Pith. sign in

REVIEW 3 major objections 5 minor 4 references

Unveiling Location-Specific Price Drivers: A Two-Stage Cluster Analysis for Interpretable House Price Predictions

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Clustering listings by location and features before fitting interpretable models gives a 36–58% mean-absolute-error advantage over global models.

desk verdict The reported gains are overstated (wrong denominator and likely target leakage in clustering/feature selection), but the two-stage cluster-then-predict design and cluster-level EBM analysis are worth a careful revision. read the letter →

arxiv 2508.03156 v1 pith:NFRZK76W submitted 2025-08-05 cs.LG

Paul Gümmer , Julian Rosenberger , Mathias Kraus , Patrick Zschech , Nico Hambauer This is my paper
classification cs.LG
keywords housepricepredictiontwo-stageclusteringinterpretablemachinelearningExplainableBoostinggeneralizedadditivemodelk-meanslocation-specificpricingGermanhousingmarket
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims the accuracy-versus-transparency tension in house price prediction can be relaxed: instead of fitting one model to an entire national market, split the market into location-defined segments and fit a simple, interpretable model inside each one. On 43,309 German house-sale listings from 2023, the authors report that their two-stage clustering scheme reduces mean absolute error for an Explainable Boosting Machine from €110,387 to €80,925, so the unclustered error is about 36% higher, and for penalized linear regression from €170,745 to €107,996, about 58% higher. The first clustering stage groups properties by latitude, longitude, and price level; the second stage refines those groups by property characteristics. The result matters because the accuracy gain does not come at the cost of opacity: every final prediction is made by a model small enough to inspect, and the clusters themselves reveal that factors like construction year and living space have different price effects in different sub-markets.

What carries the argument

The central mechanism is a two-stage cluster-then-predict pipeline. Stage one applies k-means with k=2 to location coordinates (latitude and longitude) plus price-target information, isolating two broad price regions; stage two applies k-means with k=8 to the full feature set within each region, and a regression-tree variant with depth three is compared but performs slightly worse. Each final sub-cluster is modeled separately by a linear model with L1 penalty (interpretable coefficients) or by an Explainable Boosting Machine, a generalized additive model that learns one non-linear shape function per feature. This division of labor is what carries the argument: the first stage absorbs the dominant price effect of location, and the second stage makes the remaining feature-price relationships homogeneous enough for simple models to fit accurately.

What would settle it

Re-run the full pipeline with the first-stage clusters estimated only on each cross-validation training fold, using the same 43,309 listings, the same k values, and the same models; if the mean-absolute-error advantage over unclustered models collapses substantially or disappears, the reported 36–58% gains are largely an artifact of target-informed clusters.

Watch

Extended reading notes

Core claim

The paper's central claim is that location is not just a feature to feed a model but a structural axis that should shape the model itself. Its two-stage procedure first runs k-means with k=2 on latitude and longitude together with information from the price target, separating expensive from less expensive regions, then re-clusters each region with k=8 using the full feature set; each sub-cluster receives its own L1-penalized linear regression or Explainable Boosting Machine. In the evaluation, two-stage k-means with EBM yields the lowest mean absolute error (€80,925) and RMSE (€104,189); the paper reports this as a roughly 36% improvement, meaning the unclustered EBM's MAE is about 36% higher. The same clustering lifts LR from €170,745 to €107,996, which the paper reports as about 58%. The paper also shows that EBM feature-effect curves for construction year and living space differ in sign and shape across sub-clusters, which supports the interpretation that these are genuinely distinct sub-markets rather than arbitrary partitions.

Load-bearing premise

The reported error gains assume that using the price target to form the first-stage clusters does not pass information about the evaluation-period prices into the clustering step.

Editorial extensions

If this is right

  • With two-stage k-means in both phases, EBM mean absolute error drops from €110,387 to €80,925 and LR from €170,745 to €107,996 on the 2023 German house-sale sample.
  • Location-only clustering accounts for the bulk of the gain (EBM drops to €87,276), so the second feature-based stage adds a smaller but measurable improvement; location is the dominant price driver.
  • Clustered linear regression (€107,996 MAE) slightly edges out the unclustered EBM (€110,387 MAE), giving users who need visible coefficients a competitive option.
  • EBM shape plots show construction year and living space effects that differ in sign and shape between clusters, so global models would conceal or average away these segment-specific price drivers.

