Untangling Co-Drift: Proactive Multi-Intent Failure Prediction and Root-Cause Disambiguation for Self-Driving Networks

Md. Kamrul Hossain; Walid Aljoby

REVIEW 5 major objections 5 minor 43 references

Untangling Co-Drift: Proactive Multi-Intent Failure Prediction and Root-Cause Disambiguation for Self-Driving Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read MILD claims that intent assurance can be made proactive and causal: one model predicts telemetry, analytics, and API failures and names the true root cause even when victim intents look worse.

desk verdict Serious attempt at a real gap, but the headline disambiguation claim isn't isolated and the false-positive budgets are violated; worth refereeing with changes. read the letter →

arxiv 2607.25989 v1 pith:YJYD4SEH submitted 2026-07-28 cs.NI cs.LGcs.RO

Md. Kamrul Hossain , Walid Aljoby This is my paper

classification cs.NIcs.LGcs.RO

keywords intent-basednetworkingintentdriftfailurepredictionroot-causedisambiguationmixture-of-expertsself-drivingnetworksco-driftnetworkassurance

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

The paper argues that a self-driving network's monitoring, analytics, and actuation functions form three causally coupled intents, so one fault in any of them propagates as a "co-drift"—simultaneous ambiguous anomalies in all three. It introduces MILD, a mixture-of-experts model whose explicitly supervised gating network jointly predicts each intent's failure risk within a fixed horizon and outputs a root-cause distribution over intents. Evaluated on a statistical benchmark, a microservices testbed, and an SDN edge-to-cloud testbed, MILD reports high detection rates, 8–100 minutes of lead time, and root-cause disambiguation accuracy of 67–90%. The claim is that this makes proactive, attribution-aware intent assurance feasible where reactive threshold-based drift detection cannot.

What carries the argument

The load-bearing mechanism is the teacher-augmented Mixture-of-Experts (MoE)—a set of intent-specialist expert networks, one per macro-intent, whose outputs are modulated by a gating network. The gate consumes a shared latent state plus the teacher model's coarse per-intent risk estimates, producing a root-cause probability distribution that both scales each expert's contribution to its risk head and doubles as the attribution output. A hybrid loss—lead-time-weighted focal loss, knowledge distillation, gate supervision against root-cause labels, and expert decorrelation—forces the gate to learn causal attribution rather than correlation, while the sparsity term pushes it toward decisive sing

What would settle it

Run MILD on production telemetry with independently annotated incidents, including incidents where the root cause shifts mid-incident, and compare its gate's top-1 attribution against the annotated root cause and against a trivial baseline that picks the intent with the largest smoothed KPI anomaly. If MILD's attribution accuracy is not clearly above that baseline, or if long-horizon alerts do not precede failures by the claimed lead time on real incidents, the central claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that co-drift ambiguity—where the symptomatic victim intent often shows the loudest KPI alarm—is learnable: a teacher-augmented Mixture-of-Experts with a gating network explicitly trained against root-cause labels can simultaneously predict which intents will fail within the next H minutes and concentrate probability mass on the true root-cause intent. In controlled, emulated, and SDN testbeds it reports detection rates of 91–99%, lead times of 8–100 minutes, and root-cause accuracy of 67–90%; the ablations show that removing gate supervision collapses disambiguation and removing teacher-KL alignment degrades detection. The paper presents MILD as th

Load-bearing premise

The evaluation rests on the assumption that the faults injected in the three testbeds—each with one static root cause, a fixed causal dependency set, and labels generated by the authors' own injector—faithfully mimic real self-driving-network co-drift; if production incidents have shifting root causes and unmodeled dependencies, the learned attribution could fail to transfer.

Editorial extensions

If this is right

Intent assurance shifts from reactive threshold alarms to fixed-horizon prediction with actionable lead time: average lead times of 100 minutes on the benchmark, 32 on microservices, and 8.6 on the SDN testbed, within false-positive budgets of 5–8 alerts/day.
Root-cause disambiguation becomes a trainable output rather than a post-hoc diagnostic step: 89.7%, 66.8%, and 89.0% attribution accuracy across the three environments, with ablations showing gate supervision is essential.
Multi-horizon deployment converts risk scores into a dynamically tightening urgency window, letting operators escalate from planning to mitigation as shorter-horizon models fire.
Shapley-value explanations give operators the specific KPIs behind each alert, linking intent-level attribution to concrete diagnostic evidence.
The three-macro-intent abstraction is extensible: adding intents only adds expert-head pairs, so the same architecture can cover other IBN environments such as 5G slice assurance.

