{"id":"4a4c7c93-4c0e-4b33-92f0-54ae6bffbd5b","arxiv_id":"2505.06289","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The thesis contributes the Plegma dataset, a 10-second, one-year, 13-house Greek electricity dataset, and shows that lottery-ticket-style pruning before training can shrink NILM models to 5% of their original parameters with reported performance close to the full model.","lead":"This thesis contributes a new electricity dataset from 13 Greek households and tests whether pruning deep learning models before training lets appliance-level energy disaggregation run on edge devices. A generalist might read it to judge whether regional energy data plus cheaper models can make home energy feedback practical.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GCN encoder uses test-time appliance entropy to build the graph, so the claimed NILM advantage (Table 4.2) is not a valid non-intrusive result without a leakage-free evaluation.","rationale":"The reader's weakest_assumption correctly identifies the GCN encoder's reliance on test-time appliance-level entropy. This is the single most load-bearing concern because it invalidates one of the three headline contributions: the claim to a first graph-based NILM method that outperforms sequential baselines. The limitation is explicitly disclosed in Sec. 4.1, but that does not reduce its impact: a method that requires the target appliance's ground-truth signal at inference is not a non-intrusive load monitoring method under the problem's own definition (Eq. 4.1). The reported F1 comparison in Table 4.2 therefore measures an oracle-aided upper bound, not a deployable disaggregation result. I agree with the reader's choice and with the CONDITIONAL verdict: the dataset and pruning chapters may still stand, but the GCN chapter must be reframed as a graph-based proof-of-concept with a stated oracle requirement, or re-evaluated without leakage. Other concerns the reader raised, such as the scale-imbalanced threshold metric in Eq. (5.3) and threshold selection on the test house, are also valid and strengthen the case for conditional acceptance, but they are secondary because they affect quantitative rigor rather than the fundamental validity of a claimed contribution. The proposed concrete test would settle whether the GCN advantage persists without ground-truth test-time information.","tokens_in":47907,"tokens_out":5650,"duration_ms":56759,"concrete_test":"Implement a leakage-free evaluation of the Chapter 4 GCN: fit the cluster codebook and Markov transition matrix using only training-house aggregate windows and appliance labels; at test time on UK-DALE house 2, assign each aggregate window to the nearest cluster centroid using aggregate-only features (e.g., window statistics), without accessing the washing machine's test signal. Recompute F1 and MAE for the same decoder. If F1 falls to or below the CNN baseline (0.7261), the Table 4.2 advantage is attributable to test-time ground-truth leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the GCN-encoder/transformer-decoder is the first graph-based NILM and outperforms CNN/LSTM/GRU baselines on the washing machine (Table 4.2) rests on Sec. 4.3.3, where aggregate windows are assigned to clusters 'based on the entropy metric of the corresponding appliance's consumption windows.' At test time in real NILM, only the aggregate is observed; the target appliance's signal is unavailable, so the graph nodes and Markov transition weights cannot be computed. Sec. 4.1 explicitly concedes: 'our approach could only be used to disaggregate known datasets as it provides the limitation of knowing the entropy of the appliance's signal even on the testing data.' Thus F1=0.7805 vs CNN=0.7261 is measured with access to the ground-truth appliance signal, not a fair comparison of disaggregation methods. This is not merely an external-validity caveat; it contradicts the NILM problem formulation in Eq. (4.1), which defines the task as determining ym from xagg alone. Without a leakage-free variant, the GCN contribution does not support the thesis's 'advanced DNN for NILM' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This PhD dissertation combines an interoperable IoT data-collection framework and a new residential energy dataset, the Plegma Dataset from Greece, with a set of deep-learning contributions for non-intrusive load monitoring (NILM). The dataset chapter reports 13 households, 10-second aggregate and appliance-level measurements, roughly 218 million readings, environmental data, and sociodemographic/building metadata, together with explicit cleaning rules (Algorithms 1-2). The modeling chapters claim three