Transfer Learning Across Local Housing Markets: A Research Protocol Integrating Image Fusion, Interpretable Drivers, Spillover Networks, and Mixed-Frequency Signals

1. Introduction

Automated valuation models rest on the premise that observable features can proxy the mechanisms through which housing is priced. Prior valuation research shows that these mechanisms include both macro-level factors, such as inflation and interest conditions, and micro-level factors, such as structural housing attributes and locational context. The same literature also reports that machine-learning methods are often attractive in housing applications because they can accommodate nonlinear relations among features, with tree-based models frequently performing strongly in practice. [1]

The transfer problem begins when a model estimated in one locality is applied to another. A local housing market is not simply a smaller version of a national market. High-resolution urban data from an emerging city illustrate that analysts may need variables on property characteristics, accessibility, environmental quality, public amenities, and demographic context at the residential-community level. Intercity network analysis further indicates that housing prices do not evolve as isolated local series; directed spillovers connect cities in structured ways. Aggregate forecasting work also shows that macro and financial indicators observed at different frequencies may be informative about housing conditions, while warning that predictor usefulness and index quality vary across settings. [2] [3] [4]

This manuscript addresses a narrower methodological question than house-price prediction in general. It asks how transfer learning should be designed when the target is a local housing market and when available data may combine tabular housing attributes, geospatial variables, images, and macro context. The answer proposed here is a research protocol rather than a report of completed tests. Its methodological inspiration comes from cross-domain sentiment analysis, where domain shift impairs generalization and adversarial domain adaptation is used to learn representations that are less domain-specific. Housing markets pose an analogous but not identical problem: the same observed feature can carry different valuation meanings across cities because of regulation, urban form, market maturity, or spillover exposure. [5]

2. Research question and evidence scope

The central research question is: under what conditions, and through which model design, can information learned from one or more source housing markets improve prediction in a different target market without obscuring local price-formation mechanisms? The protocol treats improvement as an empirical question to be tested later. It does not presume that transfer is beneficial. A core concern is negative transfer, meaning that source-market patterns appear statistically useful but are economically misleading for the target city.

The evidence scope is a targeted narrative synthesis, not a systematic review. Five strands are included because each bears directly on a component of cross-market transfer. First, a Hong Kong valuation study shows that exterior housing photos, street-view images, and remote-sensing images can be fused with tabular data and that image-derived features can materially affect predictive models. Second, a Nanning dataset demonstrates a fine-grained residential-community unit of analysis and a hedonic-style feature system interpreted with SHAP. Third, a spatiotemporal network study models directed spillovers among cities and reports policy-related structural reconfiguration. Fourth, an Indian forecasting study applies U-MIDAS to mixed-frequency predictors for a Housing Price Index and notes limitations in data quality and external applicability. Fifth, a sentiment-analysis study provides a transferable methodological template for adversarial domain adaptation with a gradient reversal layer. [1] [2] [3] [4] [5]

These studies are informative but heterogeneous in geography, target variable, unit of analysis, and purpose. Accordingly, this manuscript does not treat them as direct evidence that transfer learning improves local housing valuation. Rather, it uses them to motivate and bound design choices. Image fusion is supported as a feasible information channel, not as established transfer leverage. Spillover modeling supports the relevance of intercity dependence, not yet its predictive value for property-level transfer. Mixed-frequency forecasting supports explicit handling of temporal frequency mismatch, not the superiority of any specific macro predictor set for local transfer. [1] [3] [4]

3. What the prior work establishes and what remains open

The Hong Kong valuation study establishes that built and natural environment information can be operationalized from multiple image sources and fused with structured variables inside an automated valuation pipeline. It also reports that feature-configuration choices affect model performance and that image-derived variables can rank highly in interpretability analysis. For the present protocol, these findings justify treating image features as a serious candidate modality and caution against assuming that one preprocessing configuration will transfer unchanged across markets. [1]

The Nanning dataset makes a different contribution. It defines the residential community as the basic unit of analysis, compiles variables on housing attributes, accessibility, amenities, environmental quality, land use, and population context, and uses Random Forest with SHAP for interpretability. This supports the proposition that local housing markets, including emerging cities, may require granular feature systems rather than coarse city-level averages. It also shows that interpretable machine learning is feasible in a second-tier or emerging-city setting. [2]

The spatiotemporal network study shows that housing-price spillovers can be directed, hierarchical, and heterogeneous across regional sub-networks. It also reports a structural break associated with the 2016 policy directive that reconfigured transmission channels and produced a new network equilibrium. For transfer learning, the immediate implication is not that spillover metrics will necessarily improve local valuation, but that source-target similarity may depend on policy regime and network position rather than on cross-sectional feature overlap alone. [3]

