{"id":"9f3fc4dd-7970-4980-aa1a-11f6bfb37f2a","arxiv_id":"2504.18029","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Using SHAP and LIME on a public O-RAN dataset, the paper finds airtime, BSR, goodput, and throughput are the most influential parameters for RAN energy consumption.","lead":"This paper applies two explainable AI methods, SHAP and LIME, to a public Open RAN dataset to rank which radio parameters most drive base station energy consumption. It reports that airtime, buffer status, goodput, and throughput are the leading factors across three machine learning models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own single-instance SHAP/LIME figures contradict the claimed top-4 parameters across models, so the ranking generalization is unsupported.","rationale":"The reader's weakest_assumption correctly points to the local-to-global generalization problem: the paper infers 'top-4 key parameters' from single-instance explanations without aggregating over the test set or reporting variance. My stress-test confirms that this is the most load-bearing weakness, but I strengthen it: the paper's own figures and text are internally inconsistent with the claimed top-4 set. For example, Fig. 2a (LIME-Gradient Boosting) ranks bsr 5th and omits selected_airtime from the top-10, while Fig. 2e (LIME-XGBoost) ranks cqi_ul 4th and bsr_dl 7th. Therefore the claim is not merely under-supported by missing statistics; it is contradicted by the displayed evidence. This raises the seriousness of the flaw, but it remains fixable: the authors could aggregate SHAP/LIME values over the full test set and either confirm the top-4 claim or substantially weaken the conclusion. The existing CONDITIONAL verdict already demands such quantitative support, so I do not recommend a change to the verdict category. The paper should not be accepted as-is, but it is not beyond repair; a revised version with proper aggregation and honest reporting of variability could be acceptable. My concrete test would settle whether the concern lands: if the aggregated top-4 matches the claim across all models, the original conclusion survives; otherwise the paper must be revised to present a more nuanced, instance- and model-dependent account of feature influence.","tokens_in":10572,"tokens_out":6638,"duration_ms":60293,"concrete_test":"For each of the three models (Gradient Boosting, Random Forest, XGBoost) and both datasets (UL and DL/UL), compute the mean absolute SHAP value and the mean absolute LIME weight over the full test set (or a large random sample of at least 1000 instances) for every feature. Rank features by these aggregated scores for each model/method/dataset combination and compare the resulting top-4 lists to the claimed set (airtime, BSR, goodput, selected airtime). Additionally, verify whether the specific displayed instances in Figs. 2 and 3 indeed place these four features in the top-4; if the aggregated ranking varies across models or does not match the claim, the paper's central ranking conclusion is not established and must be revised or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV-B, the paper claims: 'Both LIME and SHAP in all the considered AI/ML models report that (i) airtime; (ii) Average Buffer Status Report (bsr); (iii) Average Goodput (gput) uplink; and (iv) selected airtime are top-4 key parameters that could influence energy consumption of the RAN for the DL/UL dataset.' However, the displayed evidence in Fig. 2 contradicts this. For LIME with Gradient Boosting (Fig. 2a), the top-4 contributions are airtime_ul (0.74), airtime_dl (0.13), thr_ul (0.08), gput_ul (0.08); bsr_ul is 5th and selected_airtime does not appear in the top-10 list. For LIME with Random Forest (Fig. 2c), the top-4 are airtime_ul, bsr_ul, thr_ul, turbodec_it; gput is 6th and selected_airtime is 5th. For LIME with XGBoost (Fig. 2e), the top-4 are airtime_ul, selected_airtime_ul, thr_ul, cqi_ul; bsr_dl is 7th. The SHAP descriptions in the text likewise name different leading features (dec_time, thr, turbodec_it) and explicitly downplay selected_airtime in some cases. Thus the claimed cross-model, cross-method consistency does not hold even for the single displayed instances. Because no aggregation over the test set or variance measure is provided, the paper's 'top-4' generalization is not supported by the evidence it presents. The same pattern appears in the UL dataset (Fig. 3), where nRBs, throughput, and turbodec_it frequently outrank the claimed parameters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies two model-agnostic explainability methods, SHAP and LIME, to energy-consumption prediction models (Gradient Boosting, Random Forest, and XGBoost) trained on a public O-RAN experimental dataset. It proposes integrating XAI into the O-RAN RIC architecture, reports train/test MSE for the three models on two datasets (DL/UL and UL-only), and, based on individual explanation plots, claims that a small set of RAN parameters—airtime, buffer status report, goodput, and