{"id":"3e0687a3-1f4c-4ef1-9b13-f9c42199e20c","arxiv_id":"1909.00525","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ActSense selects which homes and appliances to instrument so that energy use for each appliance can be inferred in other homes, using active tensor completion with seasonal forecasts.","lead":"This paper introduces ActSense, an active learning method that decides which homes and appliances to equip with energy sensors, so that appliance-level electricity use can be estimated in homes that have no sensors at all. A smart generalist might read it because it could reduce the hardware cost of energy breakdown programs that help households cut electricity use by roughly 15 percent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The theoretical guarantee is not established: Lemma 3.1 assumes global q-linear ALS convergence from a proper initialization that Algorithm 2 neither specifies nor verifies, and the Section 3.5 proof compares one-step upper bounds for one exchanged pair rather than accumulated error over arbitrary…","rationale":"I read the paper as making two contributions: an active sensing heuristic with a substantial empirical study on Dataport, and a theoretical proof that its selection rule is optimal in reducing estimation error and thus requires fewer sensors. The empirical part is plausible and reproducible (code is linked, multiple years, several baselines). My concern is with the theoretical part. The central theorem is never stated formally; the proof sketch in Section 3.5 relies on Lemma 3.1, whose assumptions are unverified and in tension with the algorithm: zero initialization, non-convex CP, and a regularized loss whose minimizer is not the ground truth. The cited ALS convergence result is local; no basin-of-attraction argument is given. This is not a matter of disagreeing with the community's consensus about tensor factorization being hard; it is that the proof's own required conditions are not shown. I also note that even under the lemma, Eq (9) establishes at most a comparison of one-step upper bounds for a single exchanged pair. The advertised accumulated-error/sensor-savings claim would need an extra induction or summation argument that is absent. Thus the reader's CONDITIONAL verdict is appropriate: the empirical claims can stand, but the theoretical optimality claim should be revised or clearly labeled as conditional on unverified convergence behavior. The reader identified q-linear ALS convergence as the weakest assumption; I agree that is the deepest gap. I would not lean heavily on the componentwise-dominance objection because the one-step upper-bound comparison can be recovered from the selection inequality using f(x)=x-x/sqrt(1+x^2)<=x; however this repair does not address Lemma 3.1. The proposed synthetic test would settle whether the Lemma's premise holds in the actual operating regime.","tokens_in":19950,"tokens_out":14257,"duration_ms":400585,"concrete_test":"Run a synthetic CP tensor-completion experiment with known positive factors h*, a*, s*, rank r=4, Gaussian noise, and the exact Algorithm 2 initialization (all factors zero). Over 1000 random noise draws, record (i) the ratio ||h^{n+1}-h*|| / ||h^n-h*|| across ALS sweeps and (ii) the empirical coverage of the Lemma 3.1 ellipsoid ||h^t - h*||_{A^t} <= bound at delta=0.05. If the ratio is not uniformly below 1 or the coverage is far below 0.95, Lemma 3.1's premise fails. Separately, attempt to derive a bound on the accumulated error sum_t MeanRMSE(t) from Eq (9); if the derivation requires assumptions about all pairs and all future months not present in the paper, the headline claim needs to be weakened to the one-step statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 3.1 is the foundation of the entire optimality proof, and its premise is not established. The lemma requires 'proper initialization' of the coordinate descent, a positive-definite Hessian at the optimizer h*, a*, s*, and q-linear convergence of ALS with rates q1, q2, q3 in (0,1). Algorithm 2 initializes all factors to zero; the CP objective (1) is non-convex; and the cited result [40] is a local convergence theorem, not a global q-linear guarantee for the regularized noisy tensor-completion problem. Moreover, the regularized empirical objective's minimizer is not generally the ground-truth h*, a*, s* that the lemma's inequalities (10)-(12) compare against, so it is unclear what the iterates converge to. Without Lemma 3.1, the alpha bounds used in the selection rule (5) and in Eq (8) have no basis, and Eq (9) cannot be derived. In addition, even granting Lemma 3.1, the proof