Building Function Matching Index (BFMI)

• BFMI is a composite metric that validates building function labels against independently derived POI probability fields.
• It integrates two components—Top-1 consistency for hard agreement and cosine similarity for soft distributional alignment.
• Empirical evaluations in GBA-UBF show that BFMI improves label alignment with urban activity patterns compared to baseline methods.

Searching arXiv for the cited BFMI-related paper and closely related terms. The Building Function Matching Index (BFMI) is a composite index for quantitative validation of building-level function classification against independently derived POI-based probability fields. It was introduced in the context of the Greater Bay Area Urban Building Function Dataset (GBA-UBF), where nearly four million buildings across six core Greater Bay Area cities are assigned one of five functional categories through the Multi-level Building Function Optimization (ML-BFO) pipeline. BFMI is designed to measure two distinct but complementary properties: whether an assigned building label agrees with the dominant POI-derived function in its vicinity, and whether that hard label is also consistent with the full local POI-inferred class distribution. By averaging these two aspects over all buildings, BFMI yields a scalar score in $[0,1]$ that reflects both hard-label agreement and soft probabilistic alignment (Chen et al., 10 Oct 2025).

1. Conceptual basis

BFMI is defined for a set of $N$ buildings $B_1,\dots,B_N$, each assigned a function label $y_i\in\{1,\dots,C\}$. In the GBA-UBF formulation, $C=5$. The validation reference is not another parcel-level or building-level categorical map, but a set of independently generated continuous density surfaces derived from Points of Interest (POIs). For each class $c$, a kernel density estimate produces a class-specific field $D_c(x)$, and the normalized probability field is then

$$P_{\mathrm{poi},c}(x)=\frac{D_c(x)}{\sum_{k=1}^{C} D_k(x)+\epsilon},$$

where $\epsilon$ is a smoothing constant used to avoid division by zero.

For each building footprint $B_i$, BFMI computes the zonal average of the POI-derived probability fields:

$N$0

with

$N$1

so that $N$2.

The assigned building label is encoded as a one-hot vector,

$N$3

This construction makes BFMI a comparison between a hard categorical assignment and a soft spatially varying probability distribution. In that sense, it is explicitly intended to validate both semantic correctness at the top class and compatibility with mixed or uncertain local urban activity patterns.

2. Component metrics

BFMI consists of two per-building terms: Top-1 Consistency and Cosine Similarity.

The categorical term is

$N$4

which is equal to $N$5 if the assigned building label $N$6 matches the POI-dominant class over the footprint, and $N$7 otherwise. This is the hard agreement component.

The distributional term is

$N$8

Because $N$9 is one-hot, the numerator extracts the POI probability mass assigned to the predicted class, while the denominator normalizes by the $B_1,\dots,B_N$0 norm of the POI vector. The quantity $B_1,\dots,B_N$1 lies in $B_1,\dots,B_N$2, with value $B_1,\dots,B_N$3 when the entire POI-derived probability mass is concentrated on the predicted class.

These two terms play different roles. Top-1 Consistency asks whether the dominant local activity pattern agrees with the assigned function. Cosine Similarity asks whether the assignment is compatible with the full local probability profile, including cases in which the dominant class is correct but substantial mass remains on other functions. BFMI therefore operationalizes both discrete agreement and graded distributional alignment.

3. Aggregate definition and computation workflow

The full index is the average of the two component terms:

$B_1,\dots,B_N$4

where $B_1,\dots,B_N$5 balances the categorical and distributional components. In the GBA-UBF study, the default is $B_1,\dots,B_N$6.

The computation requires four classes of inputs: building polygons with assigned labels, raw POIs with class tags, KDE parameters including bandwidth $B_1,\dots,B_N$7, smoothing parameter $B_1,\dots,B_N$8, balancing weight $B_1,\dots,B_N$9, and a raster template specifying extent and resolution. The workflow proceeds in four stages. First, POIs are rasterized by class and each class-specific density map $y_i\in\{1,\dots,C\}$0 is estimated using KDE with bandwidth $y_i\in\{1,\dots,C\}$1. Second, these density maps are normalized into probability fields $y_i\in\{1,\dots,C\}$2. Third, for each building footprint, raster cells inside the polygon are extracted, zonal averages $y_i\in\{1,\dots,C\}$3 are computed, and the Top-1 and cosine terms are evaluated. Fourth, the per-building values are aggregated to produce the final BFMI score.

The procedure is modular in the sense that the POI-derived reference can be recomputed under different kernel bandwidths, raster resolutions, or class taxonomies, while the aggregation rule remains unchanged. This suggests a reproducible validation pipeline rather than a dataset-specific heuristic.