Reading between the lines

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

  • A replication that constructs the first-stage clusters only from training-fold data would separate the benefit of segmentation from the benefit of having already seen the target prices; the paper does not run this decomposition.
  • The same two-stage recipe is a natural candidate for other spatially heterogeneous pricing tasks, such as apartment rentals or commercial property, but the paper tests it only on single-family house sale listings.
  • Because the final clusters are interpretable sub-markets, the outputs could be used for local price indices or market monitoring, not only for individual property valuations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage clustering approach for house price prediction on 43,309 German house listings. In the first stage, properties are clustered by location (k-means, k=2); in the second stage, each location cluster is further split into eight sub-clusters using either k-means or a regression tree. Cluster-specific linear regression (LR) and Explainable Boosting Machine (EBM) models are then trained and evaluated with k-fold cross-validation. The authors report that two-stage k-means clustering improves MAE by about 36% for EBM and 58% for LR compared with global (unclustered) models, and they present cluster-specific EBM feature-effect plots for construction year and living space.

Significance. If the reported gains were unbiased, the paper would provide a practically valuable result: transparent models (LR and EBM) could achieve markedly better predictive performance through a relatively simple segmentation strategy, while also yielding interpretable, location-specific price-driver insights. The paper also ships public code and an online appendix, which supports reproducibility. However, the central performance claim is currently undermined by two concrete issues: target information appears to enter the clustering and feature-selection stages before cross-validation, and the reported percentage improvements do not match the numbers in Table 1. These issues must be resolved before the quantitative contribution can be accepted.

major comments (3)
  1. [Section 4 and Section 3.3] The evaluation pipeline appears to leak target information into the training process. Section 4 states that the first clustering stage uses latitude and longitude 'combined with relying on information from the price target,' and Section 3.2 says the EBM was used on the full dataset to select the 18 features. Section 3.3 only states that 'k-fold cross-validation was employed' without specifying that clustering and feature selection are nested inside each training fold. If clusters or the feature set are derived from the full data before the CV split, the held-out folds' prices influence the cluster definitions and feature choices, so the reported MAE values are optimistically biased. To support the central claim, the entire pipeline—feature selection, first-stage k-means, and second-stage clustering—must be refit inside each training fold, with the test fold used only for evaluation.
  2. [Section 5.1 and Table 1] The claimed improvements of approximately 36% and 58% do not match Table 1. For EBM, (110,387 - 80,925) / 110,387 = 26.7%, not about 36%. For LR, (170,745 - 107,996) / 170,745 = 36.8%, not about 58%. The reported percentages appear to use the clustered MAE as the denominator (for example, 80,925 / 110,387 would be about 73%, and 107,996 / 170,745 would be about 63%, neither of which yields the stated figures either). The abstract, Section 5.1, and Section 6.1 all repeat the 36%/58% claim, so this is not a one-off typo. The authors should report the correct relative improvements or explicitly state the denominator used.
  3. [Section 3.3 and Section 5.1] Even if the leakage concern is resolved, the paper does not report the variance of the cross-validated MAE or any statistical comparison between the clustered and global models. With a single MAE per approach and no standard deviation or confidence interval, it is unclear whether the clustered-versus-global differences are meaningful relative to the noise of the evaluation. Reporting fold-level errors or confidence intervals would strengthen the empirical claim.
minor comments (5)
  1. [Section 4] The text says 'K-means and KNN identify k centroids,' but KNN is a nearest-neighbor method, not a centroid-based clustering algorithm. This wording is imprecise and should be corrected.
  2. [Section 5.1] The text contains 'num80 925' and '104 189' with an extra 'num' prefix: 'two-stage clustering with k-means (num80 925 MAE)'. This appears to be a formatting artifact and should be fixed.
  3. [References] The reference for Hui & Cheung (2009) contains the typo 'newtwork' in the title; it should read 'network'. Also, the reference for Schaffner & Thiel (2024) has a duplicated URL fragment and inconsistent formatting.
  4. [Abstract and Section 2] The abstract and several places in the text write 'German' as 'german' (e.g., 'german house property listings'). This should be capitalized consistently.
  5. [Section 3.2] The description of feature selection says 'The EBM model was leveraged to identify and retain the most important features,' but it does not specify the threshold or criterion used to decide the final set of 18 features. A brief explanation would improve reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported performance gains are not trustworthy as stated: clustering and feature selection use the price target before the k-fold split, so test prices can leak into the pipeline; the 36%/58% numbers also do not match Table 1.

