{"id":"730e5437-8528-48c6-9760-8d25baaabc65","arxiv_id":"2412.09861","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A transfer learning framework using similar-intersection matching and TrAdaBoost.R2 estimates turning movement counts from traffic controller event data with lower MAE and RMSE than eight regression baselines on 30 Tucson intersections.","lead":"This paper builds a transfer learning model that estimates how many vehicles turn left, go straight, or turn right at traffic intersections, using data from traffic signal controllers instead of new physical sensors. It reports lower average errors than eight standard machine learning models on 30 intersections in Tucson, Arizona.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported TL advantage is confounded: TL selects and reweights source training data using target-intersection input features (Eq. 10, §3.4), while baselines receive no such target-feature access, so the lowest MAE/RMSE cannot be attributed to transfer learning.","rationale":"Strictly, the paper's strongest claim is the empirical superiority of the TL model over eight baselines. That claim rests on the evaluation protocol in §4.4. The protocol is asymmetric: the TL model is given the target intersection's feature vectors during training (to choose D_S' and the matched source intersection), whereas baselines are not. This is a transductive setup; it can be legitimate, but it must be compared against transductive baselines or ablated. The reader's weakest_assumption about the similarity-label transferability premise is related but distinct: even if feature similarity did imply label similarity, the current experiment would still not show that TrA is the reason for the gains, because the baselines were denied the same target-feature information. A second, compounding ambiguity is that the paper never states whether TL trains on all 29 source intersections or only on the matched source intersection; if only one matched source is used, the comparison is additionally unbalanced. The absence of per-intersection error distributions and significance tests further weakens the support. I do not think the paper should be rejected outright—the framework is plausible and the data are real—but the central claim should be conditional on a controlled comparison that equalizes information access.","tokens_in":14863,"tokens_out":6277,"duration_ms":67622,"concrete_test":"On the same 30 intersections, run three variants: (1) TL with random source-instance selection instead of cosine/Pearson matching, keeping the same number of instances and the same TrA procedure; (2) the best baseline (e.g., LightGBM or RF) trained only on the same matched source intersection/subset that TL uses, or with cosine-similarity instance weights derived from target features; (3) TL without target-feature access, using only source features for selection. Compare per-intersection paired MAE/RMSE with a Wilcoxon signed-rank test across the 30 intersections. If TL does not beat variant 2 or loses to random selection, the reported gain is due to target-feature access, not transfer learning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 evaluates by training on 29 intersections and testing on the 30th. For each target intersection, the TL pipeline first uses the target's input features to compute the Pearson-correlation match (Eq. 10) and the cosine-similarity threshold that selects the top 10% of source instances as D_S' (§3.4); TrA then treats D_S' as target data and upweights it (§3.5). The eight baselines are trained once on all 29 source intersections and applied to the target, with no step that uses target features to choose or weight training examples. Thus the comparison is not testing 'transfer learning vs. no transfer learning'; it is testing 'training with transductive access to target features plus instance reweighting' against 'inductive training without that access.' If selecting source rows by similarity lets the model concentrate on high-volume or structurally similar patterns, MAE/RMSE can improve even when the learned function is unchanged. The paper provides no ablation with random source selection, no baseline given the same matched subset or cosine weights, and no paired statistical test across the 30 holdout intersections, so the central claim (TL achieves the lowest errors) is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven transfer learning (TL) framework for estimating turning movement counts (TMCs) at signalized intersections. The pipeline combines Lasso regression for feature selection, a 'similar-intersection matching' step based on Pearson correlation between target and source intersection features (Eq. 10), a 'target domain data substitution' step that selects the top 10% of source instances by cosine similarity to the target features (Section 3.4), and a Two-stage TrAdaBoost.R2 (TrA) regression model. The framework is evaluated on 30 intersections in Tucson, Arizona, using a leave-one-out protocol, with event-based detector data, infrastructure data, and POI data. The reported results (Tables 3 and 4) show the TL model achieving the lowest average MAE and RMSE among nine models (eight baselines plus TL) for left-turn, through, and right-turn movements.","tokens_in":15126,"tokens_out":3114,"duration_ms":32183,"significance":"If the reported accuracy gains are real, the framework would be a practically valuable contribution: it estimates TMCs without additional physical sensors, using data already available to many traffic agencies, and it explicitly addresses domain shift across intersections. The paper merits credit for assembling a realistic multi-source dataset, implementing a complete TL pipeline, and performing a 30-intersection evaluation. However, the central claim of superiority is not yet established because the comparison is confounded: the TL