Fairness Audits for Rental Recommendation Systems: A Conceptual Audit Framework Informed by Real-Estate Outcomes and Transferable Recommender Methods

1. Introduction

Rental recommendation systems affect housing access by shaping which dwellings are seen, in what order, and under what explanatory framing. In this domain, ranking is part of opportunity allocation rather than a minor convenience feature. A user who never sees a suitable listing cannot meaningfully benefit from downstream application procedures, and a listing that is repeatedly buried may receive little attention regardless of its suitability. Fairness auditing is therefore needed at the recommendation stage itself, not only at later stages of screening or contracting.

The technical problem is difficult because rental recommendation systems are heterogeneous. A platform may combine click and save histories, map interactions, listing metadata, free-text reviews, amenity tags, neighborhood relations, and sequential engagement signals. It may also operate under strong privacy limits or with many new users and newly posted listings. Methods from retail and media recommendation are therefore relevant, but they cannot be imported unchanged into housing without attention to housing-specific harms and constraints.

Direct real-estate evidence supports caution in interpreting downstream market outcomes. In the Dutch office market, buildings meeting the label-C threshold experienced a positive sales-price effect after policy adoption, while rental rates remained largely unchanged. The study discusses differences in incentives and lease structure as plausible explanations for this asymmetry rather than as a proven causal mechanism. For rental recommendation auditing, this article uses that result only as an analogy about channel asymmetry, not as evidence about residential platform behavior. The narrower point is that flat posted rents do not by themselves establish fair platform behavior. Visibility, search burden, and access to qualified options may shift before rent data show any aggregate response. [1]

2. Research question and evidence scope

The research question is: how should a fairness audit for rental recommendation systems be designed so that it detects disparate exposure, ranking, relevance, and explanation quality without treating conventional recommendation accuracy as a sufficient proxy for recommendation-stage housing access? In this article, recommendation-stage housing access means access to qualified affordable listings through the platform’s candidate generation, ranking, and explanation process, not final lease outcomes.

The evidence scope is a targeted narrative synthesis, not a systematic review. The selected studies were chosen because each offers one technical idea that can be translated into a housing audit problem. The real-estate market study provides a direct example of divergence between sales and rental responses in a related property market; in this manuscript that finding is used only as a bounded analogy warning against reliance on rent movement as a fairness proxy. EX-KAN contributes sensitivity-based explanation and regime-aware analysis. CRAFT addresses cold-start recommendation under federated privacy constraints. KGRec shows how higher-order relations among users, items, and attributes can shape sparse recommendation settings. DeepSentRec demonstrates that review sentiment, hybrid filtering, and adaptive ranking can materially alter recommendation outputs. A supplementary house-price prediction study is included because it warns that normalized error metrics are not directly interpretable in real-world units and suggests adding explainability tools; the present article extends that caution into an audit-design requirement. [1] [2] [3] [4] [5] [6]

This bounded evidence base does not prove how rental platforms currently behave. Its purpose is narrower: to construct an auditable protocol with explicit variables, rules, comparators, assumptions, and prospective validation steps. Established findings from the cited studies are distinguished from the present article’s proposals, which remain to be tested empirically.

3. Transferable foundations from prior work

The direct real-estate study is relevant because it separates market channels. It reports a sizable capitalization effect in office sales but negligible rental-price effects after the energy-performance intervention. In this manuscript, that pattern is used only as a conceptual warning that rental-price change can be a weak proxy for access conditions closer to the platform’s own actions. The transfer to residential rental recommendation is therefore analogical and limited; it does not establish how residential platforms behave or how recommendation disparities would map into market rents. A rental recommender audit should accordingly focus on exposure and qualification outcomes measured at the recommendation stage. [1]

EX-KAN addresses the tension between predictive performance and interpretability in real-estate forecasting. It uses lagged price and volume inputs and derives elasticity-style diagnostics from model Jacobians, while also distinguishing high, normal, and low market regimes. The transferable lesson is methodological rather than architectural: auditing should examine local score sensitivity to key variables and should do so within defined market regimes rather than only through global averages. [2]

CRAFT frames cold-start as a central recommender problem and describes federated training as a privacy-preserving alternative to centralized data aggregation. It notes that sparse or absent histories remain a major difficulty in both centralized and federated settings. In rental platforms, new users and new listings are common rather than exceptional, so a fairness audit must report them separately. If the production system is federated, the audit must also be feasible when raw histories cannot be pooled centrally. [3]

