DeepTravel: Neural Models for Travel Planning

• DeepTravel is a suite of travel intelligence systems that integrate neural travel time estimation, personalized recommendation, and autonomous itinerary planning.
• The framework employs grid-based sequence modeling, dual interval loss, and reinforcement learning to capture spatial, temporal, and behavioral dynamics.
• Empirical results indicate significant performance gains over traditional baselines, enhancing metrics such as MAE, RMSE, and itinerary pass rates.

DeepTravel is a name used in arXiv literature for several technically distinct travel-intelligence systems. In its original and most specific usage, it denotes an end-to-end neural model for estimating the travel time of an entire path from sequences of spatial grids and intermediate trajectory timestamps (Zhang et al., 2018). In later work, the term is also associated with itinerary-aware deep matching for online travel platforms, where FitNET is presented as exemplifying DeepTravel through intent modeling over unconsumed orders (Xu et al., 2021), with an end-to-end agentic reinforcement learning framework for autonomous travel planning agents (Ning et al., 26 Sep 2025), and with a sequential RL travel-planning baseline used in controlled comparisons against hierarchical multi-agent planners (Bui et al., 5 Mar 2026). Taken together, these usages locate DeepTravel at the intersection of travel time estimation, personalized recommendation, and autonomous itinerary generation.

1. Terminological scope and research usage

The literature does not treat DeepTravel as a single canonical architecture. Rather, the name appears in multiple research contexts with different problem formulations, inputs, and learning objectives. In the 2018 usage, DeepTravel is a path travel-time estimator over grid sequences with auxiliary supervision from intermediate timestamps. In the recommendation setting summarized in 2021, FitNET is described as operationalizing DeepTravel through itinerary-aware intent modeling at Fliggy. In the 2025 usage, DeepTravel becomes an agentic RL framework for tool-using travel planning. In the 2026 comparison study, DeepTravel denotes a sequential, monolithic LLM planning baseline in long-horizon constrained travel planning (Zhang et al., 2018, Xu et al., 2021, Ning et al., 26 Sep 2025, Bui et al., 5 Mar 2026).

Usage	Primary task	Core mechanism
DeepTravel (2018)	Travel time estimation	Grid sequence modeling with BiLSTM and dual interval loss
FitNET as DeepTravel	Travel recommendation and intent modeling	Itinerary-aware attention and multi-objective deep matching
DeepTravel (2025)	Autonomous travel planning	Agentic RL with sandbox tools, hierarchical verification, and replay
DeepTravel baseline (2026)	Sequential constrained itinerary planning	Monolithic RL agent over a single growing context

A common thread across these usages is end-to-end learning over travel-specific structured signals rather than purely heuristic assembly. This suggests that “DeepTravel” functions less as a stable product name than as a recurring label for deep models that place travel-domain structure at the center of representation learning or decision making.

2. Grid-based path travel-time estimation

The 2018 DeepTravel model addresses two limitations in prior travel-time estimation. First, segment-based methods estimate the time cost of each road segment and are not able to capture many cross-segment complex factors such as junction delays, turning penalties, and correlated congestion. Second, heuristic sub-path aggregation fails to utilize abundant temporal labels in trajectory data, namely the time stamp of each trajectory point. DeepTravel therefore estimates the travel time of an entire path directly and learns from intermediate timestamps rather than only from a final path-level regression target (Zhang et al., 2018).

Formally, the path is represented as $P=\{e_1,e_2,\dots,e_n\}$, with trajectory points $T=\{p_1,p_2,\dots,p_m\}$ and total travel time $T(P)=t_m-t_1$. To fit neural architectures and unify feature extraction, the map is partitioned into $N\times N$ equal-sized grids, and a path is represented as a sequence of traversed grids $G=\{g_1,g_2,\dots,g_n\}$. Grids without sampled GPS points are kept to maintain spatial continuity. This grid abstraction is central: it allows the model to represent a path as a variable-length sequence while preserving route continuity even under sparse sampling.