selected airtime—are the top-4 key drivers of RAN energy consumption. The paper then suggests these parameters as targets for energy-efficient RAN design.","tokens_in":10922,"tokens_out":3872,"duration_ms":38532,"significance":"The topic is timely: providing interpretable explanations for energy predictions in Open RAN could support operators and future rApps/xApps. The paper uses a public real-world dataset and applies two standard explainability techniques, which is a sensible and reproducible starting point. If the headline parameter ranking were supported by rigorous aggregation over the test set and by model diagnostics, the work would be a useful empirical benchmark for energy-oriented RAN optimization. As it stands, however, the central ranking claim is not supported by the evidence in the figures, so the scientific contribution is conditional on a substantial re-analysis.","major_comments":[{"comment":"The assertion that \"Both LIME and SHAP in all the considered AI/ML models report that (i) airtime; (ii) Average Buffer Status Report (bsr); (iii) Average Goodput (gput) uplink; and (iv) selected airtime are top-4 key parameters\" is contradicted by the displayed instances. For LIME with Gradient Boosting in Fig. 2a, the top-4 features are airtime_ul, airtime_dl, thr_ul, and gput_ul, with bsr_ul fifth and selected_airtime absent from the top-10. For LIME with XGBoost in Fig. 2e, gput_ul does not appear in the top-10 and bsr_dl ranks seventh. The SHAP panels described in the text also name different leading features, including dec_time, thr, gput, and turbodec_it. Thus the claimed cross-model and cross-method consistency does not hold even on the single displayed instances, and the central ranking claim is unsupported.","section":"Section IV-B, Figs. 2 and 3"},{"comment":"The analysis generalizes from one local explanation per model-dataset pair to a global statement about \"key parameters\" without any aggregation. SHAP and LIME values are computed for a single instance; the feature ranking can vary substantially across instances, especially for tree ensembles. The paper should report aggregated importance, for example mean absolute SHAP values or LIME weights over the full test set, together with a measure of variance (e.g., standard deviation or quantiles across instances), before drawing conclusions about which RAN parameters are globally important.","section":"Section IV-B, Figs. 2 and 3"},{"comment":"The model-performance metrics indicate overfitting, and the lack of model diagnostics weakens the credibility of the subsequent explainability analysis. On the DL/UL dataset, Random Forest has train MSE 0.00897 W versus test MSE 0.06806 W, and XGBoost has 0.01672 W versus 0.07021 W; on the UL dataset Random Forest has 0.0020 W versus 0.0143 W. No hyperparameter values, cross-validation, repeated runs, or error bars are reported, so it is unclear whether the fitted models generalize. Because SHAP and LIME explain these particular fitted models, the explanations inherit any overfitting. The authors should provide tuned models and demonstrate stability (e.g., through cross-validated performance and repeated training runs).","section":"Section IV-A, Tables I and II"},{"comment":"The conclusion that \"variations in RAN parameters ... could impact the RAN energy consumption\" is too weak and generic to be a substantive scientific finding, and the paper does not test the XAI-identified parameters against any baseline or ground truth. The claimed influence of airtime, BSR, goodput, and selected airtime should be validated, for example by retraining the models without those features and measuring the drop in prediction accuracy, or by comparing the XAI rankings with domain-known physical drivers of RAN energy consumption. Without such a step, the reported feature rankings are purely internal descriptions of fitted models and are not established as meaningful RAN energy drivers.","section":"Section IV and Section V"}],"minor_comments":[{"comment":"There are several typos and inconsistent terms: \"transmission Gain (tx again)\" should likely be \"transmission gain (txgain)\"; the dataset reference [12] contains \"Evaluatuin\" instead of \"Evaluation\"; and capitalization of \"Goodput\" and \"BSR\" is inconsistent across the text and figures.","section":"Throughout"},{"comment":"The LIME and SHAP plots are difficult to read at publication size; higher-resolution figures, a consistent color legend, and a note explaining which values are shown (e.g., a single instance vs. aggregated) would greatly improve clarity.","section":"Figures 2 and 3"},{"comment":"The paper does not state the dataset size, number of features, feature normalization, or how the 80/20 train-test split was performed. Adding these details is essential for reproducibility, especially since the analysis relies on feature-level explanations.","section":"Section IV-A"},{"comment":"The text sometimes claims a feature has a \"positive impact\" based on a SHAP summary plot without clarifying whether this refers to the global mean absolute SHAP value or to a selected instance; this should be stated explicitly in the figure captions and in the text.","section":"Section IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper's core claim is the XAI-based ranking of RAN parameters, but the current version's own figures contradict that ranking, and the local-to-global leap is not justified. The errors are fixable: aggregating explanations over the test set, adding model validation, and tempering the conclusions could make this a reasonable empirical study. The novelty is modest, but the topic and dataset fit the journal's scope. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a clean application of SHAP and LIME to a public O-RAN energy dataset, and the useful takeaway is narrow: airtime, buffer status, goodput, and throughput are the parameters that show up most often as influential across a few tree models. That is a plausible, domain-sensible result, and the authors deserve credit for using a real testbed dataset and for laying out the O-RAN architecture clearly.\n\nThe main soft spot is that the headline \"top-4\" claim is not actually backed by the figures it presents. Section IV-B says both LIME and SHAP across all models report (i) airtime, (ii) bsr, (iii) gput uplink, and (iv) selected airtime as the top-4 for the DL/UL dataset. But in Fig. 2a (LIME with Gradient Boosting), the top-4 are airtime_ul, airtime_dl, thr_ul, gput_ul; bsr_ul is fifth and selected_airtime is not in the top ten. In Fig. 2e (LIME with XGBoost), selected_airtime_ul is second, but cqi_ul is fourth and bsr_dl is seventh, not bsr_ul. The SHAP descriptions in the text name dec_time, thr, and turbodec_it as leading features and explicitly say selected_airtime has smaller impact in some cases. So the claimed cross-model consistency does not hold even for the single instances shown. The generalization would need aggregated SHAP/LIME values over the test set, with variance or at least a summary table, and the paper provides none.\n\nOther soft spots are real but smaller. The MSE tables show large train/test gaps for Random Forest (0.00897 vs 0.06806 on DL/UL) and XGBoost, which suggests overfitting; no hyperparameters, cross-validation, or error bars are given. The paper also has no code release, though the dataset is public. The reference list is mostly fine; the self-citations are relevant prior work rather than padding.\n\nI would not call this a strong paper, but neither is it a bad one. The central direction—using XAI to identify RAN energy drivers—is sound, and a revised version that reports aggregate feature importance and tempers the top-4 claim would be a useful data point for the O-RAN community. I would send it to peer review, with the expectation of major revisions. A shorter conference or workshop venue is a better fit than a journal.","headline":"A straightforward XAI application to O-RAN energy data with a plausible parameter shortlist, but the paper's top-4 claim is contradicted by its own single-instance figures.","tokens_in":11499,"tokens_out":2741,"would_cite":false,"duration_ms":25402,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims XAI reveals airtime, BSR, goodput, and selected airtime as the dominant drivers of RAN energy consumption.","keywords":["Open RAN","Explainable AI","SHAP","LIME","energy consumption","RAN parameter analysis","machine learning","5G"],"falsifier":"Take the full test set, compute mean absolute SHAP and LIME values for every feature per model and dataset, and sort them; the claim stands only if airtime, buffer status report, goodput, and selected airtime remain in the top four of the aggregated ranking.","tokens_in":10343,"feed_emoji":"📡","tokens_out":4199,"duration_ms":37575,"temperature":0.7,"pith_summary":"This paper tries to establish which radio access network parameters actually drive energy consumption, by applying two explainable-AI techniques, SHAP and LIME, to machine-learning models that predict power draw from a real O-RAN testbed dataset. It claims that airtime, buffer status report, goodput, selected airtime, and number of resource blocks consistently appear as the top influencers of predicted energy consumption across models. If right, operators could use these ranked parameters to guide energy-saving actions in the O-RAN intelligent controllers, targeting the radio quantities that matter rather than treating the network as a black box.","feed_headline":"XAI ranks the RAN parameters that burn the most energy","feed_subtitle":"SHAP and LIME on real O-RAN data put airtime, buffer status, and goodput at the top of the power-consumption list.","key_machinery":"SHAP (Shapley additive explanations) and LIME (local interpretable model-agnostic explanations), applied to three tree-based energy predictors: gradient boosting, random forest, and XGBoost. SHAP attributes each prediction to features by averaging over feature-order permutations; LIME perturbs input instances locally to approximate the