Reading between the lines

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

Beyond the paper: the gating network's success depends on the quality and completeness of root-cause annotations in training data; replacing explicit cause labels with weak supervision from incident timelines or counterfactual augmentations is a promising path toward production telemetry where such labels are scarce.
Beyond the paper: if co-drift is a general phenomenon of coupled services, MILD's architecture should transfer to any causally coupled multi-service system, such as cloud-native microservice chains; this is testable by running the same pipeline on standard microservice benchmark traces.
Beyond the paper: the static single-root-cause-per-window and fixed dependency set are the main constraints; modeling root-cause shifts over time, for example with a state-space or attention-over-time gate, is a concrete extension the paper itself flags as open.
Beyond the paper: the drop in disambiguation accuracy on the microservices testbed (66.8%) suggests real-world noise and nonlinear interactions set a ceiling unless training data include more ecologically valid co-drift examples generated from actual failure-injection campaigns rather than scripted cascades.

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

Know this: MILD is a serious attempt at a genuine gap, but the central claim about co-drift disambiguation is not isolated in the evaluation, and the false-positive rates violate the paper's own budgets. Worth refereeing despite both.

What's new: the three-macro-intent framing for the self-driving loop, and the joint task of fixed-horizon failure prediction plus root-cause disambiguation under multi-intent co-drift. I don't see that exact task in the cited prior work. MILD itself is a teacher-augmented MoE with supervised gating; the ablations show the gate supervision and teacher KL terms matter, especially on the microservices testbed where removing gate supervision collapses DA from 66.8% to 20.7%. The multi-horizon urgency ensemble is a nice operational touch, and SHAP explanations at KPI level are sensible. They also open-source the datasets and code, which is real credit.

The soft spots are significant but mostly fixable. First, the stress-test note is right: disambiguation accuracy is reported pooled over all failure types. In the statistical benchmark, only 20% of events are co-drifts; the rest are single-intent, where root-cause attribution is nearly trivial. A risk-ranking baseline could score around 80% without ever resolving co-drift ambiguity. MILD's 89.67% does not isolate the hard case, and the testbeds report no co-drift-only breakdown. The single trace in Sec. V-G.1 is suggestive, not quantitative. That needs fixing with a per-event-type DA table.

Second, the false-positive budgets are violated everywhere: 1/day budget yields 4.97/day, 3/day budgets yield 8.17 and 8.30. The paper calls these 'low' but they don't meet the stated constraint. That is an operational viability problem, not a detail.

Third, the language is occasionally stronger than the method: the gate is a supervised classifier on injected root-cause labels, so 'causal understanding' overstates it. And the baselines are generic ML and simple heuristics; MicroRCA, CauseInfer, and LLM-agent RCA baselines are cited but not compared. Fourth, all data is from the authors' own emulation and fault injection; production telemetry validation is acknowledged as future work, but that is a real limitation when the claim is operational readiness.

Who this is for: people working on IBN assurance, self-driving networks, or ML-based failure prediction in network operations. It deserves a serious referee. I'd send it to peer review with a request for co-drift-only disambiguation accuracy, threshold retuning to meet budgets, tempered causal claims, and at least one stronger RCA baseline.

Referee Report

5 major / 5 minor

Summary. The paper proposes MILD, a teacher-augmented Mixture-of-Experts framework for proactive multi-intent failure prediction and root-cause disambiguation in self-driving networks. The authors formalize three macro-intents (Telemetry, Analytics, API), define fixed-horizon binary/TTF/root-cause labels, and train a gated MoE with a hybrid loss combining lead-time-weighted focal loss, knowledge distillation, gate supervision, and expert decorrelation. The framework is evaluated on a synthetic statistical benchmark and two emulated testbeds (microservices and SDN edge-to-cloud), with reported high detection rates, lead times, and disambiguation accuracy, plus ablations and hyperparameter sensitivity studies. The paper claims to be the first to jointly address fixed-horizon intent-failure prediction and root-cause disambiguation in multi-intent co-drift scenarios.