The Indian Housing Price Index study shows that mixed-frequency predictors can be modeled explicitly through U-MIDAS and that the relevance of predictors must be assessed within a specific institutional environment. The same paper also notes that housing registration quality differs among cities, selected predictors may not fully capture local markets, and model performance may change across economic environments. These observations support caution. In the present protocol, macro and financial signals are proposed as contextual inputs to be tested, not as validated substitutes for local housing microstructure. [4]

The sentiment-analysis study identifies domain divergence as a source of poor cross-domain generalization and uses a domain discriminator with gradient reversal to encourage domain-invariant representations. That architecture was designed for text rather than multimodal spatial-economic data, so it cannot be imported directly. What transfers conceptually is a narrower idea: when source and target distributions differ, adversarial learning may help extract features that retain predictive content while reducing domain-specific overfitting, provided that source-target overlap is sufficient in the first place. [5]

4. Conceptual framework and operational definitions

The protocol represents each local housing market as a domain d, observation as i, and time index as t where applicable. The primary prediction target is y_(i,d,t) = log(price per square meter). The operational rule is exact: if p_(i,d,t) is transaction or listing price and a_(i,d,t) is floor area in square meters, then y_(i,d,t) = log(p_(i,d,t) / a_(i,d,t)). When only community-level data are available, p/a is replaced by the community’s average price per square meter. This target is proposed because housing prices are commonly right-skewed and because multiplicative effects are often easier to interpret on the log scale.

Inputs are divided into four blocks. Block 1 contains structural housing variables such as floor area, age, floor level, room count, property type, and orientation when available. Block 2 contains neighborhood and accessibility variables including distances to schools, hospitals, malls, transit stations, and central areas, together with environmental and demographic indicators. Block 3 contains image-derived variables extracted from exterior housing photos, street-view imagery, and remote-sensing imagery, using either compressed embeddings or interpretable indices such as vegetation, water, built-up, sky-view, and building-view measures. Block 4 contains macro and spillover context, including mixed-frequency economic indicators and city-level network features aligned to the observation period. [1] [2] [4] [3]

The protocol distinguishes transferable structure from local calibration. Transferable structure refers to broad predictive regularities that may recur across cities, such as nonlinear interactions between housing structure and accessibility or the relevance of image-based proxies for environmental quality. Local calibration refers to city-specific adjustments in effect size, thresholds, and residual patterns that may arise from regulation, urban morphology, data generation, or network exposure. This distinction is proposed rather than empirically established. It is the organizing premise of the model design described below.

5. Proposed transfer-learning architecture

The proposed model is a multi-branch architecture with selective parameter sharing. Branch A ingests structural and neighborhood variables. Branch B ingests image-derived features from the three image sources. Branch C ingests macro variables aligned to the local observation window through rolling aggregation or lag expansion. Branch D ingests spillover-network features. The four branch outputs are concatenated into a shared latent vector z_shared. This vector feeds a price-prediction head g(.) and a city discriminator q(.). The discriminator is connected through a gradient reversal layer so that the encoder is trained to produce features that remain useful for price prediction while making domain classification more difficult. This design borrows the logic, not the text-specific architecture, of adversarial adaptation in cross-domain sentiment analysis. [5]

The protocol does not aim for complete domain invariance. Housing prices are partly determined by local factors that should not be erased. For that reason, each city d is assigned a lightweight residual head h_d(.), producing a correction term r_d. The final prediction is y_hat = g(z_shared) + r_d. The residual head uses z_shared together with a small subset of raw local features, such as local price level, centrality in the spillover graph, or selected structural variables, to recover city-specific calibration that an invariant encoder might suppress. This is a proposed design choice intended to balance transfer and local specificity; it is not presented as a validated solution. [2] [3]

Three target-data regimes are distinguished. In a data-rich target, defined hypothetically as sufficient labeled observations to estimate a city-specific calibration curve and residual head without severe instability, the shared encoder is initialized on selected source cities and then the target residual head is trained on a held-out calibration subset. In a few-shot target, where labeled target observations are scarce, only the final layer of h_d and optionally a target-specific bias term are estimated, while the shared encoder and most residual parameters remain frozen or penalized toward zero. In a zero-shot target, no labeled target observations are available; h_d is set to zero and the model reduces to the shared prediction head applied after source selection. This zero-shot mode is treated as a stress scenario rather than the main intended use case.