KGRec argues that collaborative filtering often treats interactions too independently and therefore misses richer semantic structure among users, items, and attributes. Knowledge graphs can encode indirect relations, but those same relations can transmit socially sensitive proxies. In rental recommendation, graph structure may connect neighborhoods, transit access, building managers, furnishing status, or accessibility features. An audit must therefore inspect relation pathways and relation-type contributions rather than only final ranks. [4]

DeepSentRec shows that review text and sentiment signals can be integrated with hybrid filtering and reinforcement-learning-based ranking. The housing transfer is clear: text about maintenance, noise, landlord responsiveness, or neighborhood atmosphere may alter ranking materially. Yet text can also contain strategic manipulation or discriminatory language. A fairness audit must therefore ask not only whether sentiment changes relevance, but whether it changes affordable exposure, rank stability, or explanation behavior across groups. [5]

The supplementary house-price prediction study provides a useful caution about interpretability. It states that several reported error metrics were calculated on a normalized target scale and therefore should not be read as direct monetary errors, and it recommends adding explainability analyses in future work. The present article extends that caution to rental recommendation auditing: click-through rate, ranking loss, or normalized model error may still be useful internal diagnostics, but they do not by themselves establish equitable access to housing opportunities. The proposed audit therefore reports domain-relevant outcomes such as qualified exposure and explanation stability. [6]

4. Conceptual model of fairness in rental recommendation

The framework models a rental recommender as a five-stage pipeline: candidate generation, eligibility or business-rule filtering, ranking, explanation generation, and feedback updating. Fairness can fail at each stage. A user may receive an impoverished candidate set, qualified affordable listings may be demoted within the ranked list, explanations may vary in stability or informativeness across groups, and feedback loops may amplify early disparities when clicks or saves are used as training signals.

Three audit targets are distinguished. User-side fairness asks whether users with comparable housing-relevant needs receive comparably suitable opportunities after conditioning on legitimate constraints. Listing-side fairness asks whether some listings, especially new or sparse-feedback listings, are systematically suppressed in ways correlated with neighborhood or landlord proxies. Explanation fairness asks whether the reasons shown to users are similarly stable and interpretable across groups rather than selectively opaque.

The framework also distinguishes observed constraints from inferred eligibility. Observed constraints are directly supplied or verified platform facts, such as stated budget, move-in window, household size, accessibility needs, and explicit location preferences. Inferred eligibility consists of model-based estimates such as predicted engagement or application success. The audit gives priority to observed constraints because they are more directly reviewable. Inferred eligibility is treated as a possible source of bias that must be analyzed separately.

5. Audit variables and operational definitions

Each recommendation event should include a user vector U, listing vector L, context vector C, recommendation score s, displayed rank r, candidate-set indicator c, explanation vector E, and outcome vector Y. U should contain explicit housing constraints and interaction sparsity indicators. L should contain rent, mandatory fees, size, location, amenities, accessibility markers, lease terms, freshness, and any text-derived or graph-derived features used by the system. C should include time, channel, market regime, and cold-start status. Y may include clicks, saves, inquiries, tours, applications, or lease outcomes, depending on the platform.

Observed qualification is defined as q_obs(u,l)=1 if listing l satisfies all non-missing explicit constraints for user u, and q_obs(u,l)=0 if at least one explicit constraint is violated. If a mandatory listing field required for checking an explicit constraint is missing, set q_obs(u,l)=NA. Such cases are excluded from qualification-rate denominators but their missingness rate must be reported by group and listing segment.

Affordability is defined by the operational rule a(u,l)=1 if monthly_rent(l)+mandatory_fees(l) ≤ tau × verified_monthly_budget(u), and a(u,l)=0 otherwise. The parameter tau is a policy parameter chosen before outcome inspection. If verified budget is unavailable, a stated-budget version may be used, but the lower confidence of that substitution must be recorded.