advances: a GCN-encoder/transformer-decoder as the first graph-based NILM model, an iterative pre-training pruning scheme ('OPT-NILM') that finds subnetworks matching full-model performance at up to 95% sparsity, and structured dependency-graph pruning evaluated on the Plegma dataset. The pruning chapters report large complexity reductions on UK-DALE and Plegma, while the GCN chapter reports improved washing-machine F1 over CNN/LSTM/GRU baselines.","tokens_in":48124,"tokens_out":5779,"duration_ms":58905,"significance":"If the dataset contribution stands, it is genuinely useful: Plegma is, as claimed, one of the first public 10-second residential datasets from Greece and the Mediterranean, with underrepresented appliances such as air conditioners and electric water boilers, and the chapter is careful about missing data (6.86% NaN, 0.82% issue flags) and about reproducible preprocessing. The pruning chapters address a real deployment bottleneck and the before-training pruning idea is worth pursuing. However, the significance of the modeling claims is currently conditional: the GCN result is not a valid NILM result as evaluated, because it uses test-time appliance entropy, and the pruning 'optimal threshold' is selected on the same test-house curves used for the final reported numbers. These are load-bearing issues, not presentation details.","major_comments":[{"comment":"The GCN encoder constructs graph nodes by assigning aggregate windows to clusters 'based on the entropy metric of the corresponding appliance's consumption windows' (§4.3.3). At inference in the NILM setting defined by Eq. (4.1), only the aggregate x(t) is observed, so the target appliance's entropy is not available; §4.1 concedes that the approach 'could only be used to disaggregate known datasets.' Consequently, the washing-machine comparison in Table 4.2 (F1 0.7805 vs 0.7261 for CNN) is obtained with access to ground-truth appliance information and is not a valid non-intrusive disaggregation result. The claimed advantage over sequential baselines therefore does not support the contribution as stated; a leakage-free variant (e.g., clustering on aggregate-only features or a state estimator trained without test labels) would be needed.","section":"§4.3.3 and §4.1"},{"comment":"The optimal pruning threshold p̂t is selected by minimizing the distance in Eq. (5.3) over the 'F1 score - MACs' curves, and the experimental setup states that models are trained on houses 1,3,4,5 and tested on house 2. As presented, the F1-MAC curves in Fig. 5.1 are the performance-degradation curves on which the final results in Figs. 5.2-5.3 and Table 6.1 are reported, so p̂t is an in-sample selection rather than an independent model-selection choice. The same issue appears in Chapter 8, where Eq. (8.10) sets Pthr = Popt from the test-house curves in Fig. 8.1. Please use a validation-house split for threshold selection and report test results only for the final threshold, or explicitly label the reported numbers as oracle-selected.","section":"§5.3.1, Eq. (5.4); §8.4.3"},{"comment":"The proposed trade-off metric mixes the F1 score, which lies in [0,1], with raw MAC counts, which in Fig. 5.1 are of order 10^6. As printed, Eq. (5.3) is dist = sqrt((1-F1)^2 + (0-MACs(pt))^2); without a normalization factor for MACs, the distance is dominated by the MAC term, so the 'optimal' threshold is essentially the most aggressive sparsity that keeps F1 above zero. Please specify the scaling (e.g., normalize MACs by the baseline model's MACs) and state the exact form used to produce the reported thresholds; the current formula also appears to contain a typo (missing closing parenthesis).","section":"§5.3.1, Eq. (5.3)"}],"minor_comments":[{"comment":"The table titled 'Monitored appliances in each house' actually reports aggregate consumption, sub-metered consumption, and the percentage of sub-metered consumption; the title appears to be copied from Table 3.3 and should be corrected.","section":"§3.4, Table 3.6"},{"comment":"The MRE and MAE definitions are corrupted in the typesetting; as printed, MRE lacks an explicit summation/division structure and the MAE expression is incomplete. Please provide clean equations.","section":"§4.4.2, Eq. (4.10)"},{"comment":"The Plegma dataset is cited as [9] in Chapter 1 but as [10] in Chapter 3; please harmonize the citation numbering throughout the thesis.","section":"§1.4 and §3.1"},{"comment":"The fridge row reports the number of activations as 'continuous' in a column labeled 'Total number of appliance activations'; please clarify how 'continuous' is defined and whether it is used in downstream analyses.","section":"§3.4, Table 