Because the protocol concerns data-sparse cities, the residual heads require explicit constraints. A proposed rule is to train h_d only when the target calibration subset contains at least m labeled units in each of at least k spatial strata defined by administrative district or residential-community cluster. The values of m and k are protocol parameters to be tuned; for illustration only, one might start with m = 20 and k = 3. If this condition is not met, the target enters few-shot or zero-shot mode. In few-shot mode, the residual head is estimated with strong regularization, such as an L2 penalty or Bayesian shrinkage toward zero, and the number of trainable parameters is capped in advance. This makes the treatment of sparse targets reproducible rather than discretionary.

6. Source-market selection with an explicit transferability score

Source-city selection is treated as a formal stage rather than an implicit researcher choice. For target city T and candidate source city S, define TransferScore(S,T) = w1*COV + w2*MACRO + w3*NET, where w1 + w2 + w3 = 1. Initial weights may be set hypothetically to w1 = 0.4, w2 = 0.3, and w3 = 0.3, but these values are placeholders to be tuned by validation. A candidate source is admitted only if TransferScore(S,T) exceeds a threshold tau chosen prospectively. Any threshold reported in applications should be treated as tuned within validation, not as fixed by this article.

The COV component is defined operationally as follows. Let J be the set of harmonized continuous and ordinal covariates shared by source and target after preprocessing. For each feature j in J, winsorize both cities at common quantiles alpha and 1-alpha computed on the pooled source-target sample; alpha is a protocol parameter, for example 0.01. On the pooled winsorized support of feature j, construct B equal-width bins with edges e_0 < e_1 < ... < e_B. Let p_djb be the proportion of observations from city d falling in bin b, where d is S or T. Define the cumulative bin distribution F_dj(b) = sum_{u=1 to b} p_dju. Then Delta_j = (1/B) * sum_{b=1 to B} |F_Sj(b) - F_Tj(b)|. Because each F_dj(b) lies in [0,1], each absolute difference lies in [0,1], so Delta_j also lies in [0,1]. Define COV = 1 - (1/|J|) * sum_j Delta_j. Therefore COV lies in [0,1], with larger values indicating greater overlap in empirical feature distributions. This replaces the earlier under-specified reference to a standardized empirical cumulative distribution function summary.

The MACRO component is also defined exactly. Let M be the set of macro indicators available for both cities or for the larger economies in which they are embedded. For each indicator m, compute a calibration-window summary vector v_dm containing the mean, standard deviation, and first-order autocorrelation over the same relative time window. Standardize each summary element across all candidate cities using the training pool mean and standard deviation. Let D_macro(S,T) be the average absolute difference between the standardized summary elements of S and T across all indicators. Convert distance to similarity by MACRO = exp(-D_macro), which lies in (0,1]. This transformation avoids arbitrary truncation and preserves comparability across indicators after standardization.

The NET component uses the spillover graph as contextual information. Let G be a directed graph estimated externally from city-level price indices or obtained from a prior network-estimation stage. Define l(S,T) as the directed shortest path length from S to T. If no directed path exists, set l(S,T) = infinity. Define NET = 1/(1 + l(S,T)) when l is finite, and NET = 0 when no path exists. This yields NET in [0,1], with 1 reserved for the degenerate case S = T and smaller values for more distant or disconnected cities. If weighted edges are available, l may be computed on inverse spillover intensity. This graph-based term does not claim that network proximity improves transfer; it is included as a testable hypothesis derived from evidence that intercity housing-price relations are directed and structured. [3]

7. Feature engineering, mixed-frequency context, and spillover features

Feature engineering follows prior housing work in broad substance but is adapted for transfer. Structural features are standardized within city and time block when scale differences are mainly nominal rather than behavioral. Accessibility variables must use one distance convention across all cities, either Euclidean or network distance, to avoid hidden measurement shift. Image-derived features may include compressed embeddings extracted by pretrained vision models and interpretable indices retained separately for explanation. The motivation for retaining interpretable image indices is supported by valuation work showing that variables such as vegetation, sky-view, building-view, and remote-sensing indices can relate nonlinearly to price inside a city; their transfer value across cities, however, remains to be tested. [1]

Mixed-frequency macro variables are introduced cautiously. The Indian HPI study shows that lower-frequency housing outcomes can be forecast using higher-frequency predictors through U-MIDAS, and it also notes that predictor adequacy and external applicability are limited by institutional setting and data quality. In the present protocol, that evidence justifies a design choice: temporal frequency mismatch should be modeled explicitly rather than ignored. It does not establish that digital payment activity, consumer prices, or financial stress will improve local property-level transfer. Therefore, macro variables are optional contextual channels subject to ablation testing. [4]