The primary outcome is the Exposure Disparity Ratio at top K. The unit of analysis is the recommendation event, defined as one ranked list shown to one user at one decision time. Let TopK(u) denote the set of listings displayed to user u at ranks 1 through K in that event, and let d_exp(r)=1/log2(r+1) for r≤K. Event-level qualified affordable exposure is x(u)=Σ_{l∈TopK(u)} d_exp(r_ul)·1[q_obs(u,l)=1 and a(u,l)=1], where r_ul is the displayed rank of listing l in that event and each displayed listing contributes once. Repeated appearance of the same listing in different events is counted separately for event-level auditing; a supplementary listing-level tabulation should report the number of unique qualified affordable listings exposed by group to detect concentration on a small inventory subset. For group g, exp_g(K) is the mean of x(u) over events assigned to users in g. Then EDR(g1,g2,K)=exp_g1(K)/exp_g2(K). EDR is computed only when each group satisfies both predeclared sufficiency thresholds: a minimum number of recommendation events and a minimum number of displayed listings with non-missing qualification status. If either threshold fails, the group comparison is reported as insufficient for the primary analysis and may appear only in sensitivity tables. For each group, calculate exposure bounds L_g and U_g by assigning every missing qualification indicator zero and one, respectively, while retaining known affordability constraints. Report the observed-data estimand alongside these bounds. When U_g2 > 0 and L_g2 > 0, the worst-case EDR interval is [L_g1/U_g2, U_g1/L_g2]. A zero lower denominator gives an unbounded upper endpoint when the numerator can be positive; two groups with identically zero possible exposure produce a non-estimable comparison. Opposite group-specific assignments, rather than two common assignments applied to both groups, define the ratio bounds. These are partial-information stress bounds, not a correction for missingness bias and not a missing-at-random estimate. A one-sided zero observed group mean is reported explicitly and sent to review; it is not evidence of unlawful discrimination.

A second outcome is Qualification Parity Error. The unit of analysis is the displayed listing within a recommendation event. For group g at top K, let N_g^chk(K) be the total number of displayed top-K listings with non-missing qualification status across all audited events for users in g, and let M_g(K) be the number among them with q_obs(u,l)=1. Then QPE(g,K)=1−M_g(K)/N_g^chk(K). QPE therefore estimates the share of checkable displayed top-K positions occupied by listings that violate explicit user constraints; it is not a proportion over all displayed positions unless qualification status is complete. Events with zero checkable top-K positions are retained in the audited event set for consistency with EDR, but they contribute zero positions to N_g^chk(K) and M_g(K); they are therefore represented through the completeness statistic rather than through an arbitrary mismatch value. If N_g^chk(K)=0, QPE is not estimated for that group-stratum cell and the result is recorded as non-estimable rather than zero; the cell is excluded from primary cross-group comparison and its frequency is reported. Each QPE report must also include the completeness rate Comp_g(K)=N_g^chk(K)/N_g^disp(K), where N_g^disp(K) is the total number of displayed top-K positions in the audited events for group g. If completeness differs materially across groups, the primary QPE comparison remains the checkable-case estimand above, but it must be interpreted jointly with completeness and with the same missingness bounds used for EDR. A supplementary user-level version is computed by first averaging the displayed-listing mismatch rate within each user over that user’s audited events and then taking an equal-weight mean across users in the group; this sensitivity analysis prevents heavy users or frequent sessions from dominating the result. For the all-displayed-position sensitivity analysis, let V_g be the number of known qualification violations and R_g the number of missing qualification labels. The mismatch-rate interval is [V_g/N_g^disp, (V_g+R_g)/N_g^disp]. It supplements, rather than replaces, the checkable-case QPE estimand. No independence assumption about missingness is asserted.

Counterfactual Rank Shift (CRS) is defined here as a paired input-sensitivity diagnostic, not an identified causal effect of protected status on housing outcomes. Before access to the audit data, a versioned governance record must list permitted audit attributes, coding schemes, permitted replacement pairs, approved purposes, and the applicable data-use and ethics authorizations. The algorithm cannot establish legal permission; a missing approval record makes that attribute ineligible. For each permitted replacement, a separate semantic predicate must be specified in the schema, covering cross-field consistency and immutable task constraints. The predicate is evaluated before a statistical-support test. Historical cell support alone cannot establish semantic validity. For a valid event and replacement b, preserve its explicit housing requirements and all other recorded inputs, reconstruct the same candidate pool, and compare the rank of each common listing under the original and perturbed records. Require at least n_min=25 historical events in the registered nonprotected-covariate cell for the primary sensitivity run; values 10 and 50 are supplementary design settings. This count is a support screen, not a validity theorem. CRS(u) is the median absolute rank difference across the permitted listing-replacement pairs. If no pair is eligible, report the event as non-estimable and disclose the exclusion rate by group. Direct labels unavailable by design make CRS inapplicable; labels present but insufficiently supported make it applicable but non-estimable. Proxy cohort labels are used only for descriptive stratification and are never perturbed. Use at most B=20 distinct eligible replacements per event; if fewer exist, use all of them without manufacturing additional counterfactual records. A binary attribute may therefore yield only one distinct replacement. For stochastic scoring, run each original/perturbed pair on the same frozen list of B seed values and report seed variability separately from the number of distinct replacements. If the original pool or model version cannot be reconstructed, do not report whole-pipeline CRS.