of Eq (9) only compares one-step upper bounds for two pairs after a single swapped selection; it never bounds accumulated error over time and does not treat arbitrary deployment strategies, so the abstract's 'accumulated estimation error' and 'less sensor deployment' conclusions are not consequences of the proof as written. I do not treat the displayed 'M>=G and N>=H' as the main defect: that claim is not implied by the selection rule, but the one-step upper-bound comparison can be repaired without it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ActSense, an active learning algorithm for sensor deployment in residential energy breakdown. Energy readings are modeled as a non-negative CP-decomposed three-way tensor over homes, appliances, and months; at each month the algorithm installs sensors on the home-appliance pairs with the highest estimated prediction uncertainty, integrating historical, current, and future month uncertainties through a time-decay kernel. The central claims are (i) a theoretical guarantee that ActSense reduces accumulated estimation error faster than any other installation strategy, and (ii) an empirical demonstration on the Dataport/Austin dataset that ActSense achieves higher reconstruction accuracy than random, query-by-committee, and variational Bayes baselines under a fixed sensor budget. The paper includes proofs in a supplementary document and releases code on GitHub.","tokens_in":20284,"tokens_out":3883,"duration_ms":36599,"significance":"If the theoretical guarantee were valid, this would be a meaningful advance for active tensor completion and for practical energy-disaggregation deployment, where sensor budgets are a first-order constraint. The empirical setup is reproducible and uses a large public dataset, and the reported improvements over three baselines across four years are concrete and falsifiable. However, the theoretical argument has load-bearing gaps: Lemma 3.1 rests on unverified global convergence assumptions, and the proof of Eq. (9) uses an unjustified componentwise dominance step and proves only a one-step bound on two pairs, not the accumulated-error optimality stated in the abstract. The empirical contribution is promising, but the advertised theoretical result is not established as written.","major_comments":[{"comment":"The proof of Eq. (9) claims that the ActSense selection condition implies M ≥ G and N ≥ H, where M,N,G,H are defined immediately below the displayed bounds. The selection rule in Eq. (5) gives only the weighted inequality α1 M + α2 N ≥ α1 G + α2 H, which does not imply componentwise dominance. Without M ≥ G and N ≥ H, the comparison between the two upper bounds, specifically the ordering of α1 M / sqrt(1+M^2) + α2 N / sqrt(1+N^2) versus α1 G / sqrt(1+G^2) + α2 H / sqrt(1+H^2), does not follow. Thus the key inequality UB(EA(t+1)) ≤ UB(EO(t+1)) is not established, and the central optimality claim is unsupported.","section":"§3.5 and supplementary proof of Eq. (9)"},{"comment":"Lemma 3.1 assumes 'proper initialization' of the coordinate descent, a positive definite Hessian at the optimizer, and q-linear convergence of ALS with rates q1, q2, q3 in (0,1). Algorithm 2 initializes all latent factors to zero (lines 1–3) and provides no verification that the required conditions hold for the non-convex objective in Eq. (1). The cited result [40] is a local convergence theorem, and the supplementary text only says that the regularization parameter should be 'sufficiently large' without giving a sufficient condition. Moreover, with noisy observations the minimizer of the regularized empirical objective is not generally the ground-truth factors h*, a*, s* used in the lemma. Consequently, the uncertainty bounds α used in the selection rule (5) and in Eq. (8) have no validated basis, and the entire proof tree rooted at Lemma 3.1 is unsupported.","section":"Lemma 3.1 and Algorithm 2"},{"comment":"The paper claims a rigorous guarantee on 'accumulated estimation error' and on requiring 'less sensor deployment' than any other strategy, but the proof in §3.5 only compares upper bounds at a single future time for the single pair whose selection is exchanged, with all other errors asserted to be 'bounded by the same result.' It does not integrate the error over multiple months, does not treat a budget of L sensors per month, and does not compare against arbitrary deployment strategies over the full horizon. The empirical budget experiment (Figure 5) provides some support for the practical claim, but the stated theoretical conclusion is not a consequence of the proof as written. The authors should either weaken the stated theorem to a one-step local guarantee or provide a substantially extended proof.","section":"Abstract, §3.5, and conclusion"}],"minor_comments":[{"comment":"The initialization `E^0_t ← λ2I` appears to be a typo; for consistency with Eq. (4) and the parameter λ3, it should be `E^0_t ← λ3I`.","section":"Algorithm 2, line 3"},{"comment":"The statement 'for any ϵ1 > 0.