FARE in Diverse Research Areas

• FARE is a term that designates a cluster of unrelated research concepts spanning public transport enforcement, fare pricing, machine learning, biometrics, robotics, and symbolic tensor reduction.
• It encompasses varied methodologies such as causal inference in fare evasion detection, dynamic programming for tariff design, regression models in fare prediction, and hierarchical control in robotic exploration.
• Empirical findings include a 26% reduction in detected fare evasion with plainclothes inspections, robust fault-aware training in edge AI, and efficient tensor reduction in high-energy physics.

FARE, FaRE, FaRe, and Fare designate a cluster of unrelated research terms rather than a single settled concept. In recent arXiv usage, the label appears in public-transport enforcement and pricing, ride-hailing and taxi-fare analytics, hardware-aware GNN training, face-recognition evaluation and radar sensing, fair representation learning, robotic exploration and recovery, and symbolic tensor reduction in high-energy physics (Wallimann et al., 23 Jun 2026, Schöbel et al., 12 Feb 2025, Dhingra et al., 2024, Xu et al., 2019, Kahya et al., 14 Jan 2025, Jovanović et al., 2022, Liao et al., 21 Jan 2026, Fiorentin, 2015). The term therefore requires disambiguation by domain, capitalization, and methodological context.

Label	Domain	Referent
FARE	Public transport	Inspector visibility and fare evasion detection (Wallimann et al., 23 Jun 2026)
Fare	Transport economics	Fare structure design, zoning, and ticket pricing (Schöbel et al., 12 Feb 2025)
FARe	Edge AI hardware	Fault-aware GNN training on ReRAM-based PIM accelerators (Dhingra et al., 2024)
FaRE	Biometrics	Face recognition performance evaluation package (Xu et al., 2019)
FARE	Radar ML	Face recognition and OOD detection with short-range FMCW radar (Kahya et al., 14 Jan 2025)
FARE	Fair ML	Fairness with Restricted Encoders (Jovanović et al., 2022)
Fare / FARE	Robotics	Failure resilience and fast-slow exploration (Wang et al., 28 Oct 2025, Liao et al., 21 Jan 2026)
FaRe	High-energy physics	Tensor reduction of Feynman integrals in Mathematica (Fiorentin, 2015)

1. Fare enforcement in proof-of-payment systems

In public-transport research, FARE most directly denotes the study of fare inspection and fare evasion. "Visible or Covert? The Causal Effect of Inspector Visibility on Fare Evasion Detection: A Causal Machine Learning and Policy Learning Approach" examines whether uniformed inspections or plainclothes inspections are more effective in a proof-of-payment public transport system, using 21,727 inspection records from PostAuto, the largest regional bus operator in Switzerland (Wallimann et al., 23 Jun 2026).

The paper formalizes inspector visibility as a binary treatment,

$$D = 1 \quad \text{for uniformed inspections (Präsenzkontrolle)}, \qquad D = 0 \quad \text{for plainclothes inspections (Normalkontrolle)},$$

and defines inspection efficiency as detected fare evaders per unit of inspection time,

$$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$

The causal target is the average treatment effect

$\tau = \mathbb{E}[Y(1)-Y(0)],$

with conditional effects $\tau(x)$, GATES summaries $\gamma_k$, and an optimal policy function

$\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$

Identification is based on a selection-on-observables framework with conditional independence and common support, and estimation uses causal forests, Best Linear Predictor analysis, GATES, and a policy tree with a maximum of four leaves for interpretability (Wallimann et al., 23 Jun 2026).

The central empirical result is that plainclothes inspections outperform uniformed inspections in detected fare evaders per inspection hour. The estimated average treatment effect of uniformed versus plainclothes inspections is $-0.173$ incidents/hour with $\text{SE}=0.028$ and $p<0.001$, corresponding to a relative reduction of approximately 26%. The descriptive means point in the same direction: plainclothes inspections average 0.69 incidents/hour versus 0.46 for uniformed inspections (Wallimann et al., 23 Jun 2026).