Significance. If the empirical claims are established, MILD would be a meaningful advance: it recasts IBN assurance from reactive drift detection to fixed-horizon failure prediction and explicitly targets the cause-victim ambiguity in co-drift. The problem formalization is clear, and the experimental breadth is a strength, as are the blocked cross-validation, the ablation study, and the promised release of datasets and code. However, the central quantitative support for the co-drift disambiguation claim is currently missing: the reported DA is pooled over event types in which 80% of statistical-benchmark events are single-intent and therefore trivial for a simplex gate. In addition, the reported false-positive rates exceed the stated operational budgets, and one multi-horizon lead-time number contradicts the fixed-horizon label definition. These issues are fixable but load-bearing for the paper's main claims.

major comments (5)

[Sec. IV-A, Tables V-VII] The central co-drift claim is not isolated. In the statistical benchmark, only 20% of events are multi-intent co-drifts, while 60% are non-linear single-intent failures and 20% are simple independent drifts. For single-intent failures, y_cause is active for exactly the one failing intent (Eq. (5)) and g_t is a simplex, so a gate that selects an alerted intent is trivially correct on 80% of events before considering co-drift. The reported DA values (89.67%, 66.80%, 88.97%) and the baseline comparisons in Tables VI, VIII, and X pool over these event types. Please report DA separately for co-drift events with multiple active y_bin intents, and ideally for co-drift cases where a victim intent shows stronger KPI anomalies than the root-cause intent. The illustrative trace in Sec. V-G.1 does not replace a quantitative breakdown.
[Sec. III-E, Eqs. (18)-(19); Sec. V-A, Tables V/VII/IX] The alerting policy is claimed to be constrained by a false-positive budget Gamma_max, which is set to 1/day for the statistical benchmark and 3/day for both emulated testbeds. The reported FP rates are 4.97/day, 8.17/day, and 8.30/day, respectively. Unless 'FP Rate per Day' counts something different from the constraint in Eq. (19) (e.g., raw pre-smoothing crossings or per-intent alarms), the evaluation violates its own operational constraint by factors of 5.0, 2.7, and 2.8. Please either retune thresholds and report values satisfying the budget, or explicitly redefine the FP metric and explain why Eq. (19) does not apply.
[Sec. V-G.3 vs. Eq. (3)] The multi-horizon experiment reports first alerts at 128, 97, and 42 minutes before failure for models trained with H=120, 60, and 20 minutes. Under Eq. (3), y_bin=1 only for t in [t_fail-H, t_fail), so the H=120 model is trained on negative labels at t_fail-128 and cannot legitimately issue a pre-failure alert at that point. The reported 128-minute lead time implies either a different labeling rule, an error in alerting logic, or that a false positive outside the prediction window is counted as a true lead time. Define lead time unambiguously and reconcile the reported values with the fixed-horizon labels.
[Sec. V-B, Eq. (13), Tables VI/VIII/X] MILD is trained with explicit root-cause labels y_cause through L_gate, while the baseline neural models (MLP, LSTM) are described as having only per-intent sigmoid heads and no cause-supervision term; their 'disambiguation' appears to be derived from risk-score ranking. The comparison therefore does not isolate the architecture: it compares a cause-supervised model against attribution baselines that do not receive the same supervision. To support the claim that the teacher-augmented MoE improves root-cause disambiguation, include a baseline that receives the same y_cause labels through an auxiliary softmax head (or otherwise equalizes supervision) and report DA on co-drift events for all methods.
[Sec. V-H, Tables XII-XIV] The ablation narrative is stronger than the tables. The text claims the full model 'is the only configuration that avoids catastrophic failure on any single metric,' but several ablated variants outperform MILD on individual metrics: on microservices, w/o Teacher Aug. in Gate gives DA 69.34 vs. 66.80 and w/o Distillation gives FDR 91.57 vs. 91.32; on SDN, w/o Gate Cause Sup. gives FDR 95.64 vs. 92.06. Without a defined multi-metric selection criterion or a statistical test across folds, 'no single worst value' does not establish overall superiority. Report the intended criterion or pairwise significance comparisons.

minor comments (5)

[Fig. 8] Specify whether the plotted gating probability is raw g_t or a smoothed version, and state how representative the single trace is; a quantitative summary over co-drift test events is needed.
[Sec. III-C / V-A] The teacher model f^(T) is not fully specified (architecture, training labels, input features). Since teacher augmentation is a key component and its ablation has dataset-dependent effects, provide the teacher configuration in the experimental setup.
[Tables VI, VIII, X] Baseline results are reported without standard deviations or confidence intervals, despite notable fold-level variance in MILD's own results (e.g., Table V). Add dispersion measures for baselines to support pairwise comparisons.
[Sec. VI / VII] The limitations section honestly notes reliance on post-mortem expert labels, the static-root-cause assumption, and the absence of production-telemetry validation. These caveats should be reflected in the abstract/conclusion framing so that 'practical enabler of closed-loop assurance' is appropriately qualified.
[Tables XII-XIV] Several numeric cells are hard to parse due to missing separators (e.g., '100.098.274.13' in Table XII). Please reformat to make each column unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MILD's predictions are evaluated on held-out test blocks, and its architecture and losses are not defined in terms of the reported metrics; self-citations are not load-bearing.