ϵ2 > 0,ϵ3 > 0' uses a period instead of a comma or semicolon after '0'; this should be corrected.","section":"Section 3.3"},{"comment":"The improved-ratio plot contains unusual '□80' symbols in the y-axis tick labels, which appear to be rendering artifacts or missing negative signs; the figure should be regenerated to display the actual tick values.","section":"Figure 3(b)"},{"comment":"In the second term of the displayed bound for EA(t+1), the matrix `(C^{t+1}_{yo})^{-1}` is used where `(C^{t+1}_{ya})^{-1}` seems intended, given the definition of N; this inconsistency should be fixed.","section":"Supplementary Eq. (13)"},{"comment":"The sentence 'VBV gives worse energy breakdown quality than other approaches due to its poor performance in parameter estimation' attributes a causal explanation that the experiments do not directly isolate; consider softening the claim or adding supporting evidence.","section":"Section 4.3.1"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution of the paper is solid and reproducible, but the theoretical guarantee is the main advertised selling point and it is not proven as written. The issues in Lemma 3.1 and the proof of Eq. (9) are load-bearing; they cannot be addressed with minor edits. I recommend major_revision: the authors should either repair the proof with explicit, verifiable conditions or substantially weaken the theoretical claims to match what is actually shown. The season-pattern assumption is grounded in prior work co-authored by these authors [7], which reduces the novelty of that component, but the active-selection formulation and the empirical study are still useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the thing to know: this paper is worth reading for the problem formulation and the empirical study, not for the theory. The authors define active sensor deployment for energy breakdown as active tensor completion, with a sensible temporal-season projection, and show on the Dataport/Austin data that it beats random, QBC, and VB baselines, requiring fewer installed sensors to reach the same RMSE. Code is on GitHub, and the empirical comparison is consistent across four years. That is real value.\n\nThe soft spot is Section 3.5. Lemma 3.1 assumes q-linear convergence of ALS to the global optimizer with a proper initialization, and none of that is verified. Algorithm 2 initializes factors to zero; the objective is non-convex; the cited theorem is local. Without Lemma 3.1, the alpha uncertainty bounds have no basis. And the proof of Eq (9) claims componentwise M>=G and N>=H from the selection rule, but the rule only gives a weighted sum inequality. The one-step upper-bound comparison is also not the accumulated-error guarantee promised in the abstract. These are real gaps.\n\nThat said, I don't think the empirical contributions collapse. The improvement over baselines is consistent, and the temporal ablation makes sense. There are no error bars in the plots, which is a minor issue for a paper whose main quantitative claim is empirical. The theory could be weakened to a claim like 'the selection rule greedily reduces the current uncertainty bound' and the paper would still stand on the experiments. What I'd want from a revision: drop or substantially rewrite the optimality proof, add error bars or multiple runs, derive alpha rather than fixing it at 0.1, and be honest that the guarantee is one-step.\n\nCitation pattern: the self-citations to [7] for the season-pattern assumption are fine—that is prior work by the same group, but the assumption is also empirically supported by the paper's own Figure 2. No red flag there.