Source-linked AI summary
An Eco-Friendly Tourism Trip Planner Driven by Z-Number Uncertainty: Evidence from Baku, Dubai, Abu Dhabi, Tashkent, and Istanbul
Aziz Nuriyev, Sherzod Turaev, Saja Al-Dabet, Mary John, Kanan Qasimzade, Alihuseyn Orujzada, Alihuseyn Orujzada, Aisha Al Ali
TL;DR
Trip planners often treat travel times and ratings as exact despite variable routing times and uneven review evidence. This paper builds a Z-number planner from routing observations and review counts; across five cities, it reduces pooled trip carbon and travel time versus a crisp-mean baseline under the same optimizer.
Problem
Trip planners treat travel times and ratings as crisp values even though travel times vary by departure hour and ratings have unequal evidentiary support from different review counts.
Source: §1 · p. 2Method
The planner constructs travel-time and rating Z-numbers from routing observations and review counts, then optimizes itineraries with a hybrid loss combining score, carbon, travel-time, and reliability terms.
Source: §1 · p. 2, §3.3 · p. 4Results
Against a crisp-mean baseline under an identical optimizer, pooled trip carbon fell 19.3% and travel time fell 11.5% across five cities; both effects kept their sign under a city-clustered bootstrap.
Source: §8 · p. 10Takeaways & Limitations
The study characterizes where effects appear and where they do not across five cities.
Source: §1 · p. 2A single crisp-mean baseline and no ablation show that the representation changes selected plans, but not which component is responsible.
Source: §7.4 · p. 10
Abstract
from arXiv · showhide
Deployed tourism trip planners ingest travel times and ratings as exact numbers, although routing services return travel times that vary with the hour of departure and a platform rating may rest on a handful of reviews or on tens of thousands. We present an eco-friendly trip planner that represents both quantities as Z-numbers, which pair a fuzzy restriction on a value with a fuzzy reliability on the evidence behind it, and that constructs them directly from routing-service observations at six departure hours and from platform review counts rather than from linguistic elicitation. A hybrid loss combines score, carbon, travel-time, and reliability terms, and an adaptive large neighborhood search returns single-day itineraries with per-leg reliability annotations through a conversational front end. We compare the planner with a crisp-mean baseline under an identical optimizer in Baku, Dubai, Abu Dhabi, Tashkent, and Istanbul, with fifteen points of interest and thirty seeds per city, and we evaluate the selected itineraries on their trip carbon and their total travel time. Pooled over 150 seed-paired runs, the planner lowers trip carbon from 3.24 to 2.61 kg (19.3%, Wilcoxon p = 0.017; city-clustered bootstrap interval 6.4% to 26.7%) and travel time from 123 to 109 minutes per trip (11.5%, p = 2.8e-10; interval 5.0% to 18.2%). The travel-time reductions survive a Holm correction in Baku, Istanbul, and Abu Dhabi, the carbon reduction is significant in Baku and large but uncertain in the dispersed catalogs of Dubai and Abu Dhabi, and neither metric changes in Tashkent. Review counts were measured for Baku and estimated for the other four cities, and the comparison involves a single baseline, so the results establish feasibility and direction rather than a general advantage. Code, routing caches, frozen routing tables, and per-seed results are openly released.
1 Introduction
Trip planners rely on uncertain travel times and ratings but commonly ingest them as crisp values, obscuring how much evidence supports each estimate. This study builds a data-driven Z-number planner and compares its selected itineraries with a crisp-mean baseline across five cities.
1 Introduction
- Routing times vary by departure hour, while identical mean ratings can rest on very different review counts, dimensions crisp-input planners discard.The paper frames this as a gap in working trip planners that use operational data rather than linguistic judgments.Source: §1 · p. 2
- Travel-time and rating Z-numbers are constructed from routing observations and review counts, then used in a planner with a hybrid loss and conversational interface.Travel times use six departure hours; itineraries carry per-leg reliability annotations.Source: §1 · p. 2
- The study asks whether operational data can support Z-number construction without linguistic elicitation, whether the representation changes selected plans, and how effects vary across cities.The comparison is under an identical optimizer, loss weights, and data.Source: §1 · p. 2
- The planner is compared with a crisp-mean baseline at thirty seeds per city in Baku, Dubai, Abu Dhabi, Tashkent, and Istanbul using physical itinerary outcomes.Because the methods compute loss differently, comparison uses carbon, travel time, transport modes, and across-seed POI stability rather than loss values.Source: §1 · p. 2
- The paper contributes data-driven Z-number construction and a hybrid loss, a seed-paired five-city comparison, and an account of where effects appear and where they do not.The comparison uses per-city Holm-corrected tests and pooled tests with a city-clustered bootstrap; a companion study uses five uncertainty representations, while this paper uses one baseline.Source: §1 · p. 2
2 Related Work
Prior trip planners generally reduce travel times and ratings to crisp values, while uncertainty-aware approaches represent variability without separating it from evidence sufficiency. This paper positions operational-data-derived Z-numbers as a way to represent both.