Explanation Stability Difference measures the stability of model attributions under permitted changes to nonprotected inputs. The primary explainer is a single preregistered model-agnostic feature-permutation procedure using a frozen training-background sample and common auditable input variables. Native explanations may be reported only as supplementary results. The background selection, score scale, permutation schedule and mapping from encoded columns to parent variables must be published before evaluation. If a comparator cannot be evaluated in that common space, its explanation metric is non-estimable; a new surrogate is not selected after seeing the results. Normalize a nonzero attribution vector E by its absolute-weight sum. A zero vector is recorded as non-estimable. Set m=min(5,d), where d is the positive number of auditable parent variables, and compare their top-m sets. Event stability is the fraction of B=20 admissible nonprotected perturbations with Jaccard agreement at least 0.60; if fewer perturbations exist, use the actual count as the denominator and disclose it. The perturbations must satisfy the frozen semantic and support rules. Group differences in average stability describe attribution robustness on eligible events, not proof that the explanations are causal or that the recommender is fair. Explanation stability may be reported independently when CRS is unavailable; paired reporting uses their common eligible subset.

All thresholds in this protocol are policy parameters fixed before disparity estimation. The default audit uses tau=1.00 for the affordability rule, K=10, a minimum of 200 audited recommendation events and 1,000 non-missing displayed listings per group-stratum cell for primary group comparison, a maximum B=20 distinct permissible perturbations per event for CRS and explanation stability, and the explanation settings stated above. Threshold precedence is strict and is applied at the group-stratum level rather than to individual events: a cell enters the primary analysis only if the full audited set for that cell satisfies both the event threshold and the non-missing listing threshold. When a cell qualifies, exp_g(K), QPE(g,K), and related summaries are computed over all audited events in that cell, including events whose own top-K lists contain few or no checkable positions; event-specific sparsity is handled by the missingness conventions in the metric definitions, not by dropping those events post hoc. These defaults are operational rather than universal. A preregistered sensitivity procedure should vary one parameter family at a time over declared alternatives—tau in {0.90,1.00,1.10}, K in {5,10,20}, minimum event counts in {100,200,500}, non-missing listing thresholds in {500,1000,2000}, n_min in {10,25,50} for the CRS support rule, and B in {10,20,50}—while holding the remaining defaults fixed. The primary interpretation must rely on the default rule, with sensitivity results used to judge robustness rather than to search for a preferred conclusion.

6. Audit methodology and decision sequence

Step 1 is system mapping. The auditor identifies which architectural components are present: collaborative filtering, content features, graph propagation, sentiment extraction, sequential modeling, reinforcement-learning ranking, federated training, and post hoc explanation tools. This matters because different components create different fairness risks. Graph propagation may transmit neighborhood proxies; adaptive ranking may amplify engagement inequalities; federated training may conceal local harms behind global aggregates. [4] [5] [3]

Step 2 is task specification. The target is rental listing recommendation, not sales valuation and not generic preference prediction. This distinction is necessary because the direct real-estate evidence shows that rental and sales channels can respond differently to the same intervention. Audit logic should therefore be tied to rental access variables such as qualified exposure and lease-relevant constraints, not imported wholesale from sales analytics. [1]