For each grid, DeepTravel concatenates a feature vector

$V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$

The spatial embedding $V_{sp}\in\mathbb{R}^d$ is a learned vector tied to grid location and intended to capture spatial characteristics such as speed limits, road structure, and congestion propensity. The temporal embedding $V_{tp}\in\mathbb{R}^d$ represents time-of-day effects through learned embeddings over time bins. The driving-state vector $V_{dri}\in\mathbb{R}^4$ contains three binary indicators for starting, middle, and ending stages plus a scalar proportion of the path already traveled. Short-term traffic features $V_{short}$ are constructed from 5-minute time-bins over the hour before the query time, with each bin summarized as $T=\{p_1,p_2,\dots,p_m\}$0 and processed by an LSTM. Long-term traffic features $T=\{p_1,p_2,\dots,p_m\}$1 are built analogously across previous days at the same time-of-day, capturing weekly periodicity. Both short-term and long-term traffic features are augmented with neighbors through rings such as $T=\{p_1,p_2,\dots,p_m\}$2, and in experiments a neighborhood radius up to $T=\{p_1,p_2,\dots,p_m\}$3 was used.

Prediction is performed by a BiLSTM over the sequence of per-grid feature vectors. The hidden state at position $T=\{p_1,p_2,\dots,p_m\}$4 is $T=\{p_1,p_2,\dots,p_m\}$5, so the model has both start-to-current and end-to-current context. A shared linear head maps hidden states to local time contributions, and a cumulative summation mechanism aligns the prediction layer with the additive nature of travel time over sub-paths. This architecture is designed to capture cross-grid dependencies directly rather than relying on post hoc aggregation of per-segment estimates.

3. Auxiliary supervision, optimization, and empirical performance

A distinctive element of DeepTravel is its auxiliary supervision mechanism. Instead of using only the final target $T=\{p_1,p_2,\dots,p_m\}$6, the model extracts supervisory signals from grids that contain GPS samples. For each such grid $T=\{p_1,p_2,\dots,p_m\}$7, the leave time $T=\{p_1,p_2,\dots,p_m\}$8 is derived from trajectory points in that grid, and a mask vector $T=\{p_1,p_2,\dots,p_m\}$9 marks positions with available supervision. The forward and backward interval targets are defined as $T(P)=t_m-t_1$0 and $T(P)=t_m-t_1$1. Predicted forward intervals are produced by cumulatively summing hidden states from the start to grid $T(P)=t_m-t_1$2, while predicted backward intervals are cumulatively summed from grid $T(P)=t_m-t_1$3 to the end. The dual interval loss is the mean relative squared error over all supervised grids, masked by $T(P)=t_m-t_1$4, and uses a shared linear layer for both directions so that the per-grid transformation from hidden state to local time contribution is direction-consistent (Zhang et al., 2018).

The paper attributes three benefits to this auxiliary supervision. It augments supervision without introducing new data, provides stronger and more localized gradients for long sequences, and aligns naturally with BiLSTM’s bidirectional structure. Training minimizes the sum of dual losses over trajectories, learning network parameters and embeddings jointly with backpropagation-through-time. Reported hyperparameters include an 8:1:1 train/validation/test split, spatial and temporal embedding sizes of 100 initialized uniformly in $T(P)=t_m-t_1$5, 100 hidden units in traffic-feature LSTMs and in the BiLSTM prediction layer, Adam with initial learning rate $T(P)=t_m-t_1$6, and weight initialization in $T(P)=t_m-t_1$7.

Evaluation uses two real datasets. Porto contains 420,000 passenger trips from the Kaggle Porto taxi data, with 15s sampling interval, area $T(P)=t_m-t_1$8, and $T(P)=t_m-t_1$9 grids; its mean travel time is $N\times N$0 with standard deviation $N\times N$1. Shanghai contains 1,018,000 passenger trips from April 1–17, 2015, with 10s sampling, area $N\times N$2, and $N\times N$3 grids; its mean travel time is $N\times N$4 with standard deviation $N\times N$5. Metrics are MAE, RMSE, and MAPE. On Porto, DeepTravel reports MAE $N\times N$6, RMSE $N\times N$7, and MAPE $N\times N$8; on Shanghai, it reports MAE $N\times N$9, RMSE $G=\{g_1,g_2,\dots,g_n\}$0, and MAPE $G=\{g_1,g_2,\dots,g_n\}$1. The best sub-path baseline, PTTE, reports Porto MAE $G=\{g_1,g_2,\dots,g_n\}$2 and Shanghai MAE $G=\{g_1,g_2,\dots,g_n\}$3; the best segment baseline, spd-LSTM, reports Porto MAE $G=\{g_1,g_2,\dots,g_n\}$4 and Shanghai MAE $G=\{g_1,g_2,\dots,g_n\}$5. The reported conclusion is that DeepTravel significantly outperforms all baselines on both datasets across all metrics.