  1. fitted input called prediction [Section 4 (Method, first-stage clustering) and Section 3.3 (Evaluation and Validation)]
    "After that, the data is grouped into clusters. ... Subsequently, the cluster-level models undergo a training phase respectively. ... In the first stage, mainly latitude, and longitude are used to group geographically proximate houses with similar price levels combined with relying on information from the price target."

    The first-stage clusters are defined using the prediction target (price), and the paper describes clustering as occurring before model training. Section 3.3 only states that 'k-fold cross-validation was employed' and does not say that clustering is repeated inside each fold. If the clusters are built once on all 43,309 listings before splitting, then prices from the validation folds help determine the cluster structure used for both training and evaluation. The reported MAE gain therefore is not an unbiased estimate: the 'prediction' of a held-out property can depend, through cluster membership, on the very price that is being predicted.

  2. fitted input called prediction [Section 3.2 (Data Preprocessing, feature selection) and Section 3.3 (Evaluation and Validation)]
    "The EBM model was leveraged to identify and retain the most important features influencing the property price (Lou et al. 2013). We resulted in a final set of 18 features. ... K-fold cross-validation was employed to ensure the robustness of the models and prevent biases due to data splitting."

    Feature selection is performed by fitting a supervised EBM on the full dataset before the cross-validation split. Because the EBM's feature importance is derived from the price target, the selected 18 features can encode information from the validation folds. The subsequent cross-validated MAE for the clustered models is then optimistically biased, since the held-out prices have already influenced which features enter the model. This is a procedural leak: the cross-validated performance is not independent of the target values it claims to predict.

full rationale

The central claim—that two-stage clustering improves MAE by about 36% for EBM and about 58% for LR (Section 5.1 and abstract)—rests on an evaluation pipeline in which target information can enter model construction. Section 4 explicitly says the first clustering stage uses 'information from the price target,' and Section 3.2 says the EBM selected the final feature set on the full data. Section 3.3 only mentions k-fold cross-validation for the models and does not nest clustering or feature selection within folds. If these steps are done before splitting, the validation prices influence cluster definitions and feature selection, so the reported gains are not an unbiased estimate of predictive performance. This is the main circularity/leakage issue and justifies a score of 6. The self-citations in the paper (e.g., Hambauer et al. 2025, Kruschel et al. 2025) are not load-bearing here: they support general statements about clustering and interpretable ML, while the empirical derivation is self-contained with a public dataset. Additionally, the paper's own heading numbers are internally inconsistent with Table 1: the EBM improvement is (110,387-80,925)/110,387 = 26.7%, not approximately 36%, and the LR improvement is (170,745-107,996)/170,745 = 36.8%, not approximately 58%. That is a correctness issue rather than circularity, but it further weakens the abstract and Section 5.1 claims. Overall, the principal contribution is not circular by definition, but the target-driven preprocessing before cross-validation makes the headline performance comparison substantially less independent than claimed.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The cluster counts, tree depth, and preprocessing cutoffs are selected on the data under study, and the first-stage clustering directly consumes the price target. The evaluation therefore depends on modeling choices that are not independently constrained, plus an unstated leakage-free cross-validation assumption.

free parameters (6)
  • First-stage cluster count k = 2
    Chosen because k-means 'revealed the best results for k = 2' and to keep the number of subclusters interpretable (Section 4).
  • Second-stage cluster count k = 8
    Set to match a regression tree depth of three and validated with the Elbow method (Section 4).
  • Regression tree depth = 3
    Depth two produced nearly indistinguishable clusters; depth four produced too few observations per leaf (Section 4).
  • Duplicate tolerance = 5% difference in price and area
    Removes re-published listings that differ by up to 5% in these features (Section 3.2).
  • Plot size cutoff = 30 square meters
    Properties with plot size below 30 m2 are removed (Section 3.2).
  • Missingness threshold = 90%
    Features with more than 90% missing values are removed (Section 3.2).
assumptions (4)
  • domain assumption Listed asking prices from ImmobilienScout24 approximate transaction prices.
    The outcome variable is the listing price; Section 6.2 acknowledges that listed prices may not reflect actual transaction prices or may include unsold overpriced properties.
  • domain assumption Cluster-then-predict segmentation produces groups within which linear regression and EBM are appropriate.
    The core premise of Sections 2 and 4 is that location and property clusters capture homogeneous market segments with stable feature effects.
  • domain assumption Postal-code-derived latitude and longitude sufficiently represent location for clustering.
    Section 3.2 converts postal codes to coordinates and uses them as the primary location signal, without validation against finer-grained geodata.
  • ad hoc to paper Clustering on the full data before cross-validation introduces no target leakage.
    Section 3.3 invokes k-fold cross-validation, while Section 4 says the first-stage clustering uses the price target; no nested procedure is described, so the evaluation implicitly assumes that the cluster assignments do not leak validation labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unveiling Location-Specific Price Drivers: A Two-Stage Cluster Analysis for Interpretable House Price Predictions." pith.science (2026). https://pith.science/paper/NFRZK76W