The heterogeneity analysis is narrower than the headline effect. The distribution of estimated CATEs is centered around the negative ATE, the BLP finds that none of the pre-selected contextual covariates are significant at conventional levels, and the GATES analysis shows variation in magnitude but not in sign. The most negative group has $-0.263$ incidents/hour, the least negative group has $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$0, and the difference between the extremes is 0.203 incidents/hour with $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$1. Variable-importance results suggest that population size, different measures of GA ownership, the share of foreign residents, unemployment, and prior inspection hours matter most for treatment-effect variation, but these do not translate into strong, systematic subgroup differences in the BLP (Wallimann et al., 23 Jun 2026).

The policy-learning exercise yields a mostly covert enforcement rule. The policy tree first splits on the share of foreign residents, with plainclothes inspections recommended above the sample median of 21.9%. Below that threshold, the next split is population size: above-median population municipalities are assigned uniformed inspections, while below-median population lines remain plainclothes. Plainclothes inspections are recommended for 18,109 of 21,727 observations, or 83.3% (Wallimann et al., 23 Jun 2026).

A common misreading is that the result settles the welfare question. The paper explicitly cautions that it speaks to detection efficiency, not necessarily overall welfare: uniformed inspections might have deterrence benefits by discouraging evasion before it occurs, and they may also affect passenger perceptions of security. This suggests that immediate detection per hour and broader enforcement objectives should be analytically separated (Wallimann et al., 23 Jun 2026).

2. Fare structure, zoning, and dynamic ticket pricing

A second usage of fare concerns tariff design and pricing in transport economics. "Fare Structure Design in Public Transport" defines a fare structure as a function $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$2 assigning a price to every path $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$3 in a network, and studies the problem of minimizing the weighted sum of absolute deviations from reference prices,

$$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$4

The paper analyzes flat tariffs, affine distance tariffs, and zone tariffs, together with the no-elongation and no-stopover properties. Flat tariffs reduce to a weighted median problem and are solvable in $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$5. Affine distance tariffs become a least absolute deviations regression problem, or a 1-median-line location problem, and are also solvable in linear time $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$6. Zone tariffs are computationally harder: the arbitrary-zone variants $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$7 and $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$8 are NP-complete even for $$Y = \frac{\text{number of detected fare evaders}}{\text{inspection time}}.$$9, and the connected-zone variants $\tau = \mathbb{E}[Y(1)-Y(0)],$0 and $\tau = \mathbb{E}[Y(1)-Y(0)],$1 are NP-complete even when the underlying graph is a tree (Schöbel et al., 12 Feb 2025).

The paper’s treatment of structural constraints is operationally central. No-elongation requires longer tickets not to be cheaper than shorter ones covering part of the same journey, while no-stopover requires that splitting a journey into multiple tickets not reduce price. Flat tariffs always satisfy both properties. Affine network distance tariffs always satisfy both no-elongation and no-stopover, whereas affine beeline distance tariffs satisfy no-stopover but not no-elongation. For zone tariffs, sufficient conditions are expressed as monotonicity and subadditivity restrictions on the price list $\tau = \mathbb{E}[Y(1)-Y(0)],$2 (Schöbel et al., 12 Feb 2025).

"Fare Zone Assignment" studies a related but distinct zoning problem on tree networks. A zoning is represented by a cut set $\tau = \mathbb{E}[Y(1)-Y(0)],$3, the pricing function $\tau = \mathbb{E}[Y(1)-Y(0)],$4 is non-decreasing and concave, and commodity $\tau = \mathbb{E}[Y(1)-Y(0)],$5 contributes

$\tau = \mathbb{E}[Y(1)-Y(0)],$6

The objective is $\tau = \mathbb{E}[Y(1)-Y(0)],$7. For rooted instances, the paper gives an exact dynamic program with $\tau = \mathbb{E}[Y(1)-Y(0)],$8 complexity. For general trees, it proves a randomized $\tau = \mathbb{E}[Y(1)-Y(0)],$9-approximation and a randomized $\tau(x)$0-approximation in expectation. On paths, FZA is strongly NP-hard, but the paper outlines a PTAS and derives XP/FPT results for several natural parameters (Hoefer et al., 22 Dec 2025).