Ablation results support both the feature design and the loss design. Using only spatial and temporal embeddings yields Porto MAE $G=\{g_1,g_2,\dots,g_n\}$6 and Shanghai MAE $G=\{g_1,g_2,\dots,g_n\}$7; using only short- and long-term traffic features yields Porto MAE $G=\{g_1,g_2,\dots,g_n\}$8 and Shanghai MAE $G=\{g_1,g_2,\dots,g_n\}$9; combining them improves performance, and adding driving-state features reaches the full model result. Loss ablations show that auxiliary supervision accelerates convergence and improves accuracy, with dual forward-plus-backward supervision outperforming forward-only and backward-only variants. The paper also notes a characteristic error mode: performance degrades for rare abnormal cases such as short paths with unusually long durations caused by sudden congestion or prolonged stops, which lack predictive signatures in recent history. Reported limitations include the grid abstraction, sensitivity to GPS noise and trajectory sparsity, dependence of auxiliary supervision on grid-level GPS points, and the likely need to adapt embeddings and hyperparameters to new cities.

4. Itinerary-aware personalization and the broader DeepTravel interpretation

In the recommendation literature, the 2021 Fliggy system FitNET is presented as itinerary-aware personalized deep matching, and the technical summary explicitly frames it as a manifestation of DeepTravel in travel personalization. The central concept is the itinerary $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$0, defined for a user $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$1 as the list of the user’s unconsumed orders deemed to correspond to a travel. Modeling the itinerary as a whole addresses three major challenges for online travel platforms: sparsity, diversity, and implicitness. The claim is that travel intention is rarely captured by one order; rather, multiple orders jointly signal intent, such as outbound ticket plus hotel for sightseeing or a return ticket for transfer or return (Xu et al., 2021).

FitNET builds user and item representations with itinerary-aware attention and matches them by inner product. Inputs comprise user basic profile $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$2, user itinerary $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$3, user behavior sequence $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$4, and target item features $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$5. The embedding layer handles one-hot and multi-hot features, and category-level embeddings are concatenated to produce $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$6, $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$7, order embeddings $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$8, and behavior embeddings $V=[V_{sp},V_{tp},V_{dri},V_{short},V_{long}].$9. The first attention mechanism lets the user profile query the itinerary: with learnable projection $V_{sp}\in\mathbb{R}^d$0, attention weights are

$V_{sp}\in\mathbb{R}^d$1

and the intent vector is $V_{sp}\in\mathbb{R}^d$2 with $V_{sp}\in\mathbb{R}^d$3. A multi-head self-attention layer over the itinerary consolidates preference signals across orders, and a second scaled dot-product attention filters the user’s historical behavior sequence conditioned on the itinerary-driven intent and preferences. Three MLPs then produce sightseeing probability, user representation, and item representation, with click score $V_{sp}\in\mathbb{R}^d$4.