3.7"},{"comment":"Table 4.1 reports F1 = 0.00000 for kettle and microwave; the text explains that the model is unsuited to short-duration devices, but reporting precision and recall or activation-level detection metrics would make the failure mode more informative.","section":"§4.4.3, Table 4.1"}],"recommendation":"major_revision","confidential_remarks":"The dataset chapter is the strongest part of the submission and could stand as a publishable unit on its own. The GCN chapter's explicit limitation statement in §4.1 should be moved to the front of the contribution claims; the abstract and Chapter 1 currently overstate what is demonstrated. Please also ensure that the pruning-threshold protocol is corrected before any final acceptance, since the in-sample selection affects the headline claims of Chapters 5-8."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it mostly as a compilation thesis. The Plegma dataset is the real contribution. One year, 13 Greek households, 10-second aggregate and appliance readings, 218M records, with ACs and electric boilers plus environmental data and sociodemographic metadata. The methodology chapter is concrete: cleaning rules, NaN rates, issue flags, synchronization approach, all internally consistent. That fills a genuine regional gap; I'd trust it as a benchmark.\n\nThe GCN chapter does not support the headline claim. The encoder assigns aggregate windows to clusters using the entropy of the target appliance's consumption windows, and the author explicitly concedes in Sec 4.1 that the approach can only disaggregate known datasets because the appliance's entropy is needed on test data. So the F1 improvement over CNN/LSTM/GRU in Table 4.2 is measured with access to ground-truth appliance behavior. That's not NILM as defined in Eq. 4.1. The stress-test note got this right. The graph formulation is novel as a framing, but it needs a leakage-free variant before it can be called a NILM result.\n\nThe pruning chapters are mixed. Pre-training iterative magnitude pruning is a reasonable application of the lottery ticket idea to NILM, and the compression numbers (e.g., 5% of parameters) are plausible. But the optimal threshold is selected from the same test-house performance curves that are then reported, so the headline results are in-sample selections. The metric in Eq. 5.3 has a scale problem: MACs are in the millions while (1-F1) is around 0-1, so the Euclidean distance is dominated by the MACs term and forces the most aggressive pruning level. Also everything is single-split point estimates. These are fixable: hold out a threshold-selection split, run multiple seeds, report error bars.\n\nCitation pattern is mostly fine; the self-citations are to the author's own published chapters, which is normal in a compilation thesis. The dataset work is honestly documented.\n\nFor a reader: use the dataset chapter; treat the pruning chapters as preliminary but promising; skip the GCN's performance claims until a leakage-free evaluation appears. Worth a serious referee because the dataset deserves scrutiny and the pruning idea is worth pinning down.","headline":"Plegma dataset is a solid regional contribution; the GCN result is invalid as NILM due to test-time leakage, and pruning claims rest on in-sample threshold selection.","tokens_in":48757,"tokens_out":1773,"would_cite":true,"duration_ms":17715,"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":"Pruning before training finds subnetworks of 5% of the parameters that match full NILM model accuracy, a new Greek 10-second dataset fills the Mediterranean data gap, and a graph-based encoder beats sequential baselines on washing machines.","keywords":["non-intrusive load monitoring","energy disaggregation","Plegma dataset","iterative magnitude pruning","graph neural networks","edge deployment","Mediterranean household energy","transformer decoder"],"falsifier":"Train the GCN-encoder model on UK-DALE houses 1, 3, 4, and 5, then evaluate on house 2 without computing cluster assignments from house 2's appliance-level signal, for instance by reusing the training clusters or clustering the aggregate alone. If the washing-machine F1 collapses from the reported 0.7805 toward the CNN baseline of 0.7261, the claim's test-time dependence on knowing the appliance's entropy is confirmed as load-bearing; if the F1 holds, the graph encoder generalizes without that oracle.","tokens_in":47600,"feed_emoji":"⚡","tokens_out":13595,"duration_ms":113221,"temperature":0.7,"pith_summary":"This thesis sets out to make non-intrusive load monitoring (NILM), inferring which appliance is using what power from a single household meter, practical outside the lab. It