Dynamic ticket pricing introduces a further meaning of fare. "Fare Game: A Mean Field Model of Stochastic Intensity Control in Dynamic Ticket Pricing" models sellers of discrete perishable goods with finite inventories over a finite horizon. Inventory evolves as

$\tau(x)$1

with sale intensity $\tau(x)$2. The equilibrium is characterized by a coupled system of Hamilton-Jacobi-Bellman and Kolmogorov differential-difference equations, and the paper proves existence and uniqueness results under certain conditions. It then gives a fixed-point numerical algorithm and a preliminary comparison with airfare data for Tuesday morning flights from Chicago O’Hare to New York LaGuardia (Aydin et al., 16 Jun 2025).

Across these three lines of work, fare is not merely a price label. It is an optimization object whose admissible structure depends on network topology, behavioral constraints, inventory dynamics, and equilibrium interaction. This suggests that public-transport tariff design, revenue-maximizing zoning, and airline-style dynamic pricing are mathematically adjacent but institutionally distinct problems.

3. Ride-hailing and taxi-fare analytics

A third usage of fare concerns digital comparison and prediction systems. "Fare Comparison App of Uber, Ola and Rapido" presents a prototype fare-comparison web application intended to compare costs and ETA across ride-hailing providers. The backend is implemented in Python. Its workflow is described as Start → Environment Setup → App Installation → Data Extraction → Uber → Ola and Rapido (automation-based) → End. The data strategy is mixed: Ola uses a method aligned with its pricing policy, Rapido uses Bangalore-reported data and converts area names into continuous values to compute distance, and Uber uses random data to test the system because of API access difficulties. The paper reports that users were said to save 10–15% on average by choosing the ride recommended after comparison, but it also makes clear that the system is not using fully live public APIs for all three services (Sawant et al., 3 Dec 2025).

That limitation matters methodologically. The paper explicitly frames API access, Android Studio emulator use, Appium-based automation, location comparison, and automation-based testing as practical barriers. It also states that the comparison layer is rule-based and pragmatic rather than a formal optimizer: gather or estimate fares, compare them, and show the lowest or best option. A plausible implication is that the work is better read as an engineering prototype for transparency and unified access than as a benchmark-grade fare inference system (Sawant et al., 3 Dec 2025).

"Robust Taxi Fare Prediction Under Noisy Conditions: A Comparative Study of GAT, TimesNet, and XGBoost" addresses a different problem: supervised fare regression on the NYC Yellow Taxi Trip Records with over 55 million records. The paper evaluates GAT, XGBoost, and TimesNet on raw and denoised data, and studies predictive accuracy, calibration, uncertainty estimation, OOD robustness, and feature sensitivity. The preprocessing pipeline includes KNN imputation, Gaussian noise injection, autoencoder-based denoising, and chunk-wise processing with PySpark and Dask (Moorthy, 26 Jul 2025).

The quantitative results are strongly model-dependent. XGBoost is reported as the strongest overall performer, with clean-data MAE 0.1040, MSE 0.0279, and $\tau(x)$3, and noisy-data MAE 0.8203, MSE 1.4407, and $\tau(x)$4. GAT degrades sharply under noise, moving from MAE 1.1057, MSE 2.4414, $\tau(x)$5 on clean data to MAE 2.5036, MSE 11.4375, $\tau(x)$6 on noisy data. TimesNet shows moderate robustness but weak overall predictive quality, with clean-data MAE 1.3549, MSE 3.3534, $\tau(x)$7, and noisy-data MAE 1.7303, MSE 5.8835, $\tau(x)$8 (Moorthy, 26 Jul 2025).