full rationale

No circular step can be exhibited from the paper's own equations or citations. The supervised formulation (Sec. III-B to III-D) derives labels y_bin, y_ttf, and y_cause from historical event logs (Eqs. 3-5), trains a teacher-augmented MoE against those labels, and evaluates on blocked held-out folds (Sec. V-A). The reported detection rates, lead times, and disambiguation accuracy are empirical generalization metrics, not quantities obtained from fitted parameters by construction. In particular, disambiguation accuracy is computed as argmax g_t == i* (Eq. 20) on held-out events; the gate is trained with KL(y_cause || g_t), but supervised training on a target is not a circular derivation when the test data is separate, and here it is. The co-drift labels come from the authors' own fault-injection design (Sec. IV-A), which raises ecological-validity concerns; Sec. VI explicitly acknowledges the static-root-cause and expert-annotation limitations. This is a data-realism limitation, not a definitional reduction. Self-citations to [35], [20], and [25] appear in Sec. II and the contributions section, but no load-bearing result is imported from those papers: [35] is prior work that this manuscript extends, yet the architecture, loss functions, datasets, and evaluations are fully described and re-run here. No uniqueness theorem or ansatz is imported via citation. The strongest non-circular concern is an evidentiary gap: Sec. V-C reports 89.67% disambiguation accuracy 'of both co-drift and single-intent failures,' while Sec. IV-A states only 20% of benchmark events are multi-intent co-drifts. Co-drift-only accuracy is not reported, so the central co-drift disambiguation claim rests largely on one illustrative trace (Sec. V-G.1). That is missing quantitative support, not circularity. The multi-horizon TTF bound (TTF in (H_short, H_long]) is a definitional consequence of horizon labels, but the paper presents the individual first-alert times as empirical observations, not as fitted predictions. Alert thresholds and EWMA spans are tuned on validation splits (Eqs. 18-19), which is standard hyperparameter selection rather than fitting the reported test metrics. Overall, the derivation chain is self-contained against the paper's stated benchmarks; the weaknesses are about dataset realism and incomplete subgroup analysis, not circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. The 'three macro-intents' are a modeling abstraction over existing network functions, and 'co-drift' names a described propagation pattern rather than postulating a new mechanism. The main dependencies are domain assumptions about the structure of self-driving networks and the sufficiency of KPI data, plus a set of hyperparameters tuned on validation.

free parameters (7)

prediction horizon H = 120 min (statistical), 50 min (microservices), 15 min (SDN)
Sets the fixed prediction window; chosen per dataset via validation sensitivity; directly determines the label construction (Eq. 3) and the lead-time ceiling.
distillation mixing coefficient α = 0.9 / 0.6 / 0.7
Balances focal loss vs distillation in per-head loss (Eq. 11); tuned on held-out validation (Sec. V-A).
gate cause-supervision weight w_c = 0.7 / 0.6 / 0.6
Weights KL(y_cause || g_t) in the gate loss (Eq. 13); critical for disambiguation; chosen per dataset.
gate teacher-alignment weight w_T = 0.7 / 0.3 / 0.3
Weights KL(d^(T)_t || g_t) teacher-alignment term in gate loss.
gate sparsity coefficient λ_s = 0.005 / 0.005 / 0.002
Sparsity regularizer on gate output, encouraging confident single-intent attribution.
alert thresholds τ_i and EWMA spans W_i = tuned per intent per dataset via grid search
Grid-searched under the FP budget (Eqs. 18-19); directly determines alerts, lead time, and FP rate; exact values not reported.
false-positive budget Γ_max = 1/day (statistical), 3/day (microservices and SDN)
Operational constraint stated in Eq. 19; the reported FP rates exceed these budgets, indicating the tuning did not achieve the stated objective.

assumptions (4)

domain assumption The self-driving network can be decomposed into exactly three macro-intents (telemetry, analytics, API) whose health captures the assurance loop.
Sec. III-A introduces this decomposition without empirical evidence that it is exhaustive or representative beyond the examples; the whole framework builds on it.
domain assumption The causal dependency structure (e.g., I_anl → I_tel, I_anl → I_api) is fixed and produces co-drift with a single static root cause per failure window.
Sec. III-B and the testbed injection design fix the cascade structure; the paper acknowledges in Sec. VI that dynamic cascading root causes are out of scope.
domain assumption KPI-level observations are sufficient to identify the root-cause intent; i.e., the mapping from KPI traces to y_cause is learnable from labeled emulated data.
The model relies on KPI features and supervised root-cause labels (Sec. III-C); the evaluation assumes no unobserved confounders beyond the emulated ones.
standard math Blocked cross-validation (10-fold or 3-fold) provides a valid estimate of generalization; the KPI process is stationary across blocks except injected faults.
Sec. V-A uses blocked CV to prevent temporal leakage; standard assumption but not validated for the emulated testbeds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Untangling Co-Drift: Proactive Multi-Intent Failure Prediction and Root-Cause Disambiguation for Self-Driving Networks." pith.science (2026). https://pith.science/paper/YJYD4SEH