Training is multi-objective. One loss predicts the itinerary context’s travel intent, specifically sightseeing versus other; the second predicts user click behavior. The combined loss is $V_{sp}\in\mathbb{R}^d$5. The study further evaluates negative-sampling strategies and finds that the best setting constrains destination to the itinerary’s destination, uses 10% negatives with the same category as the positive and 90% with different categories, and adopts a 1:10 positive:negative ratio. Dataset scale is industrial: the April 2020 Fliggy production data contain 12.1M training samples, 0.27M users, 1.53M items, 0.73M orders, and 0.34M itineraries, with 3.4M test samples. Offline metrics are HitRate@k and Precision@k. Under the best setting, FitNET reports HitRate@10 $V_{sp}\in\mathbb{R}^d$6 and Precision@10 $V_{sp}\in\mathbb{R}^d$7, outperforming MIND by average hit rate $V_{sp}\in\mathbb{R}^d$8 and average precision $V_{sp}\in\mathbb{R}^d$9, and outperforming a variant using only the most recent order by average hit rate $V_{tp}\in\mathbb{R}^d$0 and precision $V_{tp}\in\mathbb{R}^d$1. An online A/B test over 7 days reports that FitNET’s average CTR was 1.3% higher than the next best method, MIND, and the system is reported as deployed at Fliggy.

This usage broadens the meaning of DeepTravel beyond mobility prediction. Here the central travel-specific object is not a path or trajectory but an itinerary composed of unconsumed orders, and the core modeling move is to make current travel intent the query that reweights both itinerary items and prior behaviors. The stated limitations remain familiar: cold-start users and items are still challenging, some travels contain few unconsumed orders, and transfer to other travel ecosystems may require re-engineering of features and labels.

5. Agentic reinforcement learning for autonomous travel planning

The 2025 system titled “DeepTravel: An End-to-End Agentic Reinforcement Learning Framework for Autonomous Travel Planning Agents” formulates travel planning as a multi-turn MDP with tool-mediated transitions. The state $V_{tp}\in\mathbb{R}^d$2 contains the user query, dialogue context and thoughts, partial itinerary, prior actions, tool responses, and verifier feedback. Actions include plan, tool_call(tool, params), reflect, revise, and finalize itinerary; tool calls cover flight_search, train_search, route_planning, hotel_search, poi_search, and web_search. The RL objective is

$V_{tp}\in\mathbb{R}^d$3

with episodes terminating on finalization or at max-turn $V_{tp}\in\mathbb{R}^d$4 (Ning et al., 26 Sep 2025).

A central engineering contribution is the sandbox environment. A daily-refreshed database caches flight, train, and hotel results and supports route, POI, and web lookups; on-demand caching stores latest query-result pairs for re-access. This removes QPS constraints and stabilizes responses across training. Tool responses are schema-constrained and reproducible for a given cached query. Tokens inside <tool_response>...</tool_response> are masked during training so the policy is trained only on agent-generated tokens.

Reward modeling is hierarchical. A trajectory-level verifier checks spatiotemporal feasibility, budget, spatial continuity, and hard user requirements such as required POIs or arrival-before constraints. A turn-level verifier then validates whether itinerary details are consistent with tool responses at the level of parameters, evidence bindings, names, times, routes, and prices. In practice, the implementation uses strict binary pass/fail: if any trajectory-level constraint fails, reward is zero and turn-level verification is skipped; if trajectory-level passes, all turns must pass for reward one. Human-authored rubrics guide a DeepSeek-R1-based verifier at both levels.

Policy optimization uses a GRPO-style clipped surrogate with KL anchoring to a reference policy, combined with reply-augmented reinforcement learning. During rollouts, if none of the $V_{tp}\in\mathbb{R}^d$5 sampled trajectories for a query passes verification, the query is appended to a failure experience buffer $V_{tp}\in\mathbb{R}^d$6. After a fixed training-step interval, queries are replayed from $V_{tp}\in\mathbb{R}^d$7 with the current policy. The paper argues that replaying hard queries after policy improvement enables correction behaviors that do not emerge from purely on-policy sampling.