These two studies illustrate a basic distinction within fare analytics. One line builds cross-platform comparison interfaces under limited data access; the other studies statistical robustness of regression models under missing values, synthetic Gaussian noise, and OOD-like perturbations. The former emphasizes aggregation and user transparency, whereas the latter emphasizes calibration, denoising, and deployment-grade predictive robustness.

4. Machine-learning, biometrics, and hardware frameworks

Outside transport, FARE and its capitalization variants often denote technical frameworks. "FARe: Fault-Aware GNN Training on ReRAM-based PIM Accelerators" addresses stuck-at faults in ReRAM-based processing-in-memory hardware used for GNN training. The paper argues that GNNs are especially sensitive because both the graph adjacency matrix and the trainable weights are mapped onto ReRAM crossbars. FARe therefore separates mitigation into two mechanisms: a fault-aware adjacency-matrix mapping algorithm for the aggregation phase, formulated as a weighted bipartite matching problem and implemented with the b-Suitor algorithm, and weight clipping for the combination phase. The framework assumes a built-in self-test circuit can identify the location and type of SAFs, with around 0.13% area/timing cost. On PPI, Reddit, OGBL, and Amazon2M, and across GCN, GAT, and GraphSAGE, FARe restores GNN test accuracy by 47.6% on faulty ReRAM hardware with a ~1% timing overhead compared to the fault-free counterpart, and can provide up to 4× speedup over existing fault-tolerant approaches (Dhingra et al., 2024).

In biometrics, "Open Source Face Recognition Performance Evaluation Package" uses FaRE to denote an evaluation toolbox rather than a recognition model. The package focuses on the evaluation stage of the face recognition pipeline, supports both online and offline evaluation, abstracts comparison, closed-set, and open-set protocols, and implements ROC, PR, ACC, EER, AUC, CMC, and DET/IET. It provides dataset APIs for LFW, CFP, UHDB31, IJB-A, IJB-B, and IJB-C, and supports set-based recognition via template fusion functions. The paper reports that processing 32 images at a time with DenseNet-121 took about 35 seconds to generate templates and compare all pairs in LFW, which it presents as evidence of practical scalability (Xu et al., 2019).

"FARE: A Deep Learning-Based Framework for Radar-based Face Recognition and Out-of-distribution Detection" reuses the acronym for a multimodal radar pipeline. Its architecture combines a Primary Path for ID face classification with Intermediate Paths, implemented as simple linear autoencoders, for OOD detection. The system operates on Range-Doppler Images and micro Range-Doppler Images, uses triplet loss for the PP, freezes the PP in a second stage, and trains the IPs with MAE-style reconstruction loss. On a custom dataset collected with a 60 GHz FMCW radar, it reports 99.30% ID classification accuracy and 96.91% AUROC for OOD detection. The paper also states that OOD training uses only ID samples, but the setup is controlled: fixed distance 25 cm, five ID identities, and a limited OOD cohort (Kahya et al., 14 Jan 2025).

In fair machine learning, "FARE: Provably Fair Representation Learning with Practical Certificates" defines FARE as "Fairness with Restricted Encoders." The core idea is to restrict the encoder representation space to a finite set of possible representations, making it possible to derive practical certificates: finite-sample, high-confidence, distribution-free, model-free upper bounds on the demographic parity distance of any downstream classifier trained on the learned embeddings. The paper instantiates the encoder with a decision tree and uses the fairness-aware split criterion

$\tau(x)$9

Its certificate has the form

$\gamma_k$0

with probability at least $\gamma_k$1, where $\gamma_k$2. The paper reports competitive accuracy-fairness tradeoffs, practical certificates that are tight and non-vacuous, and a best-accuracy gap typically around $\gamma_k$3 relative to the unfair baseline (Jovanović et al., 2022).

A recurring misconception across these works is that the shared acronym implies a shared technical lineage. It does not. Here FARE may denote hardware fault mitigation, evaluation infrastructure, multimodal radar recognition with OOD rejection, or certified fair representation learning. The commonality is nominal, not methodological.

