LSTM U-Net to other spatiotemporal rare-event domains, such as disease outbreaks or forced displacement, where fatality-like event counts are the only reliable channel.
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. This paper introduces HydraNet, a Monte Carlo Dropout LSTM U-Net for forecasting three types of political violence (state-based, non-state, one-sided) at the PRIO-grid-month level up to 36 months ahead. Using only log-transformed past fatalities as input, the model jointly solves classification and regression tasks for each violence type, producing point estimates and sampling-based predictive distributions. The main empirical claim is that HydraNet achieves higher Average Precision than the VIEWS 2020 ensemble across all three violence types in a 36-month out-of-sample evaluation (state-based AP 0.304 vs 0.272, non-state 0.134 vs 0.047, one-sided 0.162 vs 0.138; Table 2), while also providing regression and uncertainty outputs that VIEWS 2020 lacks. The paper argues that learned spatiotemporal representations outperform manually engineered conflict-history features.

Significance. The contribution is potentially significant for conflict forecasting. The architecture is a reasonable and well-motivated combination of established components (U-Net, LSTM, Monte Carlo dropout, multi-task learning), and the evaluation against VIEWS 2020 is the appropriate benchmark. If the AP improvements are robust, the result that raw past fatality grids outperform a heavily feature-engineered ensemble on the field's primary metric would be an important finding. The paper provides public replication code and data access via VIEWSER, which is a strength. However, the manuscript in its current form contains an ambiguous but potentially invalidating description of the out-of-sample protocol, and the abstract overstates the results relative to Table 2. These issues must be resolved before the contribution can be fully assessed.

major comments (4)
  1. [Section 3.5] The sentence 'During out-of-sample evaluation ... the network is exposed to the full data volume of the relevant partition, including the hold-out test set months' is directly contradicted by the next sentence, which says it processes this volume 'from the first to the last month of the training set.' If the LSTM hidden and cell states are updated on the observed test-month grids, then the AP values in Table 2 are not out-of-sample and the central claim collapses. If the first clause is a typo for 'excluding', the manuscript must say so explicitly, because this is exactly the passage a reader will check for leakage. I ask the authors to rewrite this paragraph to state unambiguously that only training months are processed as observed inputs, and that the 36 test months are generated autoregressively.
  2. [Abstract and Section 1] The claim that HydraNet 'achieves state-of-the-art performance across all tasks' is not supported by Table 2. For state-based violence, HydraNet's Brier score (0.009) is worse than both VIEWS (0.005) and the no-change baseline (0.007); for one-sided violence, its Brier (0.007) is worse than the no-change baseline (0.006). Its AUC is also lower than VIEWS for state-based (0.907 vs 0.921) and one-sided (0.884 vs 0.900) violence. The accurate claim, which the paper itself makes in Section 4.2, is that HydraNet outperforms VIEWS on Average Precision, the primary metric, for all three violence types. The abstract and conclusion should be revised to match this narrower claim.
  3. [Sections 3.2 and 4.2] The headline comparison with VIEWS 2020 assumes that the published VIEWS numbers were computed on the same data version, grid definition, test months, and metric computations. The paper does not specify the UCDP GED version accessed via VIEWSER, nor does it independently reproduce the VIEWS evaluation. The AP differences for state-based (0.304 vs 0.272) and one-sided (0.162 vs 0.138) are modest, and could be affected by minor differences in preprocessing or sample composition. Additionally, if HydraNet was trained on the full Africa+Middle East volume while the VIEWS benchmark excludes the Middle East, the comparability is not self-evident. The authors should state the exact data snapshot and either recompute the VIEWS baseline on the identical sample or provide a careful account of why the published numbers are directly comparable.
  4. [Section 4.2, Table 2] All HydraNet results are from a single 'representative' model chosen as the closest to the mean of 12 independent training runs. No variance or confidence intervals are reported. Given that the AP advantage over VIEWS is small for state-based and one-sided violence, the reader cannot tell whether this difference is systematic. The authors should report the full distribution of AP (and other metrics) across the 12 instances, or at least a standard deviation or range, and ideally a paired significance test if the metric computation allows it.
minor comments (5)
  1. [Section 3.5] The no-change baseline used in Table 2 is not defined anywhere in the text; please specify what it predicts (e.g., the previous month's conflict status or fatality count for each cell).
  2. [Section 3.4] The phrase 'the network architecture prober' appears to be a typo for 'the network architecture proper'.
  3. [Section 6] The heading 'Appedix' should be spelled 'Appendix'.
  4. [Section 4.2] The statement that the cell state is frozen during the forecasting phase is not introduced in Section 3.5 and appears to conflict with the description of autoregressive prediction; please explain the exact state-update procedure during the 36-month test phase.
  5. [General] The main text refers to an online appendix for architecture details, hyperparameters, and training specifications, but that appendix is not included in the manuscript. For a stand-alone journal submission, the essential implementation details should be in the main text or in an accessible supplement, since replication depends on them.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: HydraNet's predictions are a standard supervised shift from past conflict magnitudes to future targets, and the VIEWS 2020 benchmark is external; the only flagged issue is an internally contradictory evaluation sentence that is an out-of-sample validity risk, not a circular reduction.