method uses target-intersection input features for source selection and instance reweighting, while the baselines do not, and no statistical significance testing or ablation is provided. The transferability premise—that input-feature similarity implies label similarity—is load-bearing but unvalidated. The work is therefore promising but requires substantial additional analysis before the claims can be accepted.","major_comments":[{"comment":"The comparison between the TL model and the eight baselines is confounded: the TL pipeline uses the target intersection's input features to select a matched source intersection (Eq. 10, Section 3.3) and to select the top 10% of source instances by cosine similarity (Section 3.4), then reweights those instances via TrA (Section 3.5). The baselines are trained on all 29 source intersections with no access to target features. Consequently, the reported MAE/RMSE differences do not isolate 'transfer learning'; they compare transductive target-feature-matched instance weighting against inductive training. To support the claim that TL is the cause of improvement, the authors should include ablations such as random source-instance selection, cosine-matched source instances without TrA reweighting, and a variant that gives the baselines the same matched subset. They should also report paired statistical tests (e.g., Wilcoxon signed-rank or paired t-test over the 30 holdout intersections) with effect sizes and error bars.","section":"Section 4.2.2 and Section 4.4"},{"comment":"The hyperparameter tuning procedure is not described relative to the leave-one-out (LOO) evaluation. The paper states that 'a grid search method was initially developed to identify the optimal hyperparameter values for baseline models and the proposed TL method' (Section 4.2.2), but it does not specify whether tuning is performed inside each of the 30 LOO folds or once on all data. If the target intersection's data participates in selecting the threshold θ (top 10%), the TrA parameters S and F, or the baseline hyperparameters, the reported test errors are optimistically biased. The authors must either describe a nested cross-validation scheme or state explicitly that all hyperparameters were fixed before LOO evaluation.","section":"Section 4.2.2, Section 4.4"},{"comment":"The framework's central transferability assumption is that source intersections with high Pearson correlation on the selected input features, and source instances with high cosine similarity to target features, have similar turning movement counts. This assumption is load-bearing because the selected source instances' labels substitute for the unavailable target labels in Eq. (6). The paper does not validate this premise against random selection or alternative similarity measures, nor does it report any diagnostic showing that matched intersections/instances actually have more similar TMC values than unmatched ones. Without such a check, the reported accuracy gains could stem from the concentration of high-volume similar instances rather than from a meaningful correspondence of turning patterns.","section":"Sections 3.3 and 3.4, Eq. (10)"},{"comment":"Only point estimates of MAE and RMSE are reported—the averages across the 30 LOO test intersections. There are no confidence intervals, standard deviations, or significance tests. The paper's qualitative statements in Section 4.4, such as 'the TL model consistently shows the lowest MAE values' and 'exhibits the lowest and most consistent RMSE values', go beyond what the tables support. The authors should provide per-fold statistics and a formal comparison (e.g., paired tests across the 30 folds) so the reader can judge whether the differences are reliable given the small number of target intersections.","section":"Section 4.4, Tables 3 and 4"}],"minor_comments":[{"comment":"The abstract contains a typo: 'intersections, Accurate TMCs' should be 'intersections. Accurate TMCs'. The phrase 'state-of-the-art regression models' appears in the abstract but the paper does not define what 'state-of-the-art' means in this context; listing the eight models in Section 4.2.2 is helpful, but the abstract should align with the actual evaluation.","section":"Abstract and Section 1"},{"comment":"The sentence 'On average, data from one to four days per intersection were utilized for model training and validation' is vague. It should specify how many 15-minute intervals per intersection are in the training and test sets, and how many total instances (n) are used. This information is necessary for interpreting the MAE/RMSE magnitudes.","section":"Section 4.2.1"},{"comment":"The baseline 'LightBoost' is elsewhere called 'LightGBM' (Section 4.4 and Table captions). Please use a consistent name throughout.","section":"Section 4.2.2"},{"comment":"Reference [Dai Wenyuan, Y.Q., Guirong, X. and Yong, Y. (2007)] is a malformed duplicate of the Dai et al. (2007) entry. Please consolidate and correct the citation format.","section":"References"},{"comment":"The pseudocode for TrA is underspecified: the exact formula for the adjusted error e_i^t, the definition of β_t, and the stopping criterion in the AdaBoost.R2 call are not given. Since the proposed method relies on this algorithm, please provide the complete algorithm or a precise reference to the version used.","section":"Section 3.5, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and addresses a transportation engineering problem with a machine learning approach. The main concern is the evaluation protocol: the TL model's access to target features during source selection and instance weighting makes the comparison with inductive baselines unfair, and the lack of statistical validation weakens the central claim. The paper would be substantially improved by adding ablations, paired significance tests, and a clear description of hyperparameter