2 Related Work
- Academic planners typically use static or single-call travel times and mean ratings, treating POIs with sparse and abundant evidence alike when their means match.The paper cites P-Tour, CT-Planner5, PersTour, and other trip-planning work as examples of this literature.Source: §2 · p. 3
- Commercial planners present travel time as a single number and ratings as a single mean, and their undocumented optimization layers make them user-experience references rather than methodological baselines.Source: §2 · p. 3
- Stochastic and fuzzy trip-planning methods address travel uncertainty, but the reviewed formulations do not jointly represent value spread and the amount of evidence behind it.The paper identifies Z-numbers as doing so and says prior routing applications relied on linguistically specified parameters.Source: §2 · p. 3
3 The Z-Number Framework
The framework represents uncertain travel times and ratings as Z-numbers and constructs them from routing observations and review counts. It combines these representations in a hybrid loss, handles soft-constraint violations in two steps, and uses a heuristic reliability aggregation during search.
3 The Z-Number Framework
- A Z-number pairs a fuzzy restriction on an uncertain value with a fuzzy reliability on the evidence; the framework uses trapezoidal components, with triangular ratings and crisp monetary and carbon quantities.Reliability components are trapezoidal, while a crisp number is the special case with all four trapezoid values equal.Source: §3 · p. 3
3.1 The travel-time construction operator
- Travel-time restrictions use observed extremes as support and empirical quartiles as their core, while reliability rises with the number of observed departure hours.The bounds are fixed design parameters; support and reliability are monotone as observations are added, but the restriction as a whole is not, because its inter-quartile core may move either way.Source: §3.1 · pp. 3–4
3.2 The rating construction operator
- Rating Z-numbers use the platform rating and review count, with restriction spread decaying with the square root of review count and reliability representing evidence strength.Thus, POIs with the same mean rating but different evidence amounts are not scored identically.Source: §3.2 · p. 4
3.3 The hybrid loss functional
- The hybrid loss combines rating, carbon, travel-time, and reliability terms with penalties for total-time, opening-hours, and budget violations.The crisp baseline uses different rating and travel-time formulas and has no reliability term, so the methods' loss values are not comparable; their returned itineraries are compared instead.Source: §3.3 · p. 4
3.4 The two-step violation pattern
- Constraint handling first decides violation by ranking the restriction alone, then grades a violation's magnitude with a reliability-aware distance to the nearest feasible Z-number.The experiments do not isolate this pattern because both methods use it.Source: §3.4 · p. 4
3.5 The min- ˜B reliability rule
- During search, the min- B rule aggregates a sum's reliability by taking the componentwise minimum of its operands, as a heuristic summary rather than a bound.It can overstate trip-level reliability; reported reliabilities describe the weakest leg, and the experiments do not compare this rule with exact aggregation.Source: §3.5 · pp. 4–5
4 The Trip Planner System
The system separates ingestion, Z-number construction, optimization, and user interaction, with cached data supporting request-time optimization. Its default experimental optimizer is ALNS, and the released artifacts include data and per-seed itineraries.
4 The Trip Planner System
- A four-layer pipeline separates data ingestion, Z-number construction, optimization, and user interface, with caches allowing ingestion to run slowly while optimization serves requests from cached data.Construction applies the travel-time and rating operators and stores results in a snapshot-versioned SQLite cache.Source: §4 · p. 5
- The ingestion layer collects road distances, route durations across six departure hours and three transport modes, and POI ratings, review counts, and opening hours.Requests are rate-limited and retried with exponential backoff.Source: §4 · p. 5
- Three metaheuristics are available, but ALNS is the default and the only optimizer used in the experiments.The alternatives are random-restart local search and priority-encoded differential evolution.Source: §4 · p. 5
- A conversational assistant collects trip preferences and constraints, displays itineraries on an interactive map, and labels each leg's reliability as high, moderate, or low.Users can adjust the carbon weight to move between experience-oriented and sustainability-oriented settings.Source: §4 · p. 5
- The released bundle includes source code, POI catalogs, routing observations, frozen routing tables, run scripts, and per-seed itineraries with leg modes and durations.The authors state that every paper number is computed from the released itineraries.Source: §4 · p. 5
5 The Case-Study Design
The study compares the Z-number planner with a crisp-mean baseline across five city catalogs under a shared single-day scenario. Its design holds the optimizer and loss settings fixed, evaluates physical itinerary outcomes, and re-runs the released jobs using frozen routing data.
5.1 Cities, catalogs, and scenario
- Five catalogs each contain 15 POIs, while the shared scenario allows up to seven POIs during an eight-hour day starting at 10:00.The catalogs span contrasting urban forms; review counts were measured in Baku and estimated for the other four cities.Source: §5.1 · p. 5
5.2 Baseline
- B1 converts each travel-time and rating Z-number to its centroid and uses the same optimizer, weights, and penalties as ZN-Full.The comparison evaluates returned itineraries rather than loss values; no other baselines or ablations were run.Source: §5.2 · p. 6
5.3 Protocol and statistics
- Thirty seeds per method and city yielded 300 plans and 150 paired comparisons, tested with per-city Holm correction and a five-city cluster bootstrap.The bootstrap resamples only five city clusters, so the authors treat it as a directional check rather than a precise interval.Source: §5.3 · p. 6
5.4 Routing data and re-run
- The re-run used frozen OSRM tables and released routing caches; unavailable cache entries left some drive and transit options unavailable to both methods.It exactly reproduced 180 of 180 plans in Tashkent, Dubai, and Abu Dhabi, but only 11 of 120 in Baku and Istanbul, where OSRM tables had changed.Source: §5.4 · p. 6, Table 2 · p. 7
6 Results
Across five cities, ZN-Full lowers pooled trip carbon and travel time relative to B1, while per-city effects vary. It does not produce shorter itineraries overall, and the methods show similar selection stability.