full rationale

The paper's derivation chain is a conventional forecasting setup: the sole inputs are the historical log-fatality magnitudes cmsb, cmns, and cmos (Eq. 1), with binary presence transforms used only as targets (Eq. 2); the six outputs are future values of these same variables. Nothing defines a target as a function of the model's own output, and no fitted parameter is relabeled as a prediction. The headline comparison to VIEWS 2020 rests on AP/AUC/Brier numbers quoted from Hegre et al. (2021a) and Hegre et al. (2020c), which are external published benchmarks; the paper does not tune HydraNet to those numbers or derive them from its own training, so the comparison is independent evidence rather than a self-referential target. Self-citations are minor and non-load-bearing: the author's dissertation is mentioned only in the author's note, and the VIEWS-platform repositories are replication infrastructure, not the justification for the performance claim. One passage should be flagged for the record: Section 3.5 says 'the network is exposed to the full data volume of the relevant partition, including the hold-out test set months,' immediately followed by 'It sequentially processes this volume from the first to the last month of the training set.' If the first clause were literal, test-month grids would update the LSTM hidden/cell states before forecasting, which would make the Table 2 APs not truly out-of-sample; however, the second clause and the surrounding protocol (hold-out 'not used for training', evaluation 'based exclusively on out-of-sample performance') indicate a wording ambiguity rather than a constructed circularity. This is an evaluation-validity concern to resolve, not a case where the model's prediction is equivalent to its inputs by definition. I therefore find no circular step meeting the quoted-evidence standard.

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

The central claims rest on the inputs used (UCDP GED via VIEWSER), the modeling premise that past conflict patterns are sufficient, the architectural premise that LSTM U-Net with MC dropout learns useful patterns from rare events, and the comparability premise that HydraNet's evaluation matches VIEWS 2020. Hyperparameters and sampling choices are free parameters not fully reported.

free parameters (5)
  • Network hyperparameters (depth, filters, kernel sizes, dropout rate, learning rate, loss weights) = Not reported in paper (deferred to online appendix)
    Tuned on the calibration partition (Section 3.5); central to reported performance but not reproducible from the text alone.
  • Patch size for spatial sampling = 32 x 32
    Training uses 32x32x276 patches (Section 3.5); affects which spatial context the model sees.
  • Number of Monte Carlo Dropout samples = 128
    Point estimates are means of 128 posterior samples (Section 3.5); a different sample count would change estimates slightly.
  • Curriculum learning schedule parameters = Not specified
    Predefined Curriculum Learning biases sampling toward conflict areas early (Section 3.5); the schedule is not detailed in the paper.
  • Number of independently trained model instances = 12
    One of 12 instances closest to the sample mean is reported (Section 4.2); this selection rule affects results.
assumptions (5)
  • domain assumption UCDP GED fatality data accurately measures conflict presence and magnitude at the priogrid-month level.
    The entire approach and evaluation rest on the validity of this data (Section 3.2).
  • domain assumption Past conflict fatalities contain sufficient predictive signal to forecast future conflict at SOTA level.
    This is the paper's central premise, argued in Section 2; if false, the SOTA claim fails.
  • domain assumption The LSTM U-Net architecture with MC dropout can learn spatiotemporal patterns from limited rare-event data.
    The architecture's design assumes that convolutional plus LSTM structures generalize from a small number of conflict events (Section 3.4).
  • standard math Monte Carlo Dropout approximates a valid Bayesian predictive posterior.
    This is a standard ML technique, but treating dropout as a Bayesian posterior is a known approximation that is not validated with distributional metrics in this paper (Sections 3.4 and 4.1).
  • domain assumption The evaluation scheme is identical to Hegre et al. (2021a), including data partitioning and metric definitions.
    Used to justify direct comparability with VIEWS 2020 (Section 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Next-Generation Conflict Forecasting: Unleashing Predictive Patterns through Spatiotemporal Learning." pith.science (2026). https://pith.science/paper/UZTZG4PM