tuning relative to the LOO procedure. I do not see a load-bearing error that would justify rejection; the issues are correctable within the manuscript's scope. I would recommend major revision. Also note that the claim of being 'the first attempt to employ TL for TMC estimation' seems inconsistent with the authors' own prior TL work on ramp flow estimation (Ma et al., 2023c, 2024), and should be moderated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate first application of transfer learning to turning movement count estimation, using real controller event data from 30 Tucson intersections. The data work is real, and the comparison against eight baselines is extensive. But the main claim — that TL beats generic models — is not actually established by the numbers as reported, because the TL pipeline sees the target intersection's features when it selects which source data to use, and the baselines don't.\n\nWhat's genuinely new here is applying the source-selection-and-substitution scheme to TMC data; I don't see that in the cited literature. The feature set (detector occupancy, green time, cycle counts, lane geometry, POI) is reasonable, and Lasso gives a sensible ranking. The evaluation is on real intersections, not simulations, and the reported error numbers are internally consistent.\n\nThe stress-test note is fair. The TL model computes Pearson correlation between each target and each source intersection (Eq. 10) and then uses cosine similarity to keep the top 10% of source instances that are closest to the target features (§3.4). That's transductive use of target inputs. Baselines are trained once on all 29 source intersections and applied to the target, with no analogous feature-based selection. So the comparison conflates transfer learning with target-aware data selection. An ablation with random source selection, or giving baselines the same matched subset, would be needed. As it stands, the reported 20-30% error reduction could come mostly from selecting rows that are structurally similar to the target, not from the TrAdaBoost.R2 mechanism itself.\n\nAlso missing: error bars or significance tests across the 30 holdout intersections, no description of hyperparameter tuning for the weak learner (the weak learner G is never specified), and no code/data release. Minor issues: the text spends a lot of space on generic introductions of Lasso and AdaBoost, and the claim to 'guarantee' high accuracy in the intro is overblown. None of this is fatal; it's fixable with a proper experimental design.\n\nWho this is for: transportation researchers and practitioners working on TMC estimation or sensor-free traffic data; also a good teaching case for why transductive vs inductive baselines matter. It deserves a serious referee — the application is relevant, the data are real, and the gap is correctable. But the revision needs to fix the comparison, add significance tests, and ideally release code/data.","headline":"A plausible first application of transfer learning to turning movement counts, undermined by a confounded evaluation: the TL model gets target-feature access during source selection while the baselines don't.","tokens_in":15650,"tokens_out":2072,"would_cite":false,"duration_ms":21164,"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":"A transfer learning model that borrows data from similar intersections estimates turning movement counts with lower errors than eight standard regression models.","keywords":["transfer learning","turning movement counts","traffic flow estimation","TrAdaBoost.R2","traffic controller event data","point-of-interest data","lasso regression","intersection traffic"],"falsifier":"Re-run the exact TrA pipeline on the same 30 intersections, but replace the Pearson-correlation intersection matching and the cosine-similarity top-10% selection with random matching and random selection of the same sizes; if the average MAE remains close to 9.57/28.59/10.25 and RMSE close to 13.65/40.62/13.98, the similarity mechanism is not doing the work and the central claim fails.","tokens_in":14627,"feed_emoji":"🚦","tokens_out":8211,"duration_ms":75655,"temperature":0.7,"pith_summary":"Turning movement counts (left, through, right turns) at signalized intersections are usually collected by sensors or manual counting, which is costly to scale. This paper proposes a transfer learning framework that estimates those counts for intersections without sensors by reusing data from instrumented 'source' intersections. The framework selects important traffic variables with Lasso regression, matches each target intersection to a similar source intersection by Pearson correlation, picks the most similar 10% of source instances by cosine similarity, and fits a Two-stage TrAdaBoost.R2 model. On 30 Tucson intersections, the model reports the lowest average MAE and RMSE for all three movement types compared with eight baseline regression models. If the result holds, traffic agencies could obtain turning movement counts at unmonitored intersections from existing controller event data, road geometry, and points-of-interest data.","feed_headline":"Transfer learning beats 8 models at turning-movement counts","feed_subtitle":"The transfer-learning model posts the lowest error for left, through, and right turns at 30 Tucson intersections.","key_machinery":"The load-bearing mechanism is Two-stage TrAdaBoost.R2 (TrA), a boosting procedure that begins with a combined set of source and selected source instances, assigns equal initial weights, and at each round reweights source instances by their adjusted error while keeping target-like instances' weights controlled by a pre-specified schedule, finally outputting the weak estimator with the lowest cross-validation error. Before boosting, the framework identifies comparable intersections by maximizing the sum of Pearson correlations between the