argues that real deployment is blocked by two gaps: a data gap, because public datasets come mostly from the USA and UK and miss Mediterranean staples like air conditioners and electric water boilers, and a compute gap, because deep disaggregation models are trained and run on cloud infrastructure, raising cost and privacy concerns. To close the first gap it introduces the Plegma dataset, one year of 10-second readings from 13 Greek households with roughly 218 million measurements. To close the second it claims that iterative magnitude pruning before full training can isolate subnetworks with only 5% of the original parameters that disaggregate as well as the full model, and that a graph-based architecture with a transformer decoder outperforms sequential baselines on multi-state appliances. If these claims hold, appliance-level feedback could run on low-cost devices in the home, including in regions whose consumption patterns current models have never seen.","feed_headline":"Pruning before training cuts NILM models to 5% with no accuracy loss","feed_subtitle":"A 10-second Greek household dataset plus pre-training pruning makes appliance-level monitoring practical on edge devices.","key_machinery":"The argument runs on two mechanisms. The first is iterative magnitude pruning before training, in the lottery-ticket style: the network is trained for a single epoch, the weights with the smallest L1 norm below a threshold $p_t$ are masked out via a binary mask $\\mu \\in \\{0,1\\}^N$ so that $\\hat{w} = \\mu \\odot w$, and the surviving weights are reset to their initial values $w_0$; repeating this for a few rounds $R \\ll K$ yields a sparse subnetwork that is then fully trained. This is what lets the scheme claim compute savings in both the training and the inference phase, since the expensive full model is never trained. The second is the entropy-based graph construction for the GCN encoder: aggregate windows are assigned to clusters according to the sample entropy of the target appliance's consumption windows, cluster means seed the node embeddings, Markov-chain transition probabilities between clusters define the directed edge weights, and stacked graph convolutional layers propagate information across nodes before a transformer decoder produces the appliance signal.","core_discovery":"The thesis's central claim is that the two obstacles standing between NILM research and real deployment, missing data for Mediterranean consumption patterns and the high compute cost of deep disaggregation models, are both removable. On the data side, it presents the Plegma dataset: a year-long, 10-second-resolution record of aggregate and appliance-level power from 13 Greek households, roughly 218 million readings, covering air conditioners (18 units) and electric water boilers (12 units) that rarely appear in other public datasets, alongside environmental and sociodemographic metadata. On the compute side, it argues that iterative L1-magnitude pruning performed before full training, with surviving weights reset to their initialization, isolates subnetworks containing 5% of the original parameters whose disaggregation performance matches the full model on the UK-DALE benchmark, and that the same idea in structured and dependency-graph form cuts model size up to 90% when evaluated on Plegma. It further claims that a sequence-to-sequence model with a graph convolutional encoder and a transformer decoder is the first graph-based NILM, recovering multi-state appliances like the washing machine better than CNN, LSTM, and GRU baselines (F1 0.7805 versus 0.7261 for the CNN baseline).","pith_inferences":["The mask-reset pruning recipe is demonstrated on one CNN architecture; a direct extension is testing whether the same schedule holds for transformer or GCN backbones, whose per-layer sparsity tolerances likely differ.","Plegma's combination of UK-DALE-comparable granularity and Mediterranean appliance mix makes a cross-dataset transfer experiment the natural next test: train on UK-DALE, test on Plegma, and quantify how much air-conditioner and boiler patterns degrade models that never saw them.","The disclosed test-time-entropy limitation suggests a concrete fix the thesis does not pursue: learning the cluster assignment from the aggregate signal alone, which would make the graph encoder truly non-intrusive.","The questionnaire metadata (occupancy, income, heating type) enables conditioning disaggregation on household context, a direction the dataset opens but the thesis leaves unexplored."],"forward_implications":["Training NILM models on the edge becomes plausible: if subnetworks found before training match full models, the