6.1 Per-city carbon and travel time
- Travel time fell significantly after Holm correction in Abu Dhabi (23.5%), Istanbul (15.2%), and Baku (10.7%); Dubai’s 6.6% reduction was not significant, and Tashkent was unchanged.Carbon fell significantly in Baku by 18.1%; larger mean reductions in Dubai and Abu Dhabi were not significant, while Tashkent and Istanbul were unchanged.Source: §6.1 · p. 7, Table 3 · p. 7, Table 4 · p. 7
6.2 Pooled analysis
- 19.3% lower trip carbon (3.24 to 2.61 kg) and 11.5% lower travel time (123.0 to 108.9 minutes) for ZN-Full versus B1 across 150 pairs.Both pooled reductions keep their sign under the city-clustered bootstrap; plans average 6.06 versus 6.24 POIs, and carbon per visited POI is lower in four cities.Source: §6.2 · pp. 7–8, Table 5 · p. 8
- ZN-Full itineraries spent 26.6% of travel time walking and 26.4% in transit, versus 20.6% and 34.3% for B1, with similar car shares.Table 5 reports mode shares as proportions of total travel minutes.Source: §6.2 · p. 8, Table 5 · p. 8
6.3 Selection stability and constraints
- Both methods averaged 1.20 POIs in the strict across-seed core, and soft-constraint violations occurred on 7 ZN-Full runs and 9 B1 runs.The authors conclude that selection stability is similar and draw no conclusion from it.Source: §6.3 · p. 8
7 Discussion
The planner’s effects vary across cities, with the largest effects in the dispersed Dubai and Abu Dhabi catalogs, while the authors caution that the evidence does not establish why. The discussion links the changes to travel-time spread and itinerary selection, and scopes the findings by the data and comparison design.
7.1 Where the effect appears
- The largest effects appear in dispersed Dubai and Abu Dhabi catalogs, while Tashkent’s methods select different POIs but yield nearly the same carbon and travel time.The authors note that five cities and unmeasured catalog dispersion cannot establish a relationship between dispersion and effect.Source: §7.1 · pp. 8–9
7.2 The mechanism
- The comparison isolates the Z-number representation as a whole, so it cannot identify which component drives the changes in selected plans.The travel-time spread term favors predictable, short legs; Abu Dhabi’s baseline more often admits distant POIs, while the planner retains a compact set near the hotel base.Source: §7.2 · p. 9
7.3 The carbon weight
- The default score weight is 3.5 times the carbon weight, but the study does not characterize the score–carbon trade-off across carbon-weight settings.The implementation exposes the carbon weight as a request parameter, making a sweep a straightforward follow-up.Source: §7.3 · p. 9
7.4 Limitations
- The findings are scoped by estimated review counts outside Baku, one baseline with no ablation, fifteen-POI catalogs, and a single reported scenario.The pooled inference uses five city clusters, and the computational evaluation does not test whether tourists understand, trust, or follow the reliability annotations.Source: §7.4 · pp. 9–10
8 Conclusion
The paper presents a data-driven Z-number trip planner and compares it with a crisp-mean baseline across five cities. The planner lowers pooled trip carbon and travel time, but the study supports feasibility and favorable direction rather than general superiority.
8 Conclusion
- The planner lowers pooled trip carbon by 19.3% and pooled travel time by 11.5% versus a crisp-mean baseline, with both effects keeping their sign under a city-clustered bootstrap.Travel time falls significantly in Baku, Istanbul, and Abu Dhabi; neither metric changes in Tashkent.Source: §8 · p. 10
- The study establishes feasibility and favorable direction, not general superiority, given one baseline, estimated review counts in four cities, and fifteen POIs per city.The re-run reproduces original itineraries exactly in three cities; remaining differences change the sign of a negligible Istanbul carbon difference.Source: §8 · p. 10
- The planner constructs travel-time and rating Z-numbers from routing observations and review counts, then returns itineraries with per-leg reliability annotations through a conversational interface.Ablations, larger catalogs with measured review counts, a carbon-weight sweep, and a study of tourist responses are identified as next steps.Source: §8 · p. 10
Funding
The study was supported by the United Arab Emirates University through the UAEU-ASOIU Collaborative SDG Research Program Grant.
Funding
- The United Arab Emirates University supported the work through the UAEU-ASOIU Collaborative SDG Research Program Grant G00005329.The funder had no role in the study design, data collection or interpretation, writing, or submission decision.Source: Funding · p. 10