@misc{pith2026260725989,
  author       = {Pith},
  title        = {Pith review of: Untangling Co-Drift: Proactive Multi-Intent Failure Prediction and Root-Cause Disambiguation for Self-Driving Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJYD4SEH}},
  note         = {Machine review of arXiv:2607.25989}
}

read the original abstract

The vision of self-driving networks that monitor, reason, and act upon themselves with minimal human intervention relies on tightly coupled monitoring, analytics, and actuation functions. In this work, we treat these functions as three operational macro-intents: continuous telemetry, real-time analytics, and programmatic actuation, and formalize the health of each function as an intent that the network must continuously satisfy. A critical, yet underexplored, challenge stems from the causal coupling among these intents, where a singular fault within one macro-intent propagates as a co-drift and subsequently triggers cascading, symptomatic anomalies across the remaining intents. This ambiguity makes it exceedingly difficult for existing, reactive approaches to distinguish the true root-cause intent from symptomatic victim intents, and their reliance on threshold-crossing detection leaves insufficient time for proactive remediation. We introduce MILD, a novel framework that reformulates intent assurance from reactive drift detection to proactive failure prediction. Grounded in our three-macro-intent formulation of the self-driving control loop, MILD employs a teacher-augmented Mixture-of-Experts architecture with a hybrid objective that jointly optimizes intent failure prediction and root-cause attribution. MILD enables KPI-level diagnostics via SHAP explainability and dynamic intent failure urgency estimation via multi-horizon modeling. Our extensive evaluation of MILD across three environments of increasing realism, from a controlled statistical benchmark, to a microservices application, to an SDN-based edge-to-cloud testbed, demonstrates that MILD achieves high failure detection rates, strong remediation lead times, and accurate intent-level root-cause disambiguation. This positions MILD as a practical enabler of closed-loop assurance in next-generation autonomous networks.

Figures

Figures reproduced from arXiv: 2607.25989 by the authors.

**Figure 1.** Overview of the proposed framework. KPI streams from the three macro-intents of a self-driving network (Telemetry, Analytics, and API) are [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Closed-loop control architecture for a self-driving network. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Example of a canonical co-drift cascade. A CPU bottleneck in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: MILD: teacher-augmented MoE with risk prediction, root-cause disambiguation, and per-alert intelligence. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Multi-horizon MILD deployment. Multiple models trained at different [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Containerized microservices emulation testbed for KPI generation and [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: SDN-based edge-to-cloud emulation testbed. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Example of root-cause disambiguation under a co-drift event, showing [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: SHAP explanation for an analytics intent failure alert on the SDN [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 4 linked inside Pith

[1]

Why (and how) networks should run themselves,

N. Feamster and J. Rexford, “Why (and how) networks should run themselves,”arXiv preprint arXiv:1710.11583, 2017

arXiv 2017
[2]

Workshop on self-driving networks — workshop report,

Feamster and Rexford, “Workshop on self-driving networks — workshop report,” https://www.cs.princeton.edu/ ∼jrex/papers/ self-driving-networks18.pdf, 2018, [Accessed 07-06-2026]

2018
[3]

Towards nwdaf-enabled analytics and closed-loop automation in 5g networks,

F. S. Ardestani, N. Saha, N. Limam, and R. Boutaba, “Towards nwdaf-enabled analytics and closed-loop automation in 5g networks,”
[4]

Toward integrated sensing and communications for 6g: Key enabling technologies, standardization, and challenges,

A. Kaushik, R. Singh, S. Dayarathna, R. Senanayake, M. Di Renzo, M. Dajer, H. Ji, Y . Kim, V . Sciancalepore, A. Zappone, and W. Shin, “Toward integrated sensing and communications for 6g: Key enabling technologies, standardization, and challenges,”IEEE Communications Standards Magazine, vol. 8, no. 2, pp. 52–59, 2024

2024
[5]

Intent-driven 6g end-to-end network orchestration,

Y . Ouyang, C. Li, J. Zhang, X. Zhao, and C. Yang, “Intent-driven 6g end-to-end network orchestration,” inProc. IEEE INFOCOM Workshop, 2024, pp. 1–2

