Retrieval-Enhanced Real Estate Appraisal

Simon Popelier; Matthieu X. B. Sarazin; Maximilien Bohm; Mathieu Gierski; Hanna Mergui; Matthieu Ospici; Adrien Bernhardt

cs.LGMar 13, 2026

Retrieval-Enhanced Real Estate Appraisal

Authors: Simon Popelier, Matthieu X. B. Sarazin, Maximilien Bohm, Mathieu Gierski, Hanna Mergui, Matthieu Ospici, Adrien Bernhardt

Organizations: Homiwoo, 10 rue de la Boétie, 75008 Paris, France · École Polytechnique, Route de Saclay, 91128 Palaiseau, France

Abstract

The Sales Comparison Approach (SCA) is one of the most popular when it comes to real estate appraisal. Used as a reference in real estate expertise and as one of the major types of Automatic Valuation Models (AVM), it recently gained popularity within machine learning methods. The performance of models able to use data represented as sets and graphs made it possible to adapt this methodology efficiently, yielding substantial results. SCA relies on taking past transactions (comparables) as references, selected according to their similarity with the target property's sale. In this study, we focus on the selection of these comparables for real estate appraisal. We demonstrate that the selection of comparables used in many state-of-the-art algorithms can be significantly improved by learning a selection policy instead of imposing it. Our method relies on a hybrid vector-geographical retrieval module capable of adapting to different datasets and optimized jointly with an estimation module. We further show that the use of carefully selected comparables makes it possible to build models that require fewer comparables and fewer parameters with performance close to state-of-the-art models. All our evaluations are made on five datasets which span areas in the United States, Brazil, and France.

Figures & tables

Figure 1: Model architecture. Red rows represent the target property whose price is estimated; blue rows represent comparables picked via vector retrieval; yellow rows represent comparables picked via geographic retrieval; bi-color rows represent features relative to the target; grey shaded rectangles and dashed lines are the modules added to REA to form EREA.

Region	Attr.	Samples	Value (€/ $m^{2}$ , $, R$ )	Dist. (m)
IV	22	45,554	2,613 $\pm$ 1,278	353 $\pm$ 628
KC	19	21,608	540,098 $\pm$ 367,156	604 $\pm$ 940
FC	12	83,067	155,288 $\pm$ 76,420	98 $\pm$ 148
POA	8	15,368	443,798 $\pm$ 228,518	202 $\pm$ 268
SP	8	68,848	741,952 $\pm$ 411,643	202 $\pm$ 246

Table 1: Dataset statistics. Values and distances are calculated w.r.t the 40 geographically closest comparables.

Figure 2: The temporal split of the IV dataset into offset, train, validation, test.

Figure 3: Intrinsic retrieval evaluation with REA. Geographic retrieval (yellow), hybrid retrieval (green), and vector retrieval (blue), are compared with different number of comparables. The average and 95% CI of 20 simulations are shown here.

Model	IV		KC		FC		POA		SP
	MdAE	MdABRE	MdAE	MdABRE	MdAE	MdABRE	MdAE	MdABRE	MdAE	MdABRE
LR	574	21.21	68568	16.96	20259	16.66	82831	25.57	136263	25.70
kNN	467	19.24	68112	16.83	12255	9.55	40000	11.18	108381	19.73
XGB	408	16.08	39485	9.26	9990	7.72	44548	12.41	69212	11.93
ASI	357.16*	13.24*	36559.8	8.37	8608.5	6.55	34281.1	9.57	58194.7	10.0
ASI	$\pm$ 2.75	$\pm$ 0.09	$\pm$ 224.3	$\pm$ 0.05	$\pm$ 40.4	$\pm$ 0.03	$\pm$ 408.2	$\pm$ 0.1	$\pm$ 1496.1	$\pm$ 0.24

Table 2: Baseline comparison table on all 5 datasets. MdAE is in €/

m^{2}

for IV,

for KC/FC and R

for POA/SP, MdABRE is in %. The average and 95% CI of 20 simulations are shown here for the last 3 models. ASI was adapted into PyTorch for the IV dataset and might differ slightly, hence the asterisk.

Model	IV	KC	FC	POA	SP
ASI	11,660	21,700	17,140	19,240	3,565
REA	900	800	660	580	580
EREA	1,472	1,142	792	592	592

Table 3: Number of parameters per model and dataset.

Hybrid Retrieval

Hybrid retrieval combines lexical, dense, and structured search signals to find relevant evidence more reliably.

See all 91 papers

RowNet: A Memory Transformer for Tabular Regression
Jun 3, 2026Askat Rakhymbekov, Gulshat MuhametjanovaMemory-Augmented Neural Networks Tabular Regression
Department of Applied Mathematics and Informatics Kyrgyz-Turkish Manas University Bishkek, Kyrgyzstan · Department of Applied Mathematics and Informatics2026 Kyrgyz-Turkish Manas University Bishkek, KyrgyzstanJun
Benchmarking on Tasks That Matter: Dataset Selection for Preserving Model Rankings
Jun 26, 2026Rostislav Gusev, Alexey ZaytsevBenchmark Design
Applied AI Institute Moscow, Moscow region, Russian Federation
ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering
Apr 20, 2026Yindong Zhang, Wenmian Yang, Yiquan Zhang +1LLM Agent Orchestration AI Agent Benchmarks
Hong Kong Baptist University · Beijing Normal-Hong Kong Baptist University · Beijing Normal University

Jun 3, 2026J/K move · Enter open · S save

RowNet: A Memory Transformer for Tabular Regression

Askat Rakhymbekov, Gulshat Muhametjanova

Department of Applied Mathematics and Informatics Kyrgyz-Turkish Manas University Bishkek, Kyrgyzstan · Department of Applied Mathematics and Informatics2026 Kyrgyz-Turkish Manas University Bishkek, KyrgyzstanJun

Memory-Augmented Neural Networks Tabular Regression

Open paper AskPDF arXiv

Real estate valuation is a structured regression problem in which prices are governed by heterogeneous feature types, sparse regional effects, nonlinear interactions, and the practical logic of comparable properties. Standard multilayer perceptrons treat each row as an isolated vector and must learn locality, scale sensitivity, and categorical matching from supervision alone. Gradient-boosted decision trees provide strong tabular baselines, but their feature-centric splitting mechanism does not explicitly model the retrieval of similar historical observations. This paper presents RowNet, a retrieval-based neural architecture for real estate price-per-square-meter prediction. RowNet represents a query property through pairwise similarity features against a memory bank of labeled properties. A first retrieval layer estimates a coarse target from feature-only similarities. A second layer augments the memory comparison with target-consistency features and uses multiple learned attention heads to retrieve complementary comparable sets. A final mixture-of-experts module combines learned gating, residual correction, entropy regularization, and head-diversity regularization to produce the prediction.

Abstract

Hybrid Retrieval

Explore similar work

RowNet: A Memory Transformer for Tabular Regression

Benchmarking on Tasks That Matter: Dataset Selection for Preserving Model Rankings

ReCoQA: A Benchmark for Tool-Augmented and Multi-Step Reasoning in Real Estate Question and Answering