@misc{pith2026250614817,
  author       = {Pith},
  title        = {Pith review of: Next-Generation Conflict Forecasting: Unleashing Predictive Patterns through Spatiotemporal Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZTZG4PM}},
  note         = {Machine review of arXiv:2506.14817}
}
read the original abstract

Forecasting violent conflict at high spatial and temporal resolution remains a central challenge for both researchers and policymakers. This study presents a novel neural network architecture for forecasting three distinct types of violence -- state-based, non-state, and one-sided -- at the subnational (priogrid-month) level, up to 36 months in advance. The model jointly performs classification and regression tasks, producing both probabilistic estimates and expected magnitudes of future events. It achieves state-of-the-art performance across all tasks and generates approximate predictive posterior distributions to quantify forecast uncertainty. The architecture is built on a Monte Carlo Dropout Long Short-Term Memory (LSTM) U-Net, integrating convolutional layers to capture spatial dependencies with recurrent structures to model temporal dynamics. Unlike many existing approaches, it requires no manual feature engineering and relies solely on historical conflict data. This design enables the model to autonomously learn complex spatiotemporal patterns underlying violent conflict. Beyond achieving state-of-the-art predictive performance, the model is also highly extensible: it can readily integrate additional data sources and jointly forecast auxiliary variables. These capabilities make it a promising tool for early warning systems, humanitarian response planning, and evidence-based peacebuilding initiatives.

Figures

Figures reproduced from arXiv: 2506.14817 by the authors.

Figure 1
Figure 1. Toy example of a z-stack visualized as a series of 2D grids covering Africa. Each grid [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Simplified schematic of HydraNet architecture featuring a deep convolutional en [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. A toy example of the patch sampling process. The orange squares represent 15 sampled [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: MSE, AP, AUC, and Brier scores for state-based violence over the 36 hold-out test [PITH_FULL_IMAGE:figures/full_fig_p030_4.png]
Figure 5
Figure 5. Figure 5: MSE, AP, AUC, and Brier scores for non-state violence over the 36 hold-out test [PITH_FULL_IMAGE:figures/full_fig_p031_5.png]
Figure 6
Figure 6. Figure 6: MSE, AP, AUC, and Brier scores for one-sided violence across 36 validation months in [PITH_FULL_IMAGE:figures/full_fig_p031_6.png]
Figure 7
Figure 7. Figure 7: MSE, AP, AUC, and Brier scores for state-based violence across the full test partition of [PITH_FULL_IMAGE:figures/full_fig_p032_7.png]
Figure 8
Figure 8. Figure 8: Evaluation metrics (MSE, AP, AUC, Brier) for non-state violence over 36 hold-out [PITH_FULL_IMAGE:figures/full_fig_p032_8.png]
Figure 9
Figure 9. Figure 9: MSE, AP, AUC, and Brier scores for one-sided violence over 36 months of out-of-sample [PITH_FULL_IMAGE:figures/full_fig_p033_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 61 canonical work pages

  1. [1]

    Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017). Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence , 39(12):2481--2495

    work page 2017

  2. [2]

    Bara, C. (2018). Legacies of violence: Conflict-specific capital and the postconflict diffusion of civil war. Journal of Conflict Resolution , 62(9):1991--2016

    work page 2018

  3. [3]

    and Rupert, M

    Bayer, R. and Rupert, M. C. (2004). Effects of civil wars on international trade, 1950-92. Journal of Peace Research , 41(6):699--713

    work page 2004

  4. [4]

    N., and Tucker, R

    Beck, N., Katz, J. N., and Tucker, R. (1998). Taking time seriously: Time-series-cross-section analysis with a binary dependent variable. American Journal of Political Science , 42(4):1260--1288

    work page 1998

  5. [5]