2024
[6]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2023

2023
[7]

Intent-Based Networking - Concepts and Definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent-Based Networking - Concepts and Definitions,” RFC 9315, Oct. 2022. [Online]. Available: https://www.rfc-editor.org/info/rfc9315

2022
[8]

Orchestran: Orchestrat- ing network intelligence in the open ran,

S. D’Oro, L. Bonati, M. Polese, and T. Melodia, “Orchestran: Orchestrat- ing network intelligence in the open ran,”IEEE Transactions on Mobile Computing, vol. 23, no. 7, pp. 7952–7968, 2023

2023

Show all 43 references

[9]

A comprehensive survey on network resource management in sdn enabled data centre network,

A. Sharma, S. Tokekar, and S. Varma, “A comprehensive survey on network resource management in sdn enabled data centre network,”6G Enabled Fog Computing in IoT: Applications and Opportunities, pp. 333–353, 2023

2023
[10]

Intent assurance using llms guided by intent drift,

K. Dzeparoska, A. Tizghadam, and A. Leon-Garcia, “Intent assurance using llms guided by intent drift,” inProc. IEEE/IFIP NOMS, 2024, pp. 1–7

2024
[11]

Assurance and conflict detection in intent-based networking: A compre- hensive survey and insights on standards and open-source tools,

M. Gharbaoui, F. Sciarrone, M. Fontana, P. Castoldi, and B. Martini, “Assurance and conflict detection in intent-based networking: A compre- hensive survey and insights on standards and open-source tools,”IEEE Transactions on Network and Service Management, vol. 23, pp. 1891– ...

1912
[12]

Kpi assurance and llms for intent- based management,

K. Dzeparoska and A. Leon-Garcia, “Kpi assurance and llms for intent- based management,” inProc. IEEE/IFIP NOMS, 2025, pp. 1–9

2025
[13]

Network assurance in intent-based net- working data centers with machine learning techniques,

X. Zheng and A. Leivadeas, “Network assurance in intent-based net- working data centers with machine learning techniques,” inProc. IEEE CNSM, 2021, pp. 14–20

2021
[14]

Intent-based multi- agent reinforcement learning for service assurance in cellular networks,

S. K. Perepu, J. P. Martins, R. S. S, and K. Dey, “Intent-based multi- agent reinforcement learning for service assurance in cellular networks,” inProc. IEEE GLOBECOM, 2022, pp. 2879–2884

2022
[15]

Performance analysis of learning-based intent drift detection algorithms in next generation networks,

C. Muonagor, M. Bensalem, and A. Jukan, “Performance analysis of learning-based intent drift detection algorithms in next generation networks,” inProc. IEEE LATINCOM, 2024, pp. 1–6

2024
[16]

Anomaly detection and failure root cause analysis in (micro) service-based cloud applications: A survey,

J. Soldani and A. Brogi, “Anomaly detection and failure root cause analysis in (micro) service-based cloud applications: A survey,”ACM Comput. Surv., vol. 55, no. 3, Feb. 2022. [Online]. Available: https://doi.org/10.1145/3501297

2022 doi
[17]

Root cause analysis for microservice systems via hierarchical reinforcement learning from human feedback,

L. Wang, C. Zhang, R. Ding, Y . Xu, Q. Chen, W. Zou, Q. Chen, M. Zhang, X. Gao, H. Fan, S. Rajmohan, Q. Lin, and D. Zhang, “Root cause analysis for microservice systems via hierarchical reinforcement learning from human feedback,” inProc. ACM SIGKDD, ser. KDD ’23. New York, NY...

2023
[18]

Towards understanding the mixture-of-experts layer in deep learning,

Z. Chen, Y . Deng, Y . Wu, Q. Gu, and Y . Li, “Towards understanding the mixture-of-experts layer in deep learning,”Advances in neural information processing systems, vol. 35, pp. 23 049–23 062, 2022

2022
[19]

A unified approach to interpreting model predictions,

S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” inProc. NIPS, ser. NIPS’17. Red Hook, NY , USA: Curran Associates Inc., 2017, p. 4768–4777

2017
[20]

Netintent: Leveraging large language models for end-to-end intent-based sdn automation,

M. K. Hossain and W. Aljoby, “Netintent: Leveraging large language models for end-to-end intent-based sdn automation,”IEEE Open Journal of the Communications Society, vol. 6, pp. 10 512–10 541, 2025

2025
[21]