Spillover variables are likewise framed as hypotheses rather than established transfer gains. Candidate features include out-degree, in-degree, standardized betweenness centrality, net spillover status, and policy-regime indicators marking known structural breaks. The network study reports radiation hubs, chokepoints, and a policy-induced reconfiguration after 2016, which supports the relevance of intercity context. It does not directly show that such variables improve local automated valuation transfer. In this protocol they serve two roles to be tested prospectively: source-selection context and an auxiliary input block that may help explain residual city-level variation. [3]

8. Comparator models, training decisions, and outcome measures

The primary comparator is a target-only model trained on the same target-city labels without transfer. Transfer is defensible only if it improves performance, improves data efficiency, or preserves calibration better than this local benchmark. Secondary comparators are a pooled multi-city model without adaptation, a source-only model applied directly to the target, and a non-adversarial multi-task model with a shared encoder and city-specific heads but no domain discriminator. If image data are available, an additional comparator without image features is recommended to isolate the contribution of multimodality.

Training proceeds in stages. Stage 1 harmonizes variables and computes TransferScore for source selection. Stage 2 pretrains the shared encoder and price head on admitted source cities, using the domain discriminator only after the prediction loss has reached a stable range on source validation data. Stage 3 adapts to the target according to data regime: full residual-head training in data-rich mode, restricted last-layer or bias adaptation in few-shot mode, and no target-head training in zero-shot mode. Stage 4 compares performance against all comparator models on a temporally later and geographically distinct target test set. These stages are proposed to make the role of target labels explicit in a setting where sparse supervision is central.

Outcome measures should match intended use. For valuation accuracy, the protocol recommends root mean squared error on log price per square meter, mean absolute error, and median absolute percentage error on the original price-per-square-meter scale. Calibration should also be assessed across property type, price decile, and central versus peripheral areas. For ranking uses, Spearman correlation may be reported. A transfer-specific metric is defined exactly as TransferGain = Error(target-only) - Error(transfer model). Positive values indicate beneficial transfer on the chosen error metric; negative values indicate negative transfer.

The discriminator should not be governed by an arbitrary fixed AUC cutoff. Instead, its role is diagnostic and tied to the prediction objective. The protocol proposes the following placeholder rule for later tuning: select the adversarial weight lambda that minimizes target validation error among candidate lambda values, subject to the constraint that discriminator performance after training is not substantially above the empirical distribution obtained from label-permuted city identities under the same architecture. In practice, one may compute a null distribution of discriminator AUC under permutation and treat persistent separation beyond an upper quantile as evidence that the encoder remains strongly domain-revealing. This is a tunable diagnostic, not a validated universal decision rule.

9. Prospective validation protocol

Because this article reports no new experiments, the evaluation plan must be prospective and falsifiable. The protocol recommends multiple transfer scenarios. Scenario A concerns within-country transfer between cities with high covariate overlap and strong network connectivity. Scenario B concerns transfer across different spillover sub-networks or across policy periods separated by a structural break. Scenario C concerns emerging-city transfer in which the target has limited labeled data but usable GIS and image coverage. Scenario D is a stress test near the source-selection threshold, designed to examine negative transfer when similarity is weak. These scenarios reflect distinct hypotheses rather than presumed success cases. [3] [2]

Validation should include three orthogonal holdouts. First, temporal holdout trains on earlier periods and tests on later periods to assess robustness to market evolution. Second, spatial holdout withholds neighborhoods or residential communities within the target city to assess geographic extrapolation. Third, modality ablation removes image features, spillover features, and macro features in turn. This structure is justified by prior evidence that image features can matter in city-level valuation, that intercity spillovers are structured, and that mixed-frequency predictors can be informative in aggregate forecasting. The actual contribution of each modality to transfer remains an empirical question. [1] [3] [4]

Interpretability validation is also required. Global and local attribution methods, comparable in spirit to SHAP, should be computed for the final transferred model. A model that lowers error but relies on substantively implausible features in the target city should be treated with caution. Attribution profiles should be examined for sign reversals and abrupt neighborhood-to-neighborhood instability. The Nanning dataset and the Hong Kong image-fusion study support the practical use of interpretability tools in housing settings; this protocol extends that logic to transfer diagnostics rather than asserting that attributions guarantee causal validity. [2] [1]

Robustness checks should cover subgroup error and missing-modality conditions. Errors should be compared across central and peripheral districts, dense and sparse communities, and areas with high versus low image coverage. Missing-modality tests should mask one image source or one macro block at evaluation time. A transferred model that performs acceptably only when all modalities are complete may be impractical for cities with uneven data infrastructure.