    Bengio, Y., Louradour, J., Collobert, R., and Weston, J. (2009). Curriculum learning. In Proceedings of the 26th annual international conference on machine learning , pages 41--48

    work page 2009

  6. [6]

    Bestgen, Y. (2015). Exact expected average precision of the random baseline for system evaluation. The Prague Bulletin of Mathematical Linguistics , 103(1):131--138

    work page 2015

  7. [7]

    G., Davies, A

    Brodrick, P. G., Davies, A. B., and Asner, G. P. (2019). Uncovering ecological patterns with convolutional neural networks. Trends in Ecology & Evolution , 34(8):734--745

    work page 2019

  8. [8]

    and Gleditsch, K

    Buhaug, H. and Gleditsch, K. S. (2008). Contagion or confusion? W hy conflicts cluster in space. International Studies Quarterly , 52(2):215--233

    work page 2008

Show all 63 references
  1. [9]

    Burkov, A. (2020). Machine learning engineering , volume 1. True Positive Incorporated Montreal, QC, Canada

    2020

  2. [10]

    S., and Buhaug, H

    Cederman, L.-E., Gleditsch, K. S., and Buhaug, H. (2013). Inequality, Grievances, and Civil War . Cambridge University Press

    2013

  3. [11]

    S., and Wucherpfennig, J

    Cederman, L.-E., Gleditsch, K. S., and Wucherpfennig, J. (2017). Predicting the decline of ethnic civil war: Was G urr right and for the right reasons? Journal of Peace Research , 54(2):262--274

    2017

  4. [12]

    Chadefaux, T. (2014). Early warning signals for war in the news. Journal of Peace Research , 51(1):5--18

    2014

  5. [13]

    Chadefaux, T. (2017). Conflict forecasting and its limits. Data Science , 1(1-2):7--17

    2017

  6. [14]

    Collier, P. et al. (2003). Breaking the Conflict Trap: Civil War and Development Policy . World Bank Publications

    2003

  7. [15]

    and Hoeffler, A

    Collier, P. and Hoeffler, A. (2004). Greed and grievance in civil war. Oxford Economic Papers , 56(4):563--595

    2004

  8. [16]

    and Sambanis, N

    Collier, P. and Sambanis, N. (2002). Understanding civil war: A new agenda. Journal of Conflict Resolution , 46(1):3--12

    2002

  9. [17]

    Croicu, M., Dale, J., and Landsverk, P. (2023). viewser 6.0.0. https://pypi.org/project/viewser/\#description. Accessed: 2023-06-16

    2023

  10. [18]

    Crost, B., Felter, J., et al. (2015). Is conflict contagious? E vidence from a natural experiment. Technical report, Households in Conflict Network

    2015

  11. [19]

    Fearon, J. D. and Laitin, D. D. (2003). Ethnicity, insurgency, and civil war. American Political Science Review , 97(1):75–90

    2003

  12. [20]

    and Ghahramani, Z

    Gal, Y. and Ghahramani, Z. (2016). Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning , pages 1050--1059

    2016

  13. [21]

    A., Bates, R

    Goldstone, J. A., Bates, R. H., Epstein, D. L., Gurr, T. R., Lustik, M. B., Marshall, M. G., Ulfelder, J., and Woodward, M. (2010). A global model for forecasting political instability. American Journal of Political Science , 54(1):190--208

    2010

  14. [22]

    Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning . MIT press

    2016

  15. [23]

    Hegre, H., Allansson, M., Basedau, M., Colaresi, M., Croicu, M., Fjelde, H., Hoyles, F., Hultman, L., H \"o gbladh, S., Jansen, R., et al. (2019). Views: A political violence early-warning system. Journal of Peace Research , 56(2):155--174

    2019

  16. [24]

    R., Lindqvist-McGowan, A., Randahl, D., R d, E

    Hegre, H., Bell, C., Colaresi, M., Croicu, M., Hoyles, F., Jansen, R., Leis, M. R., Lindqvist-McGowan, A., Randahl, D., R d, E. G., and Vesco, P. (2021a). Views2020: Revising and evaluating the views political violence early-warning system. Journal of Peace Research , 58(3):599--611