Opendaylight: A linux foundation collaborative project,

O. Project, “Opendaylight: A linux foundation collaborative project,” 2015, accessed: 2025-01-15. [Online]. Available: https://www.opendaylight.org/

2015
[22]

Interactivity anomaly detection in remote work scenarios using lstm,

J. Arellano-Uson, E. Maga ˜na, D. Morat ´o, and M. Izal, “Interactivity anomaly detection in remote work scenarios using lstm,”IEEE Access, vol. 12, pp. 34 402–34 416, 2024

2024
[23]

Towards intent assurance: A traffic prediction model for software-defined networks,

M. Gharbaoui, B. Martini, D. Berardi, and P. Castoldi, “Towards intent assurance: A traffic prediction model for software-defined networks,” in Proc. IEEE ICIN, 2025, pp. 135–139

2025
[24]

Telecom foundation models: Applications, challenges, and future trends,

T. Zanouda, M. Masoudi, F. G. Gebre, and M. Dohler, “Telecom foundation models: Applications, challenges, and future trends,” Aug. 2024. [Online]. Available: http://dx.doi.org/10.36227/techrxiv. 172296173.35282215/v1

2024
[25]

Lead-drift: Real-time and explainable intent drift detection by learning a data-driven risk score,

M. K. Hossain and W. Aljoby, “Lead-drift: Real-time and explainable intent drift detection by learning a data-driven risk score,” 2026. [Online]. Available: https://arxiv.org/abs/2602.13672

2026
[26]

Refined link-level intent drift forecasting through advanced link performance prediction and path similarity ap- proaches,

H. Liu, H. Li, and Y . Ai, “Refined link-level intent drift forecasting through advanced link performance prediction and path similarity ap- proaches,”Computer Networks, vol. 275, p. 111857, 2026

2026
[27]

Autonomous conflict handling in intent-based management,

E. Dehghan Biyar, M. D’Angelo, J. C. Cisneros, A. Nascimento, M. Orlic, A. Likhyani, A. Zahemszky, A. C. Baktir, D. A. Temesgene, and D. Roeland, “Autonomous conflict handling in intent-based management,”Computer Networks, vol. 271, p. 111561, 2025. [Online]. Available: https:...

2025
[28]

Detecting application transitions and identifying application types for intent- based network assurance: A machine learning perspective,

J. Violos, F. V outsas, C. Diou, and A. Leivadeas, “Detecting application transitions and identifying application types for intent- based network assurance: A machine learning perspective,”Computer Networks, vol. 274, p. 111872, 2026. [Online]. Available: https: //www.scienced...

2026
[29]

Microrca: Root cause localization of performance issues in microservices,

L. Wu, J. Tordsson, E. Elmroth, and O. Kao, “Microrca: Root cause localization of performance issues in microservices,” inProc. IEEE/IFIP NOMS, 2020, pp. 1–9

2020
[30]

Causeinfer: Automated end-to-end perfor- mance diagnosis with hierarchical causality graph in cloud environment,

P. Chen, Y . Qi, and D. Hou, “Causeinfer: Automated end-to-end perfor- mance diagnosis with hierarchical causality graph in cloud environment,” IEEE Transactions on Services Computing, vol. 12, no. 2, pp. 214–230, 2019

2019
[31]

Microrca-agent: Microservice root cause analysis method based on large language model agents,

P. Tang, S. Tang, H. Pu, Z. Miao, and Z. Wang, “Microrca-agent: Microservice root cause analysis method based on large language model agents,” 2025. [Online]. Available: https://arxiv.org/abs/2509.15635

2025
[32]

Tamo:fine-grained root cause analysis via tool-assisted llm agent with multi-modality observation data in cloud-native systems,

X. Zhang, Q. Wang, M. Li, Y . Yuan, M. Xiao, F. Zhuang, and D. Yu, “Tamo:fine-grained root cause analysis via tool-assisted llm agent with multi-modality observation data in cloud-native systems,”IEEE Trans- actions on Services Computing, vol. 18, no. 6, pp. 4221–4233, 2025

2025
[33]

Towards zero downtime: Using machine learning to predict network failure in 5g and beyond,

E. Basikolo and T. Basikolo, “Towards zero downtime: Using machine learning to predict network failure in 5g and beyond,”ITU Journal on Future and Evolving Technologies, vol. 4, no. 3, pp. 434–446, 2023

2023
[34]

A predictive maintenance methodology: Predicting the time-to-failure of machines in industry 4.0,

M. Z ¨ufle, J. Agne, J. Grohmann, I. D ¨ortoluk, and S. Kounev, “A predictive maintenance methodology: Predicting the time-to-failure of machines in industry 4.0,” inProc. IEEE INDIN, 2021, pp. 1–8