same Lasso-selected traffic variables of source and target intersections (Eq. 10), then substitutes the top 10% of source instances ranked by cosine similarity to the target instances as pseudo-target labeled data. This chain carries the argument: feature selection, similarity matching, instance substitution, and TrA reweighting together let labeled source data stand in for missing target labels.","core_discovery":"The paper's central claim is that an instance-based transfer learning model, Two-stage TrAdaBoost.R2, can estimate turning movement counts at a target intersection using no observed target labels, by reweighting source-domain instances that resemble the target and by first matching the target to a similar source intersection. It reports average test errors on 30 Tucson intersections of MAE 9.57 for left turns, 28.59 for through movements, and 10.25 for right turns, with RMSE 13.65, 40.62, and 13.98, all lower than the corresponding errors of KNN, SVR, RF, MLP, AdaBoost, XGBoost, CatBoost, and LightGBM. The paper argues that the gains come from relaxing the assumption that source and target data distributions are identical, which lets the model build scene-specific estimators for intersections with different geometry, signal timing, and local behavior.","pith_inferences":["The paper does not test whether its Pearson-correlation intersection matching and cosine-similarity instance selection beat random selection; a reader should treat the transferability premise as unproven until such an ablation is run.","Because the evaluation protocol trains on the other 29 intersections for each target, the source library always contains near-neighbors of the target; deployed systems with a smaller or less similar source pool would likely see smaller gains than the 9.57/28.59/10.25 MAE values reported.","The same instance-reweighting chain could be applied to other intersection-level quantities such as queue lengths, delays, or pedestrian volumes, but the similarity premise would need to be re-validated for each target quantity.","A direct comparison against a model fine-tuned with a small number of real target labels would clarify how much of the gain comes from transfer versus from the boosting procedure itself."],"forward_implications":["If the central claim is right, turning movement counts can be estimated at intersections that lack loop detectors or cameras, using data already collected by traffic signal controllers.","A city could build a library of instrumented intersections and, for any new intersection, automatically select similar sources and obtain movement-specific estimates without field counting.","The reported accuracy gains imply that scene-specific models, rather than one global model, are the better default for TMC estimation across heterogeneous intersections.","Because the method uses controller event data plus road and POI features, it can be extended to new intersections quickly at low marginal cost.","The framework's reweighting mechanism could be re-run whenever new controller data arrive, giving a scalable path to continuous traffic monitoring."],"supporting_citations":[{"why":"Supplies the TrAdaBoost transfer learning algorithm that the proposed Two-stage TrAdaBoost.R2 is built on.","marker":"Dai et al., 2007"},{"why":"Extends boosting to regression transfer, providing the two-stage regression variant used here.","marker":"Pardoe & Stone, 2010"},{"why":"Defines AdaBoost.R2, the base regressor-boosting procedure called inside TrA.","marker":"Drucker, 1997"},{"why":"Introduces Lasso regression, used for variable selection before matching and transfer.","marker":"Tibshirani, 1996"},{"why":"Given as the source for the top-10% cosine-similarity threshold used to select substitute source instances.","marker":"Dai Wenyuan, Y.Q., Guirong, X. and Yong, 2007"},{"why":"Shows that traffic controller event-based data can support network-level TMC estimation, the data foundation the framework relies on.","marker":"Xu et al., 2023"},{"why":"Prior transfer learning framework for ramp flows whose instance-weighting idea the paper adapts to intersections.","marker":"X. Ma et al., 2024"}],"fun_headline_variants":["Transfer learning with no target labels beats 8 models","Label-free transfer learning beats 8 models at turn counts","Zero-shot transfer learning for turning-movement counts beats 8","Transfer learning wins at turning-movement counts without target data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that an intersection or a single data row that resembles the target on detector activity, lane counts, POI, and time will also have similar turning movement counts, so the labels of the most similar source rows can stand in for the missing target labels.","fun_headline_variants_meta":{"raw":{"variants":["Transfer learning with no target labels beats 8 models","Label-free transfer learning beats 8 models at turn counts","Zero-shot transfer learning for turning-movement counts beats 8","Transfer learning wins at turning-movement counts without target data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001851,"raw_usage":{"total_tokens":7275,"prompt_tokens":950,"completion_tokens":6325,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":6258}},"tokens_in":566,"tokens_out":6325,"duration_ms":43929,"temperature":1.0,"reasoning_tokens":6258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:38:27.682086+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact TrA pipeline on the same 30 intersections, but replace the Pearson-correlation intersection matching and the cosine-similarity top-10% selection with random matching and random selection of the same sizes; if the average MAE remains close to 9.57/28.59/10.25 and RMSE close to 13.65/40.62/13.98, the similarity mechanism is not doing the work and the central claim fails.","supporting_citations":[],"review_version":1}