    2021

  17. [25]

    R., Lindqvist-McGowan, A., Randahl, D., R d, E

    Hegre, H., Colaresi, M., Croicu, M., Hoyles, F., Jansen, R., Leis, M. R., Lindqvist-McGowan, A., Randahl, D., R d, E. G., and Vesco, P. (2020a). Views: Data setup and notation. appendix a to views2020: Revising and evaluating the views political violence early-warning system

    2020

  18. [26]

    R., Lindqvist-McGowan, A., Randahl, D., R d, E

    Hegre, H., Colaresi, M., Croicu, M., Hoyles, F., Jansen, R., Leis, M. R., Lindqvist-McGowan, A., Randahl, D., R d, E. G., and Vesco, P. (2020b). Views: Data setup and notation. appendix c to views2020: Revising and evaluating the views political violence early-warning system

    2020

  19. [27]

    R., Lindqvist-McGowan, A., Randahl, D., R d, E

    Hegre, H., Colaresi, M., Croicu, M., Hoyles, F., Jansen, R., Leis, M. R., Lindqvist-McGowan, A., Randahl, D., R d, E. G., and Vesco, P. (2020c). Views: Data setup and notation. appendix d to views2020: Revising and evaluating the views political violence early-warning system

    2020

  20. [28]

    W., Nyg rd, H

    Hegre, H., Metternich, N. W., Nyg rd, H. M., and Wucherpfennig, J. (2017a). Introduction: Forecasting in peace research

    2017

  21. [29]

    M., and Landsverk, P

    Hegre, H., Nyg rd, H. M., and Landsverk, P. (2021b). Can we predict armed conflict? H ow the first 9 years of published forecasts stand up to reality. International Studies Quarterly , 65(3):660--668

    2021

  22. [30]

    M., and R der, R

    Hegre, H., Nyg rd, H. M., and R der, R. F. (2017b). Evaluating the scope and intensity of the conflict trap: A dynamic simulation approach. Journal of Peace Research , 54(2):243--261

    2017

  23. [31]

    Hegre, H., Vesco, P., and Colaresi, M. (2022). Lessons from an escalation prediction competition. International Interactions , pages 1--34

    2022

  24. [32]

    and Schmidhuber, J

    Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural Computation , 9(8):1735--1780

    1997

  25. [33]

    and Weinstein, J

    Humphreys, M. and Weinstein, J. M. (2007). Demobilization and reintegration. Journal of Conflict Resolution , 51(4):531--567

    2007

  26. [34]

    Kendall, A., Badrinarayanan, V., and Cipolla, R. (2015). Bayesian segnet: Model uncertainty in deep convolutional encoder-decoder architectures for scene understanding. arXiv preprint arXiv:1511.02680

    2015 arXiv

  27. [35]

    Kendall, A., Gal, Y., and Cipolla, R. (2018). Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 7482--7491

    2018

  28. [36]

    R., Maier-Hein, K., Eslami, S

    Kohl, S., Romera-Paredes, B., Meyer, C., De Fauw, J., Ledsam, J. R., Maier-Hein, K., Eslami, S. A., Rezende, D. J., and Ronneberger, O. (2018). A probabilistic u-net for segmentation of ambiguous images. In Advances in Neural Information Processing Systems , pages 6965--6975

    2018

  29. [37]

    Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Doll \'a r, P. (2017). Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision , pages 2980--2988

    2017

  30. [38]

    Lock, P. (1997). Armed conflicts and small arms proliferation: Refocusing the research agenda. Policy Sciences , 30(3):117--132

    1997

  31. [39]

    Lu, X., Ma, C., Ni, B., Yang, X., Reid, I., and Yang, M.-H. (2018). Deep regression tracking with shrinkage loss. In Proceedings of the European conference on computer vision (ECCV) , pages 353--369

    2018

  32. [40]

    Malone, I. (2022). Recurrent neural networks for conflict forecasting. International Interactions , pages 1--20

    2022

  33. [41]

    Mueller, H. F. and Rauh, C. (2016). Reading between the lines: Prediction of political violence using newspaper text. American Political Science Review , 2(112):358--375

    2016

  34. [42]