Step 3 is cohort and stratum construction. Users should be grouped by lawful protected attributes when available. When they are unavailable, proxy cohorts may be used only for descriptive disparity stratification and only under a preregistered three-part rule. First, the proxy must be constructed from non-outcome administrative variables available at recommendation time rather than from clicks, applications, tours, or lease outcomes. Second, the variable set and coding scheme must be frozen before disparity estimation; no post hoc search over multiple proxy definitions is permitted. Third, the proxy must satisfy a documented governance justification showing that it approximates a legally or policy-relevant cohort construct more closely than any single eligibility variable such as budget alone; if such justification is absent, proxy-based subgroup reporting is not undertaken. Proxy cohorts are never used as treatment variables in counterfactual procedures: they may define descriptive subgroup tables for EDR, QPE, or explanation stability, but they are excluded from CRS by rule. Separate strata are needed for new users, sparse-history users, and repeat users. Listings should likewise be stratified as new, established, or sparse-feedback items, reflecting the centrality of cold-start conditions in recommender performance and privacy-preserving system design. [3]

Step 4 is comparator selection. The primary attribution rule is one-component-at-a-time ablation. For each audited high-risk component h, construct comparator M^(−h) by removing or neutralizing h while holding fixed the candidate-generation rules, training and audit data partitions, feature coding outside h, optimization objective, random-seed schedule where feasible, rank cutoff K, and evaluation cohorts. If removal of h changes input dimensionality, the replacement representation must be the simplest neutral alternative declared in advance, such as zeroing sentiment features, bypassing graph propagation while retaining base item and user features, or disabling reinforcement updates while keeping the same static ranker. A comparator remains admissible for ranking-stage attribution only if it preserves the audited candidate set and the common auditable feature space. If removing h necessarily alters candidate retrieval or deletes features needed to score part of the original candidate pool, the comparator is reclassified as a candidate-generation or support-shift audit rather than a ranking-stage comparator, and its results are reported separately. To address interaction effects, each one-component comparator is accompanied by a residual interaction check: if two single ablations produce materially different fairness conclusions and the architecture couples those components directly, add a joint ablation M^(−h1,−h2) and report that attribution is non-separable. [5] [4]

Step 5 is metric estimation by pipeline stage. Candidate-generation fairness is measured at the event level by the rate at which qualified affordable listings enter the candidate pool before ranking. Ranking fairness is then estimated on the same event set using EDR, QPE, and CRS under the denominator conventions defined above. Explanation fairness is estimated on the subset of events for which the declared explanation method is available and auditable. Feedback fairness is not inferred from aggregate engagement alone. Instead, if logs permit replay, the auditor iterates the observed recommendation-update cycle over fixed historical horizons and records whether group disparities widen after model updating. The primary estimand throughout the audit is the event-level group summary, because the recommendation event is the platform action under review. User-level summaries are secondary sensitivity analyses and are computed by first averaging the relevant event-level metric within each user over that user’s audited events and then taking an equal-weight mean across users in the group. Repeated display of the same listing within a user remains counted at the event stage; the user-level sensitivity changes only the weighting of frequent sessions. The event-level estimand remains primary when event- and user-weighted results differ. The user-level result is a sensitivity analysis of session concentration, not a second gate that can erase a primary event-level disparity. Report the direction and threshold crossing under both weighting schemes without merging them into a single robustness label.

Step 6 is local sensitivity analysis. Borrowing the explanatory logic of EX-KAN, the auditor estimates how recommendation scores change with respect to variables such as rent, fees, distance, freshness, review sentiment, and recent interaction counts. Derivatives may be used where available; otherwise finite differences are acceptable if applied consistently. The audit should compute these sensitivities within predeclared market regimes rather than only on the pooled sample, because variable effects may differ across market conditions. Group comparisons then ask whether the same variable produces systematically different score responses across audited cohorts after conditioning on the declared regime and observed constraints. [2]

The proposed adjudication rule uses three triage states: no threshold trigger detected, review required, and high-priority review. These are research-workflow labels, not findings of legal discrimination. In an estimable group comparison, flag the primary EDR when it falls outside [2/3,1.5], and flag qualification mismatch when the absolute checkable-case QPE difference exceeds 0.10. Either flag is sufficient for high-priority review; a one-sided zero exposure mean is also sent to that state. If a primary metric is non-estimable, or if only CRS or explanation stability triggers its separately registered diagnostic threshold, assign review required rather than a favorable finding. No trigger detected requires estimable primary metrics with neither primary flag. These numerical bands are illustrative governance choices requiring prospective validation, not accepted fairness standards. Report cluster-bootstrap uncertainty and missingness bounds beside every point-based triage result. Diagnostics cannot compensate numerically for an exposure or qualification flag. Any remediation remains a proposed intervention whose effects must be evaluated afresh.