Training uses Qwen3-8B and Qwen3-32B backbones. A cold-start SFT phase uses 1K verified tool-integrated trajectories distilled from DeepSeek-R1 with batch size $V_{tp}\in\mathbb{R}^d$8, learning rate $V_{tp}\in\mathbb{R}^d$9, and $V_{dri}\in\mathbb{R}^4$0 epochs. RL then uses 500 high-quality samples, with 450 for training and 50 for validation, rollout size $V_{dri}\in\mathbb{R}^4$1, learning rate $V_{dri}\in\mathbb{R}^4$2, max response length $V_{dri}\in\mathbb{R}^4$3K tokens, and max turns $V_{dri}\in\mathbb{R}^4$4. Evaluation includes 6,224 real user queries from DiDi Enterprise Solutions and 1,000 synthesized offline queries. The primary metric is Final Pass Rate. DeepTravel-32B-RL reports offline Final Pass Rate of 69.34% on easy no-constraint queries, 54.74% on medium, 29.17% on hard, and 73.21%, 62.22%, and 35.75% respectively on easy, medium, and hard constrained queries; its online score is 62.77%, with 82% on a 50-case human evaluation. DeepTravel-8B-RL reports online 49.75%, improving over base Qwen3-8B from 26.72% by +23.0 points. The paper states that the 32B and 8B variants outperform OpenAI o1, o3, and DeepSeek-R1 on the reported travel-planning tasks.

Ablations identify several necessary components. Removing experience replay drops online performance from 49.75% to 40.00% for Qwen3-8B; removing cold-start drops it to 32.45%; removing the trajectory-level verifier yields 26.52%; removing the turn-level verifier yields 32.45%, with hard splits suffering especially severely. Reported limitations include reliance on carefully authored rubrics, dependence on cached APIs, and weaker reward gains when training directly against live noisy tools. The safety discussion notes that RL reduced factual errors from approximately 50% to below 20% in a 50-case user study, and that no PII was stored in deployment.

6. Sequential DeepTravel as a long-horizon planning baseline

The 2026 HiMAP-Travel study re-implements DeepTravel as a sequential, monolithic LLM agent trained via GRPO for controlled comparison against a hierarchical multi-agent framework. In this setting, DeepTravel generates the full itinerary token-by-token in a single context, day after day, without hierarchical decomposition, global transactional enforcement, or parallelism. The policy is written as $V_{dri}\in\mathbb{R}^4$5, where all tool outputs and reasoning accumulate in one growing history. Constraint tracking is self-maintained in text and evaluated post hoc; there is no external atomic constraint monitor for budget, diversity, or transportation consistency (Bui et al., 5 Mar 2026).

The study’s central diagnosis is “Constraint Drift under Long Tool Traces.” As travel horizon lengthens, attention to initial global constraints degrades. Reported evidence is that budget adherence in sequential baselines drops from about 98% on Day 1 to about 42% by Day 5. Observed failure modes include constraint drift, cascading failures from early overspending, high variance across random seeds, and latency that scales linearly in the number of days, $V_{dri}\in\mathbb{R}^4$6. Under identical backbone, training, tools, decoding, and TravelPlanner harness, DeepTravel achieves validation Final Pass Rate of 45.56% and test Final Pass Rate of 43.98% for Qwen3-8B, with test standard deviation 7.18%; the corresponding Qwen3-4B scores are 35.28% validation and 34.92% test. In the same controlled setup, HiMAP-Travel reaches 52.65% test Final Pass Rate on Qwen3-8B, or +8.67 percentage points over DeepTravel, and reduces variance from 7.18% to 0.48%.

The comparison is not uniformly negative. The same study notes that DeepTravel’s commonsense micro score is already high at 93.28%, close to HiMAP’s 94.62%; it uses fewer tokens than HiMAP on 7-day trips, approximately 12.3k versus 14.2k; and its runtime stack is simpler, with fewer moving parts. The paper therefore identifies conditions under which the sequential DeepTravel formulation remains appropriate: short itineraries of 1–3 days with few constraints, environments where external monitors or multi-agent orchestration are infeasible, and cases where simpler deployment is preferred over proactive correctness-by-construction.

Across these usages, DeepTravel denotes a set of travel-domain deep-learning paradigms rather than a single immutable model. In travel-time estimation it is a grid-sequence BiLSTM with auxiliary interval supervision; in itinerary-aware recommendation it becomes attention-based deep matching over unconsumed orders; in autonomous planning it is an agentic RL system with sandboxed tools and hierarchical verification; and in long-horizon comparisons it serves as the sequential baseline against which transactional and hierarchical designs are measured. The shared pattern is the use of domain-structured representations—paths, itineraries, tool traces, or day-level plans—to replace purely heuristic aggregation with learned inference over travel-specific constraints and signals.