    D., and Linke, A

    O'Loughlin, J., Witmer, F. D., and Linke, A. M. (2010). The A fghanistan- P akistan wars, 2008-2009: Micro-geographies, conflict diffusion, and clusters of violence. Eurasian Geography and Economics , 51(4):437--471

    2010

  35. [43]

    Perry, C. (2013). Machine learning and conflict prediction: A use case. Stability: International Journal of Security & Development , 56(2(3))

    2013

  36. [44]

    Radford, B. J. (2022). High resolution conflict forecasting with spatial convolutions and long short-term memory. International Interactions , pages 1--20

    2022

  37. [45]

    and Hegre, H

    Raknerud, A. and Hegre, H. (1997). The hazard of war: Reassessing the evidence for the democratic peace. Journal of Peace Research , 34(4):385--404

    1997

  38. [46]

    Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 234--241. Springer

    2015

  39. [47]

    Schrodt, P. A. (2014). Seven deadly sins of contemporary quantitative political analysis. Journal of Peace Research , 51(2):287--300

    2014

  40. [48]

    and Weidmann, N

    Schutte, S. and Weidmann, N. B. (2011). Diffusion patterns of violence in civil wars. Political Geography , 30(3):143--152

    2011

  41. [49]

    Singer, J. D. (1973). The peace researcher and foreign policy prediction. Peace Science Society (International) , 21:1--13

    1973

  42. [50]

    Slantchev, B. L. (2012). Borrowed power: Debt finance and the resort to arms. American Political Science Review , pages 787--809

    2012

  43. [51]

    T., Rota, P., and Sebe, N

    Soviany, P., Ionescu, R. T., Rota, P., and Sebe, N. (2022). Curriculum learning: A survey. International Journal of Computer Vision , 130(6):1526--1565

    2022

  44. [52]

    Stedman, S. J. (1997). Spoiler problems in peace processes. International Security , 22(2):5--53

    1997

  45. [53]

    Su, W., Yuan, Y., and Zhu, M. (2015). A relationship between the average precision and the area under the roc curve. In Proceedings of the 2015 International Conference on the Theory of Information Retrieval , pages 349--352. ACM

    2015

  46. [54]

    F., Strand, H., and Buhaug, H

    Tollefsen, A. F., Strand, H., and Buhaug, H. (2012). PRIO -grid: A unified spatial data structure. Journal of Peace Research , 49(2):363--374

    2012

  47. [55]

    UCDP georeferenced event dataset ( GED ) global version 18.1

    UCDP (2017). UCDP georeferenced event dataset ( GED ) global version 18.1. http://www.ucdp.uu.se/downloads/. Accessed: 2018-11-13

    2017

  48. [56]

    B., Lo, A., Reisch, G., and Weidmann, N

    Vesco, P., Hegre, H., Colaresi, M., Jansen, R. B., Lo, A., Reisch, G., and Weidmann, N. B. (2022). United they stand: Findings from an escalation prediction competition. International Interactions , pages 1--37

    2022

  49. [57]

    von der Maase, S. P. (2023). A Lens to Learn Through: Conflict Studies in the Age of Computational Methods . Ph.d. dissertation, University of Copenhagen, Department of Political Science and the Center for Social Data Science (SODAS). Supervised by Lene Hansen and Jacob Gerner...

    2023

  50. [58]

    Walter, B. F. (2004). Does conflict beget conflict? E xplaining recurring civil war. Journal of Peace Research , 41(3):371--388

    2004

  51. [59]

    Wang, X., Chen, Y., and Zhu, W. (2021). A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence , 44(9):4555--4576

    2021

  52. [60]

    Ward, M. D. and Beger, A. (2017). Lessons from near real-time forecasting of irregular leadership changes. Journal of Peace Research , 54(2):141--156

    2017

  53. [61]

    D., Greenhill, B

    Ward, M. D., Greenhill, B. D., and Bakke, K. M. (2010). The perils of policy by p-value: Predicting civil conflicts. Journal of Peace Research , 47(4):363--375

    2010

  54. [62]

    Weidmann, N. B. and Ward, M. D. (2010). Predicting conflict in space and time. Journal of Conflict Resolution , 54(6):883--901

    2010

  55. [63]

    Wood, E. J. (2008). The social processes of civil war: The wartime transformation of social networks. Annual Review of Political Science , 11:539--561

    2008

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.