5. Robotics and control frameworks

In robotics, Fare and FARE again denote distinct architectures. "Fare: Failure Resilience in Learned Visual Navigation Control" augments imitation-learned visual navigation policies with OOD detection, failure recognition, and heuristic recovery. The OOD-aware policy is learned with a Variational Information Bottleneck objective, test-time OOD detection uses functional conformal prediction on a latent KL score, and recognition uses Grad-CAM on the KL score to produce a heatmap identifying the image regions responsible for failure. Recovery actions are selected from backtrack, rotate_left, rotate_right, and get_help, based on a coarse three-bin discretization of the heatmap. The paper reports real-world recovery across two policy architectures, improved informed recovery relative to a random variant, and a 300 m route completed across indoor and outdoor environments (Wang et al., 28 Oct 2025).

The design has explicit limits. The recovery policy is heuristic, heatmap discretization is coarse, and spurious heatmap activations can hurt recovery, especially in local minima. The paper nevertheless argues that failure resilience can be embedded into learned policies without explicit failure trajectories in training data (Wang et al., 28 Oct 2025).

"FARE: Fast-Slow Agentic Robotic Exploration" uses the acronym for a hierarchical exploration system that integrates a LLM for global reasoning with a reinforcement learning policy for local decision making. The slow-thinking module interprets a concise textual description of the unknown environment, synthesizes an agent-level exploration strategy, and grounds it into a sequence of global waypoints through a pruned global belief graph. Pruning is based on community modularity, retaining only the top-$\gamma_k$4 communities with the highest modularity contributions. The fast-thinking module executes exploration from local graph observations plus waypoint guidance and is shaped by an instruction-following reward

$\gamma_k$5

In simulation, FARE reports $\gamma_k$6 m and $\gamma_k$7 s in indoor environments, $\gamma_k$8 m and $\gamma_k$9 s in forest environments, and $\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$0 m and $\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$1 s in warehouse environments. It is also deployed on hardware in a $\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$2 building environment (Liao et al., 21 Jan 2026).

Both robotics papers use a dual-scale architecture, but for different purposes. Fare couples OOD recognition to recovery in learned navigation control; FARE couples global semantic reasoning to local geometric execution in exploration. This suggests a broader pattern in contemporary robotics research: the acronym is repeatedly attached to systems that separate slow, diagnostic, or strategic modules from fast closed-loop control.

6. FaRe in symbolic high-energy physics

In symbolic computation, FaRe refers to a Mathematica package rather than a transport or machine-learning framework. "FaRe: a Mathematica package for tensor reduction of Feynman integrals" implements the decomposition of a generic tensor Feynman integral, with arbitrary loop number and arbitrary tensor rank, into scalar integrals in shifted, higher dimensions, with altered propagator powers (Fiorentin, 2015).

The package follows the dimensional tensor-reduction algorithm of Tarasov/Tausk and related work. Starting from a Schwinger-parameter representation, it diagonalizes the quadratic form by sequential loop-momentum shifts,

$\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$3

and then absorbs parameter factors into propagator powers and shifted dimensions. FeynCalc is required so that Lorentz structures are preserved and manipulated explicitly through objects such as Pair[LorentzIndex[...], Momentum[...]], MetricTensor[...], and FourVector[...] (Fiorentin, 2015).

Its principal functions are QP[loops], which returns $\pi^* = \arg\max_{\pi}\mathbb{E}[Y(\pi(X))].$4 after loop-momentum shifts; TRed[D, num, den, loopMomenta], the tensor-reduction routine; and FIREType[expr, irr, zeros, fLetter], which converts output into a FIRE-compatible form. The intended workflow is to define the tensor integral, run TRed, use FeynCalc to extract Lorentz coefficients if needed, reformat with FIREType, and then reduce scalar integrals to masters with FIRE. The package is therefore a tensor-decomposition front end, not a master-integral reducer itself (Fiorentin, 2015).

Within the broader disambiguation of FARE, FaRe is the clearest example of acronymic divergence across disciplines. It shares neither the application domain nor the computational objective of the transport, robotics, fairness, or recognition papers; its unifying feature is only the label.