\n\nFor a reader: this is a CIKM-type applied ML paper. It will be most useful to people working on collaborative sensing, energy disaggregation, or active tensor completion. I would send it out if I were the editor; it needs revision, but it is a serious paper with a clear contribution. My vote: major revision, not reject.","headline":"ActSense is a genuinely new active sensor-deployment formulation with solid empirical results on public data, but the theoretical optimality proof doesn't hold up and shouldn't be the paper's selling point.","tokens_in":20794,"tokens_out":1799,"would_cite":true,"duration_ms":17428,"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 that ActSense, which selects home-appliance pairs by quantified uncertainty in a low-rank tensor model, provably reduces accumulated energy-breakdown error faster than any other sensor installation strategy.","keywords":["active learning","tensor completion","energy breakdown","collaborative sensing","sensor deployment","uncertainty quantification","CP decomposition","alternating least squares"],"falsifier":"Run ActSense on the Dataport data, track the distance between successive ALS iterates and the final solution, and test whether each factor's error shrinks by a constant factor q<1 per round with a positive definite Hessian at the optimum; if the ratio fails to drop geometrically for any factor, Lemma 3.1's uncertainty bounds are not valid and the selection rule loses its proven guarantee.","tokens_in":19750,"feed_emoji":"⚡","tokens_out":6042,"duration_ms":52262,"temperature":0.7,"pith_summary":"The paper asks which homes and appliances should receive sub-meters so that appliance-level energy use in all homes can be reconstructed with the least instrumentation. It models readings as a low-rank home-by-appliance-by-month tensor and selects pairs with the highest quantified reconstruction uncertainty, using historical season patterns to anticipate future months. It claims a proof that this selection reduces accumulated estimation error at least as fast as any other strategy, with high probability. If correct, utilities could obtain the same energy-breakdown accuracy with far fewer installed sensors, lowering the cost barrier that currently prevents widespread deployment.","feed_headline":"Smarter sub-meter placement cuts energy-breakdown cost","feed_subtitle":"Choosing the most uncertain home-appliance pairs to instrument provably beats random deployment with fewer sensors.","key_machinery":"The central object is a three-way energy tensor with homes, appliances, and months as axes, completed by non-negative CP decomposition into home, appliance, and season factor matrices. The uncertainty bound in Lemma 3.1 turns the q-linear convergence rates of alternating least squares into per-factor confidence ellipsoids, and the selection rule picks the unobserved home-appliance pair maximizing a sum of two ellipsoid-norm terms. A triangle kernel integrates those uncertainty estimates over a 12-month window, so past, current, and projected future months all contribute to the decision; the proof that this one-step greedy rule is error-optimal relies on a Sherman-Morrison update of the Hessian after observing the selected pair.","core_discovery":"ActSense treats energy breakdown as active tensor completion. Readings from M homes, N appliances, and T months form a three-way tensor that is assumed low rank and factored by non-negative CP decomposition into home, appliance, and season factors. The paper derives analytic confidence ellipsoids for each factor (Lemma 3.1) under a Gaussian noise model, and at each month selects the L unobserved home-appliance pairs with largest combined uncertainty, where uncertainty is a weighted sum over a 12-month window that uses historical season factors for future months. The theoretical result is that, with probability at least 1-delta, the upper bound on the prediction error after ActSense's selection is no larger than the bound after any other selection; empirically, on four years of Austin Dataport data, ActSense beats random, QBC, and VBV baselines, reaching the same year-RMSE with about 3 sensors per month where QBC needs 8 and random needs 10.","pith_inferences":["Beyond the paper, the same greedy uncertainty rule transfers to other behind-the-meter monitoring problems, such as water, gas, or multi-zone HVAC, where aggregate readings are always available, time series are seasonal, and sensor installation is one-time.","Beyond the paper, because the proof relies on q-linear ALS convergence, the guarantee is unlikely to survive when the tensor rank is misspecified or data is too sparse to make the Hessian positive definite; in those regimes ActSense remains a reasonable heuristic but not a proven optimum.","Beyond the paper, a testable extension is to replace the fixed monthly budget L with a budget allocated across months by projected uncertainty; under the paper's assumptions this should lower the total sensors needed to reach a fixed accuracy."],"forward_implications":["Fixed-budget comparisons on the Austin Dataport data favor ActSense in every year tested, with mean-RMSE improvements over random selection reaching 11.88% on average and 35.06% in the peak month