resource-heavy full model never has to be trained centrally, and the reported 5% parameter subnetworks match full-model performance on UK-DALE.","The Plegma dataset gives researchers a Mediterranean benchmark with 10-second granularity, enabling NILM evaluation on air conditioners and electric water boilers and comparisons of cross-region transfer against UK-DALE and REFIT.","Structured and dependency-graph pruning variants, validated on Plegma, reach up to 90% model-size reduction and a 10x compute saving, a level suited to Raspberry-Pi-class devices.","For multi-state appliances such as washing machines, graph-based encoding with a transformer decoder beats sequential baselines on activation detection, with F1 of 0.7805 compared with 0.7261 for the CNN baseline."],"supporting_citations":[{"why":"Supplies the lottery-ticket-style hypothesis that a randomly initialized network contains a subnetwork matching the full model's accuracy; the foundation of the pre-training pruning scheme.","marker":"[60]"},{"why":"The after-training magnitude-pruning edge-NILM baseline whose performance and compression the proposed scheme is compared against.","marker":"[105]"},{"why":"Another post-training compression baseline (quantization of the seq2point CNN) that the thesis cites as failing to reduce training-phase cost.","marker":"[2]"},{"why":"UK-DALE, the benchmark dataset on which the pre-training pruning experiments in Chapters 5 and 6 are run.","marker":"[95]"},{"why":"The CNN seq2seq NILM work that serves as the main baseline in Table 4.2 and the compressed architecture's reference point.","marker":"[191]"},{"why":"Defines the seq2seq CNN formulation used as the base architecture for the pruning experiments.","marker":"[88]"},{"why":"Provides the graph convolutional network formalism, message passing over nodes, that the encoder builds on.","marker":"[99]"},{"why":"Establishes using GCNs inside sequence-to-sequence models, the inspiration for modeling NILM as a graph problem.","marker":"[184]"},{"why":"The dataset collection and storage recommendations the Plegma dataset is explicitly designed to satisfy.","marker":"[101]"},{"why":"REFIT, the comparable 8-second UK dataset used for granularity choice and for benchmarking the NaN rate in technical validation.","marker":"[125]"}],"fun_headline_variants":["Pruning before training yields 5% NILM models with no accuracy drop","Plegma dataset and pre-training pruning make edge NILM practical","Graph-transformer NILM recovers appliances better than CNN baselines","Pruning to 5% before training preserves NILM accuracy on UK-DALE","Edge NILM with 10-second Greek data and 90% smaller models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, disclosed by the thesis itself, is that the graph encoder knows the target appliance's consumption entropy on the test data; in a genuine deployment, where the appliance signal is exactly what is being inferred, that information is not available and the graph cannot be built.","fun_headline_variants_meta":{"raw":{"variants":["Pruning before training yields 5% NILM models with no accuracy drop","Plegma dataset and pre-training pruning make edge NILM practical","Graph-transformer NILM recovers appliances better than CNN baselines","Pruning to 5% before training preserves NILM accuracy on UK-DALE","Edge NILM with 10-second Greek data and 90% smaller models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3478,"prompt_tokens":1069,"completion_tokens":2409,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":2306}},"tokens_in":685,"tokens_out":2409,"duration_ms":16914,"temperature":1.0,"reasoning_tokens":2306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:31:39.128765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the GCN-encoder model on UK-DALE houses 1, 3, 4, and 5, then evaluate on house 2 without computing cluster assignments from house 2's appliance-level signal, for instance by reusing the training clusters or clustering the aggregate alone. If the washing-machine F1 collapses from the reported 0.7805 toward the CNN baseline of 0.7261, the claim's test-time dependence on knowing the appliance's entropy is confirmed as load-bearing; if the F1 holds, the graph encoder generalizes without that oracle.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The CNN seq2seq NILM work that serves as the main baseline in Table 4.2 and the compressed architecture's reference point."},{"cited_title":"Document Graph for Neural Machine Translation","cited_arxiv_id":"2012.03477","evidence_quote":"Establishes using GCNs inside sequence-to-sequence models, the inspiration for modeling NILM as a graph problem."}],"review_version":1}