10. Failure cases, boundary conditions, and limits of generalization

Negative transfer is the central failure case. It may occur when an image cue has different social meaning across cities, when accessibility interacts differently with transport systems, or when a policy change alters transmission channels. The network study’s report of a structural break after the 2016 policy directive is an important reminder that transfer across time can fail even within one national system. Accordingly, this protocol treats regime mismatch as a substantive risk, not a nuisance detail. [3]

Data-generation differences are another limit. The Nanning dataset combines multiple listing platforms and applies rule-based deduplication. Other cities may differ in platform penetration, geocoding quality, or variable definitions. The Indian HPI study also notes that housing registration quality varies among cities and that selected predictors may not fully capture local conditions. Transfer learning can amplify these differences if preprocessing harmonization is weak. Variable name agreement should therefore not be mistaken for measurement equivalence. [2] [4]

A further boundary condition concerns modality imbalance. The valuation literature supports the relevance of image features within a city, but image availability, season, camera viewpoint, and spatial coverage can vary sharply across localities. Apparent transfer failure may therefore reflect missing or noisy modalities rather than the invalidity of transfer learning itself. This is why the protocol includes ablation and masking tests rather than interpreting aggregate error alone. [1]

There is also an unresolved tension between domain invariance and local interpretability. Adversarial adaptation tries to suppress domain-specific information, yet some domain-specific information is economically meaningful in housing markets. The protocol’s residual heads and source-selection stage are proposed mechanisms for managing that tension, but they do not resolve it in principle. Their adequacy remains to be tested. Generalization beyond the markets eventually studied would also remain limited, because the motivating evidence comes from specific Chinese and Indian settings and from a methodological analogy to text classification rather than from completed cross-market housing-transfer experiments. [5] [1] [2] [3] [4]

11. Conclusion

This manuscript has proposed, rather than validated, a research protocol for transfer learning across local housing markets. The protocol integrates four candidate information layers: micro-level structural and neighborhood covariates, image-derived environmental signals, macro indicators observed at mixed frequencies, and intercity spillover context. It also proposes that cross-market transfer should be selective rather than automatic, using explicit source-city screening and a model structure that separates shared representation from city-specific calibration.

The methodological contribution is explicit and falsifiable. The article defines the prediction target as log price per square meter, specifies a multi-branch architecture with adversarial adaptation and residual heads, operationalizes a reproducible source-selection score, distinguishes data-rich, few-shot, and zero-shot target regimes, and sets out comparator models, outcome measures, and a prospective validation design. It also states failure conditions: negative transfer, persistent domain separability after adaptation, unstable attributions, subgroup calibration loss, and dependence on unavailable modalities.

What remains untested is substantial. No empirical evidence is offered here that the proposed protocol improves valuation accuracy, reduces label requirements in sparse markets, or preserves fairness better than target-only models. The usefulness of spillover and mixed-frequency features for local transfer, the best weighting of the transferability score, and the proper balance between domain invariance and local specificity all require future empirical study. The present contribution is therefore bounded: it establishes a rigorous framework for evaluating cross-market transfer in housing, not a demonstrated performance gain.

References

1. Deng, Lin (2025). Real estate valuation with multi-source image fusion and enhanced machine learning pipeline. PLOS One. 20(5), e0321951. doi:10.1371/journal.pone.0321951.
2. Lin, Wenbo; Wu, Kaichuang; Zhang, Xiaolu; Gu, Xiaoling; Jiang, Ming; Wang, Jun; Han, Feng (2026). A comprehensive dataset for uncovering housing price drivers in emerging cities: Nanning as a case study. Data in Brief. 66, 112884. doi:10.1016/j.dib.2026.112884.
3. Qiu, Lu; Jiao, Yanzhe; Dong, Gege; Cui, Guangcan (2026). Finite-Length Spatiotemporal Modelling for Housing Price Network Spillovers. Entropy. 28(5), 537. doi:10.3390/e28050537.
4. Singh, Nutan; Shanmugam, Kavitha (2026). Exploring predictive models to improve the accuracy of Housing Price Index forecasts in India’s real estate sector. PLOS One. 21(1), e0341026. doi:10.1371/journal.pone.0341026.
5. Kumar, Ajeet; Abhishek, Kumar; Shafeeq B M, Ahamed (2025). SentXFormer: a transformer-enhanced hybrid deep learning framework for cross-domain sentiment analysis of customer reviews. Scientific Reports. 16(1), 3528. doi:10.1038/s41598-025-33526-1.