7. Comparison with specific prior methods

The proposed framework differs from DeepSentRec in objective. DeepSentRec integrates review sentiment, hybrid filtering, and reinforcement-learning-based ranking to improve product recommendation outputs. The present framework treats those same components as audit targets and asks whether they change qualified affordable exposure, counterfactual rank shift, or explanation stability unequally across groups. [5]

The framework also differs from CRAFT and KGRec in how it handles sparsity and relational structure. CRAFT is relevant because it emphasizes cold-start difficulty and federated training under privacy constraints; the audit converts those facts into separate reporting rules for new users, new listings, and decentralized computation of fairness statistics when raw logs cannot be pooled. KGRec is relevant because higher-order relations can improve recommendation under sparse data, yet the same relations may carry neighborhood or institutional proxies; the audit therefore adds relation-type ablations and path inspection rather than relying on accuracy alone. [3] [4]

The supplementary house-price study supports a narrower caution: internal error metrics calculated on transformed scales may be useful for model comparison yet still be hard to interpret in practical units. The audit therefore reports housing-access outcomes directly and does not treat engagement or normalized model error as evidence of equitable access. [6]

8. Prospective validation protocol

Because this article is conceptual, validation is stated as a prospective protocol rather than as completed evidence. The protocol assumes access either to historical platform logs or to a replay environment derived from them under appropriate governance.

Phase A is retrospective disparity estimation. The auditor computes EDR, QPE, CRS, and explanation stability within predeclared user and listing strata, disaggregated by geography, affordability band, cold-start status, and market regime. In federated settings, locally computed sufficient statistics may be aggregated when raw interaction data cannot be centralized. [3]

Comparator replay is restricted to archived candidate pools, frozen model versions and common-random-number seed schedules. A one-component ablation supports a ranking-stage software comparison only when removing that component leaves retrieval, score support and input availability unchanged. It does not identify the causal effect of deployment on housing access. An ablation affecting retrieval is a separate end-to-end system comparison. If the full pool or seeded retrieval process cannot be reconstructed, candidate-generation metrics and whole-pipeline CRS are non-estimable. The audit may still report exposure and qualification from displayed lists, explicitly conditional on that observed stage. Incomplete pool logs are never treated as a complete inventory. List which comparator claims remain available under each access condition. [5] [4]

Phase C is explanation and sensitivity auditing. For sampled events, the auditor generates local explanations in the common auditable feature space, runs the preregistered perturbation tests, and estimates score sensitivities for key variables such as rent, fees, distance, freshness, sentiment features, and recent interaction counts. Particular attention is given to whether apparently legitimate drivers are displaced by opaque proxies such as neighborhood embeddings or latent text features. [2] [4]

Before any prospective deployment, normalize qualified exposure by M_K=sum from r=1 to K of 1/log2(r+1), the maximum weighted exposure of a fully qualified affordable top-K list. For each sufficiently observed group, compare this normalized mean under the revision and the frozen comparator. The illustrative non-inferiority margin is 0.02 of M_K: the lower endpoint of a preregistered 95% user-cluster bootstrap interval for revision-minus-comparator exposure must be at least -0.02 in every audited primary stratum. A non-estimable stratum or a high-priority primary-metric trigger suspends deployment under this proposed rule. The margin and bootstrap design are protocol settings, not validated social or legal adequacy thresholds. A failed guardrail leads to further research and review rather than an automatic adverse decision about any user.

9. Assumptions, failure cases, and limits of generalization

The framework assumes that recommendation events can be linked to candidate sets and displayed ranks, that explicit housing constraints are recorded with tolerable error, and that some lawful basis exists for defining protected or proxy cohorts. It audits a specific construct: disparity in recommendation-stage access to qualified affordable options and in the stability of rank and explanation behavior under controlled perturbations. If the identification assumptions fail, the audit becomes a structured stress test rather than a definitive fairness finding.

One failure case arises when clicks, saves, or applications already reflect discriminatory behavior by users, landlords, or earlier platform design. If those outcomes are treated as ground truth for relevance, the audit may inherit bias from the behavior it aims to evaluate. A second failure case is differential missingness in rent components, fees, or accessibility features. Because such missingness directly affects qualification and affordability checks, it must be reported as an audit result rather than hidden in preprocessing.