4. Interpretation and reported empirical behavior

BFMI takes values in $y_i\in\{1,\dots,C\}$4. If $y_i\in\{1,\dots,C\}$5, the classification agrees almost perfectly with the POI-derived distributions: dominant classes match broadly and the probability mass concentrates on the assigned label. If $y_i\in\{1,\dots,C\}$6, most buildings are assigned labels that do not correspond to local POI clusters, and the POI mass is concentrated on other functions.

The weighting parameter controls what the score emphasizes. As $y_i\in\{1,\dots,C\}$7, BFMI approaches pure Top-1 accuracy; as $y_i\in\{1,\dots,C\}$8, it approaches pure distributional alignment. Setting $y_i\in\{1,\dots,C\}$9 equally rewards hard agreement and soft consistency.

In the GBA-UBF evaluation, BFMI was used to compare ML-BFO building labels with a baseline inherited from parcel-level EULUC-China 2.0 labels. The reported values are: ML-BFO/GBA-UBF, $C=5$0; baseline, $C=5$1; Top-1 consistency, GBA-UBF $C=5$2 versus baseline $C=5$3; cosine similarity, GBA-UBF $C=5$4 versus baseline $C=5$5. These gaps were interpreted as showing that the fine-grained ML-BFO labels align more closely with urban activity patterns as captured by POI density (Chen et al., 10 Oct 2025).

No further ablation on $C=5$6 or KDE bandwidth was reported. The paper notes, however, that changing $C=5$7 shifts sensitivity between strict label matching and soft distributional agreement.

5. Practical considerations and methodological extensions

Several implementation choices materially affect BFMI.

Heat-map resolution and bandwidth determine the trade-off between local detail and smoothness. Small $C=5$8 or high raster resolution may produce noisy probability fields in POI-sparse areas, whereas large $C=5$9 may oversmooth mixed-use zones. A spatial cross-validation strategy or a rule-of-thumb such as Silverman’s may guide bandwidth selection.

The smoothing constant $c$0 prevents division by zero in areas without POIs. Values on the order of $c$1 to $c$2 of the typical density sum are described as sufficient.

The balancing weight $c$3 controls the relative emphasis on exact label agreement versus probabilistic alignment. If exact label matching is the primary concern, one chooses $c$4; if soft alignment is more important, including settings with multi-function structures, one chooses $c$5.

Building footprint–raster alignment matters because zonal averages are computed over the raster cells inside each polygon. Accurate estimation therefore depends on raster coverage of the footprint; partial-cell weighting or a finer grid may be used.

Class imbalance affects the POI density fields because rare function classes yield sparser $c$6 surfaces. The methodological notes indicate that one may optionally re-weight cosine terms by class frequency or use a class-specific bandwidth in KDE.

Mixed-use buildings represent a limitation of the baseline formulation because $c$7 is one-hot. The proposed extension is to replace the one-hot encoding with a soft-label distribution and compute cosine similarity directly against $c$8. This would preserve the distributional logic of BFMI while relaxing the single-function assumption.

Taken together, these considerations show that BFMI is not only a score but also a validation protocol whose behavior depends on spatial smoothing, rasterization, and class modeling choices.

6. Scope, significance, and disambiguation

Within building-level urban function analysis, BFMI serves as a validation metric that bridges coarse categorical maps and fine-grained urban analytics. Its significance lies in pairing a hard classification output with an independently generated probabilistic spatial reference rather than comparing only against another categorical layer. This makes it particularly suitable when parcel boundaries are coarse, intra-block heterogeneity is substantial, and POIs provide a richer proxy for urban activity.

A common source of confusion is acronym overlap. In the surrounding arXiv literature, unrelated methods address incremental FM-index construction for long sequence reads (Li, 2014), index-estimation precision and matching calipers in causal inference (Hansen, 2023), and pseudo-minimal automata for subset-matching queries (Fredriksson, 2010). This suggests that BFMI should be interpreted in a domain-specific manner: in the GBA-UBF setting it denotes the Building Function Matching Index, a composite metric for validating building-level function classification against POI-derived probability fields, rather than a sequence index, a matching-caliper construction, or a subset-matching automaton.

In that domain-specific sense, BFMI provides a reproducible and tunable framework for evaluating whether building labels are both categorically correct and distributionally plausible. Its formulation is sufficiently explicit to support direct reuse across cities and classification workflows, provided that the underlying POI-derived probability fields and building labels are defined on compatible class systems.