model and reads off contributions. The work uses these attributions, displayed per test instance, to rank which measured RAN parameters (airtime, BSR, goodput, throughput, subframe decoding time, number of resource blocks, and others) push predicted power up or down.","core_discovery":"On the paper's own terms, the central discovery is that SHAP and LIME explanations, applied to gradient boosting, random forest, and XGBoost energy predictors, converge on a small set of RAN parameters — airtime, buffer status report, goodput, selected airtime, and number of resource blocks — as the dominant drivers of RAN energy consumption for both uplink-only and combined uplink/downlink data. The analysis concludes that variations in these parameters could impact RAN energy consumption, and that the identified ranking provides a benchmark for designing energy-efficient protocols.","pith_inferences":["The displayed explanations cover one instance per model; a defensible global ranking would aggregate SHAP and LIME values over the full test set, so the paper's top-4 claim is stronger than its current evidence.","The parameters named as top drivers, such as airtime and selected airtime, are likely correlated with each other, so the individual importance values may split a shared causal effect.","A direct test of the paper's claim would use the top-ranked parameters as control inputs in an O-RAN testbed and measure whether constraining them cuts measured power.","If the ranking generalizes, XAI-based energy monitoring could move from post-hoc analysis to closed-loop control within the Near-RT RIC's feedback loop."],"forward_implications":["Energy-saving rApps and xApps in the O-RAN intelligent controllers can be steered by the ranked parameters, for example by triggering power-reduction actions when airtime or buffer status rises.","Operators can prioritize protocol changes that minimize unnecessary airtime, since it appears as the strongest positive driver of predicted power.","The convergence of SHAP and LIME rankings across three different tree models suggests the identified parameters are not an artifact of one ML algorithm.","The same XAI pipeline can be rerun on live O-RAN data to track when the energy drivers change with traffic conditions."],"supporting_citations":[{"why":"Cited as the SHAP technique and as an ML energy-prediction approach; supplies the explanation method the analysis applies.","marker":"[10]"},{"why":"Provides the LIME method whose local attributions the paper uses.","marker":"[11]"},{"why":"Supplies the real O-RAN testbed energy and performance dataset (UL and DL/UL) analyzed throughout.","marker":"[12]"},{"why":"Defines gradient boosting, one of the three models whose predictions are explained.","marker":"[13]"},{"why":"Defines random forest, another explained model.","marker":"[14]"},{"why":"Defines XGBoost, the third explained model.","marker":"[15]"}],"fun_headline_variants":["Airtime, buffer, goodput top XAI's RAN energy driver list","RAN energy waste: XAI cites airtime, buffer status, goodput","XAI reveals RAN's top energy consumers: airtime, buffer, goodput","Explainable AI points to airtime and buffer as RAN power drains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the single example explanation shown for each model fairly represents which features matter across the whole dataset, without averaging the explanation scores over all test cases.","fun_headline_variants_meta":{"raw":{"variants":["Airtime, buffer, goodput top XAI's RAN energy driver list","RAN energy waste: XAI cites airtime, buffer status, goodput","XAI reveals RAN's top energy consumers: airtime, buffer, goodput","Explainable AI points to airtime and buffer as RAN power drains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00062,"raw_usage":{"total_tokens":2820,"prompt_tokens":838,"completion_tokens":1982,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":1895}},"tokens_in":454,"tokens_out":1982,"duration_ms":16251,"temperature":1.0,"reasoning_tokens":1895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:25:12.779216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the full test set, compute mean absolute SHAP and LIME values for every feature per model and dataset, and sort them; the claim stands only if airtime, buffer status report, goodput, and selected airtime remain in the top four of the aggregated ranking.","supporting_citations":[{"cited_title":"Energy consumption prediction by using machine learning for smart building: Case study in Malaysia,","cited_arxiv_id":null,"evidence_quote":"Cited as the SHAP technique and as an ML energy-prediction approach; supplies the explanation method the analysis applies."},{"cited_title":"Stochastic Gradient Boosting,","cited_arxiv_id":null,"evidence_quote":"Defines gradient boosting, one of the three models whose predictions are explained."},{"cited_title":"XGBoost: A Scalable Tree Boosting System,","cited_arxiv_id":null,"evidence_quote":"Defines XGBoost, the third explained model."}],"review_version":1}