Graph-based systems pose a further difficulty. Removing one relation type may not eliminate a proxy if the signal has already diffused through other edges or embeddings. KGRec’s discussion of graph scale and propagation cost implies that fairness debugging in these systems may also be structurally complex. Sentiment-aware systems create another boundary condition, since discriminatory or manipulated review text may distort ranking before the final ranker acts on it. [4] [5]

Federated systems create additional limits. CRAFT emphasizes client-specific models together with global aggregation. An audit based only on global averages may therefore miss harms concentrated in particular clients or regions. Where governance permits, fairness reporting should be repeated at the client or regional level using the same preregistered estimands rather than ad hoc local metrics. [3]

Generalization is also limited by market regime and by domain transfer. EX-KAN shows that explanatory relations differ across market regimes, so fairness results from one rental market condition should not be assumed to hold in another. The direct real-estate study concerns commercial offices rather than residential rental platforms. Its relevance here is analogical and bounded: it illustrates that sales and rental outcomes can diverge after the same intervention, not that residential recommendation disparities have any demonstrated effect on rents or leasing outcomes. [2] [1]

10. Conclusion

This article proposes a conceptual fairness-audit framework for rental recommendation systems. The direct real-estate study is used for a limited inference: sales and rental outcomes may respond differently to the same policy signal, so rental-platform auditing should not rely on posted-rent movement as its main fairness proxy. The transferable recommender studies contribute distinct technical elements—regime-aware sensitivity analysis, cold-start and federated constraints, higher-order relational structure, and sentiment-aware ranking—but none of them supplies a complete audit of recommendation-stage housing access on its own. [1] [2] [3] [4] [5]

The manuscript therefore proposes operational definitions for observed qualification and affordability, default event-level estimands for exposure and qualification mismatch, a reproducible cell-support rule for counterfactual rank shift, a common-space requirement for explanation auditing, proxy-cohort restrictions, and a prospective validation sequence. The contribution is methodological rather than empirical. No claim is made that the protocol has yet been validated on observed rental-platform data, nor that any proposed threshold is uniquely correct. The defaults are governance choices to be fixed in advance and checked by preregistered sensitivity analysis before the protocol is used for substantive platform assessment. The framework also states two limits directly: counterfactual rank auditing is unavailable when lawful protected-attribute coverage is inadequate, and disagreement between event-level and user-level summaries limits the strength of any final fairness inference.

References

1. Eichholtz, Piet; Kok, Nils; Sun, Xudong (2026). The impact of minimum energy performance standards on the commercial real estate market. Nature Communications. 17(1), 4038. doi:10.1038/s41467-026-70684-w.
2. Le, Dat; Rajasegarar, Sutharshan; Luo, Wei; Nguyen, Thanh Thi; Angelova, Maia (2026). From prediction to explanation for price–volume dynamics in real estate market. Interface Focus. 16(3), 20250069. doi:10.1098/rsfs.2025.0069.
3. Sivakumar, Navaneeth; John, Reuben Stephen; Bijo, Allen; Suganeshwari, G.; Anbalagan, Sudha; Thandapani, Sujithra (2026). CRAFT: cold-start recommender with attention and federated training. Scientific Reports. 16(1), 17150. doi:10.1038/s41598-026-47175-5.
4. Hoan, Trinh Duong; Hung, Bui Thanh (2026). KGRec: A knowledge graph attention-based model for recommender system. PLOS One. 21(3), e0344585. doi:10.1371/journal.pone.0344585.
5. Kollu, Siva Rama Prasad; Garapati, Yugandhar (2026). DeepSentRec: a deep learning-based sentiment-aware product recommendation system. Scientific Reports. 16(1), 10580. doi:10.1038/s41598-026-45953-9.
6. Alhammad, Sarah M.; Fouad, Yasser; Mahmoud, Amira A.; Osman, Ahmed M.; El-Bakry, Hazem M.; Elshewey, Ahmed M. (2026). House price prediction using a hybrid GRU–MLP based on binary whale optimization algorithm and ant colony optimization for hyperparameter tuning. Scientific Reports. 16(1), 23225. doi:10.1038/s41598-026-62854-z.