of 2015.","For a target year-RMSE of 50, ActSense needs roughly 3 new sub-metered pairs per month; QBC needs 8 and random needs 10, so the same breakdown accuracy can be had at a fraction of the instrumentation cost.","Injecting historical season factors and summing uncertainty over past, current, and projected months improves selection; the ablation shows the largest gain comes from projecting future months, letting the algorithm install HVAC sensors in May rather than waiting until summer.","The Sherman-Morrison argument in the proof implies the one-step greedy selection is optimal in the sense of worst-case error upper bound, so no single-step replacement of the uncertainty rule can improve the guarantee."],"supporting_citations":[{"why":"Supplies the tensor factorization approach for energy breakdown and the cross-year season-pattern assumption that ActSense uses for future projection.","marker":"[7]"},{"why":"Provides the matrix factorization collaborative sensing method that the tensor formulation extends.","marker":"[9]"},{"why":"Supplies the q-linear convergence result for alternating least squares that Lemma 3.1 and the selection proof rely on.","marker":"[40]"},{"why":"Supplies the self-normalized martingale bound used to control the noise term inside the uncertainty bounds.","marker":"[1]"},{"why":"Provides the Dataport dataset used for all empirical evaluations.","marker":"[30]"},{"why":"Provides the low-rank active learning approach that the VBV baseline extends from matrix to tensor completion.","marker":"[38]"},{"why":"Provides the Bayesian tensor factorization used to implement the VBV baseline's posterior-variance selection.","marker":"[42]"},{"why":"Provides an active matrix completion method that motivates the committee-based uncertainty querying used by the QBC baseline.","marker":"[11]"}],"fun_headline_variants":["Active sensor selection cuts energy breakdown cost","Active learning picks uncertain pairs to cut cost","Fewer sensors, same insight: active selection wins","Proven active selection beats random with fewer meters","Active tensor completion: less metering, same breakdown"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the alternating least squares routine used to fit the tensor converges geometrically to the true factors from its initialization; if that convergence does not hold on real data, the uncertainty bounds and the optimality guarantee collapse.","fun_headline_variants_meta":{"raw":{"variants":["Active sensor selection cuts energy breakdown cost","Active learning picks uncertain pairs to cut cost","Fewer sensors, same insight: active selection wins","Proven active selection beats random with fewer meters","Active tensor completion: less metering, same breakdown"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001052,"raw_usage":{"total_tokens":4424,"prompt_tokens":954,"completion_tokens":3470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":3400}},"tokens_in":570,"tokens_out":3470,"duration_ms":21474,"temperature":1.0,"reasoning_tokens":3400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:45:37.646871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ActSense on the Dataport data, track the distance between successive ALS iterates and the final solution, and test whether each factor's error shrinks by a constant factor q<1 per round with a positive definite Hessian at the optimum; if the ratio fails to drop geometrically for any factor, Lemma 3.1's uncertainty bounds are not valid and the selection rule loses its proven guarantee.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the low-rank active learning approach that the VBV baseline extends from matrix to tensor completion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Bayesian tensor factorization used to implement the VBV baseline's posterior-variance selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the tensor factorization approach for energy breakdown and the cross-year season-pattern assumption that ActSense uses for future projection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the q-linear convergence result for alternating least squares that Lemma 3.1 and the selection proof rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the self-normalized martingale bound used to control the noise term inside the uncertainty bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Dataport dataset used for all empirical evaluations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides an active matrix completion method that motivates the committee-based uncertainty querying used by the QBC baseline."}],"review_version":1}