@misc{pith2026250803156,
  author       = {Pith},
  title        = {Pith review of: Unveiling Location-Specific Price Drivers: A Two-Stage Cluster Analysis for Interpretable House Price Predictions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NFRZK76W}},
  note         = {Machine review of arXiv:2508.03156}
}
read the original abstract

House price valuation remains challenging due to localized market variations. Existing approaches often rely on black-box machine learning models, which lack interpretability, or simplistic methods like linear regression (LR), which fail to capture market heterogeneity. To address this, we propose a machine learning approach that applies two-stage clustering, first grouping properties based on minimal location-based features before incorporating additional features. Each cluster is then modeled using either LR or a generalized additive model (GAM), balancing predictive performance with interpretability. Constructing and evaluating our models on 43,309 German house property listings from 2023, we achieve a 36% improvement for the GAM and 58% for LR in mean absolute error compared to models without clustering. Additionally, graphical analyses unveil pattern shifts between clusters. These findings emphasize the importance of cluster-specific insights, enhancing interpretability and offering practical value for buyers, sellers, and real estate analysts seeking more reliable property valuations.

Figures

Figures reproduced from arXiv: 2508.03156 by the authors.

Figure 1
Figure 1. Results of the location clusters for different algorithms 5.3 Assessment of Cluster-specific Price Effects The application of location-based clustering effectively reduced the dominant influence of location in the subsequent sub-clustering step, allowing other property characteristics to become more prominent in the modeling phase using LR and EBM. An analysis of the feature relationships learned by the EBM model re… view at source ↗
Figure 2
Figure 2. Feature effect of construction year on the price using the EBM model based on two distinct clusters respectively. increases slowly, but almost linearly in relation to living space. This could indicate that in sub-cluster 2, demand for very large properties tapers off, whereas in sub-cluster 5, demand remains stable even for larger properties. These varying feature-price relationships across sub-clusters highlight th… view at source ↗
Figure 3
Figure 3. Feature effect of living space on the price using the EBM model based on two distinct clusters respectively. Third, we provide cluster-specific insights into price drivers, revealing variations in the influence of factors such as living space and construction year. This enhances model transparency and supports better decision-making for real estate professionals by providing actionable insights. Finally, our work co… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

  1. [1]

    O’Reilly Media, Inc

    Alfaro-Navarro, J.-L., Cano, E. L., Alfaro-Cortés, E., García, N., Gámez, M. & Larraz, B. (2020), ‘A fully automated adjustment of ensemble methods in machine learning for modeling complex real estate systems’, Complexity 2020(1), 5287263. Azimlu, F., Rahnamayan, S. & Makrehchi, M. (2021), House price prediction using clustering and genetic programming al...

    arXiv 2020

  2. [2]

    (2019), ‘Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead’,Nature machine intelligence1(5), 206–

    Rudin, C. (2019), ‘Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead’,Nature machine intelligence1(5), 206–

    work page 2019

  3. [4]

    (1996), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288

    Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288. Wang, P.-Y ., Chen, C.-T., Su, J.-W., Wang, T.-Y . & Huang, S.-H. (2021), ‘Deep learning model for house price prediction using heterogeneous data analysis along with joint self-attention mec...

    work page 1996

  4. [215]

    & Thiel, P

    Schaffner, S. & Thiel, P. (2024), Fdz data description: Real-estate data for germany campus files (rwi-geo-red panel and rwi-geo-red cross v5)-advertisements on the internet platform immobilienscout24 for teaching purposes, Technical report, RWI Datenbeschreibung. Shi, D., Zhang, H., Guan, J., Zurada, J., Chen, Z. & Li, X. (2023), Deep learning in predict...

    work page 2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.