2021
[35]

Mild: Multi-intent learning and disambiguation for proactive failure prediction in intent-based networking,

M. K. Hossain and W. Aljoby, “Mild: Multi-intent learning and disambiguation for proactive failure prediction in intent-based networking,” 2026. [Online]. Available: https://arxiv.org/abs/2602.14283

2026
[36]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProc. ICCV, 2017, pp. 2980–2988

2017
[37]

Distilling the knowledge in a neural network,

G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”arXiv preprint arXiv:1503.02531, 2015

2015 arXiv
[38]

Re- ducing overfitting in deep networks by decorrelating representations,

M. Cogswell, F. Ahmed, R. Girshick, C. L. Zitnick, and D. Batra, “Re- ducing overfitting in deep networks by decorrelating representations,” in 4th International Conference on Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2016

2016
[39]

R. J. Hyndman and G. Athanasopoulos,Forecasting: Principles and practice, 3rd ed. OTexts, 2021

2021
[40]

Deep learning for multi-label learning: A comprehensive survey,

A. N. Tarekegn, M. Ullah, and F. A. Cheikh, “Deep learning for multi-label learning: A comprehensive survey,” 2024. [Online]. Available: https://arxiv.org/abs/2401.16549

2024 arXiv
[41]

A deep neural network-based multi- label classifier for sla violation prediction in a latency sensitive nfv application,

N. Jalodia, M. Taneja, and A. Davy, “A deep neural network-based multi- label classifier for sla violation prediction in a latency sensitive nfv application,”IEEE Open Journal of the Communications Society, vol. 2, pp. 2469–2493, 2021

2021
[42]

Mild: Multi-intent learning and disambiguation for proactive failure prediction in intent-based networking

M. K. Hossain and W. Aljoby, “Mild: Multi-intent learning and disambiguation for proactive failure prediction in intent-based networking.” [Online]. Available: https://github.com/Muhammadkamrul/ Extended MILD
[2025]

Available: https://arxiv.org/abs/2505.06789

[Online]. Available: https://arxiv.org/abs/2505.06789

arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Why (and how) networks should run themselves,

N. Feamster and J. Rexford, “Why (and how) networks should run themselves,”arXiv preprint arXiv:1710.11583, 2017

arXiv 2017

[2] [2]

Workshop on self-driving networks — workshop report,

Feamster and Rexford, “Workshop on self-driving networks — workshop report,” https://www.cs.princeton.edu/ ∼jrex/papers/ self-driving-networks18.pdf, 2018, [Accessed 07-06-2026]

2018

[3] [3]

Towards nwdaf-enabled analytics and closed-loop automation in 5g networks,

F. S. Ardestani, N. Saha, N. Limam, and R. Boutaba, “Towards nwdaf-enabled analytics and closed-loop automation in 5g networks,”

[4] [4]

Toward integrated sensing and communications for 6g: Key enabling technologies, standardization, and challenges,

A. Kaushik, R. Singh, S. Dayarathna, R. Senanayake, M. Di Renzo, M. Dajer, H. Ji, Y . Kim, V . Sciancalepore, A. Zappone, and W. Shin, “Toward integrated sensing and communications for 6g: Key enabling technologies, standardization, and challenges,”IEEE Communications Standards Magazine, vol. 8, no. 2, pp. 52–59, 2024

2024

[5] [5]

Intent-driven 6g end-to-end network orchestration,

Y . Ouyang, C. Li, J. Zhang, X. Zhao, and C. Yang, “Intent-driven 6g end-to-end network orchestration,” inProc. IEEE INFOCOM Workshop, 2024, pp. 1–2

2024

[6] [6]

A survey on intent-based networking,

A. Leivadeas and M. Falkner, “A survey on intent-based networking,” IEEE Communications Surveys & Tutorials, vol. 25, no. 1, pp. 625–655, 2023

2023

[7] [7]

Intent-Based Networking - Concepts and Definitions,

A. Clemm, L. Ciavaglia, L. Z. Granville, and J. Tantsura, “Intent-Based Networking - Concepts and Definitions,” RFC 9315, Oct. 2022. [Online]. Available: https://www.rfc-editor.org/info/rfc9315

2022

[8] [8]

Orchestran: Orchestrat- ing network intelligence in the open ran,

S. D’Oro, L. Bonati, M. Polese, and T. Melodia, “Orchestran: Orchestrat- ing network intelligence in the open ran,”IEEE Transactions on Mobile Computing, vol. 23, no. 7, pp. 7952–7968, 2023

2023