A travel-AI paper, explained from the beginning
Can AI guess where you will go next?
It can learn patterns from past check-ins. But in our independent reconstruction, the complex model did not consistently beat “suggest places you often visit.”
No coding or memorized equations needed. Start with a list of suggested stops. Finish able to explain the paper in your own words—and say what evidence is still missing.
Move the list-size slider. Numbers average 20 full-model runs on real NYC or Tokyo tests, counting targets already in the training catalogue. A hit means the next check-in is on the list; it does not mean a purchase or a satisfied traveller.
Translate the long title into an everyday question
You arrive in a new city. Your phone knows thousands of places, but you have only one afternoon. You want it to guess what you might visit, then arrange a route you can actually finish. The paper tries to connect these two jobs.
| Words in the title | What they mean here |
|---|---|
| Tourism consumer behaviour prediction | Use past records to predict the next place. The paper also discusses visit/booking scores, but check-ins themselves do not supply purchase labels. |
| Personalized recommendation | Rank places differently for different people, instead of giving everyone one popularity list. |
| Deep learning | Let a multi-layer model adjust its internal numbers using examples to reduce prediction errors. |
| Big data | Many behaviour records across people and time, plus ways to process them. More rows do not automatically mean better evidence. |
The paper calls its system TourRec-Stack. Think of a pipeline: organize records → learn travel patterns → rank places → check time and budget.
What does one row actually tell us?
A check-in records that a user was at a place at a time. A sequence of check-ins gives the activity order we learn from. It is not a complete travel spending diary.
| Data | Size | What it can tell us |
|---|---|---|
| NYC check-ins | 227,428 records; 1,083 users | Place, category, time and location; not every movement each user made. |
| Tokyo check-ins | 573,703 records; 2,293 users | Analyzed separately from NYC; user IDs cannot simply be merged across cities. |
| Hotel bookings | 119,390 records | Whether an existing booking was cancelled. There is no usable shared user identity linking it to the check-ins. |
The check-ins span April 2012 to February 2013. The two cities total 801,131 records, not 801,131 independent people: one person can contribute many rows. Stations, offices and homes also appear, so we cannot assume everyone is a tourist.
Understand the AI through three ideas
① Turn experiences into numbers, then remember their order
A computer represents a place using a learnable list of numbers called an . Think of a compact calling card. Individual numbers do not necessarily mean “likes coffee” or “likes quiet places.”
Next, pass the cards through a model in time order. Our reconstruction uses a , a model that updates a running summary after each event. It can distinguish “left the station to eat” from “finished eating and headed for a train.”
is like looking back at relevant notes: the model weights past records differently for a prediction. It must not peek into the future, and this does not mean it understands a person’s mind.
Go deeper: why use time and location?
The same café may mean something different before work and late at night. Hour, weekday, time gaps and spatial cells give the model possible clues. To test their value, compare with a version that removes them; a plausible story is not enough.
② Learning means adjusting a rule after mistakes
During training, hide the next stop, let the model guess, then reveal the recorded stop. A scores how wrong it was. An optimizer makes small changes to the internal numbers. After many passes, we hope it also predicts unseen later events better.
Our reconstruction uses three exercises: identify the next place from the whole catalogue; distinguish it from randomly chosen non-target places; and rank the recorded target above random places. Training tasks together is called multi-task learning.
Go deeper: what are CE, BCE and BPR?
CE penalizes low probability for the recorded next place. BCE penalizes incorrect binary decisions about recorded versus sampled places. BPR penalizes failing to rank the recorded target above a sampled non-target. Our loss is 0.5×CE + 0.5×BCE + 0.1×BPR—an explicit reconstruction, not recovered author code. A sampled negative may be a place the user visited before or visits later; it does not mean dislike.
Each point is a real validation score after a training epoch, showing one fixed run per city. More training is not guaranteed to help forever: memorizing practice questions can fail on new ones.
③ Separate practice, choosing a method and the final exam
The act like practice questions, a mock exam and a final exam. For each user, we use the earlier 70% of check-ins to learn, the next 10% to choose the training epoch, and the final 20% to report performance.
A large score is not enough: know what it measures
Imagine AI suggests ten places, and the recorded next stop is third. We can ask three different questions: is it on the list? How many listed places are correct for this event? How high is the correct one ranked?
| Metric | Plain meaning | In this experiment |
|---|---|---|
| HR@10 / Recall@10 | Did the top ten contain the next stop? | With one recorded next stop per test event: score 1 for a hit, 0 otherwise, then average. |
| Precision@10 | How many of the ten were correct for this event? | At most one relevant target per event, so the maximum is 1÷10 = 0.1. |
| NDCG@10 | When correct, how high is the target ranked? | Rank 1 scores 1; later ranks score less; ranks beyond 10 score 0. |
Remember that rewards putting the correct answer near the top. A score of 0.34 is neither “34% of people are satisfied” nor “34% of places were correct.”
Five options make a much easier test than thousands
The chart uses measured NDCG@5 from the first fixed seed in each city. The full catalogues contain 7,384 places in NYC and 11,233 in Tokyo. The sampled protocol keeps only the target and four random non-targets. With five candidates, the answer must be in the top five, so Recall@5 = 1 says nothing about model quality.
What did the experiments actually show?
The model learned patterns, but did not consistently beat a simple rule
A is a simple method used for comparison. Popularity gives everyone the same ranking. Personal frequency prioritizes places this person often visited in training. Complexity needs evidence from a fair comparison, not just an impressive architecture.
The full model was trained 20 times per city. Each other neural baseline and ablation was trained once per city; traditional baselines do not require this neural training. Main NDCG scores evaluate targets already in the training catalogue, not arbitrary new places.
Counting only known places gives a more favourable score
If the next place never entered the training catalogue, this model cannot recommend it. This is item . Only 63.46% of NYC test targets and 77.89% of Tokyo targets are covered. Counting the rest as misses lowers mean HR@10 from 50.49% to 32.04% in NYC, and from 52.34% to 40.76% in Tokyo.
It is better at predicting returns than first-time visits
Both groups use places in the training catalogue. “Not visited” means absent from this user’s training history, not unknown to the platform. Full-model HR@10 for these targets is about 1.66% in NYC and 6.87% in Tokyo. Predicting routines and helping visitors discover new attractions are different challenges.
Are the results reproducible—and how stable are they?
A neural network starts with random values and uses random training samples. Changing the is like varying a student’s starting point and practice order. We retrained 20 times per city rather than choosing the prettiest run. These are still the same data, not 20 independent city studies.
NYC NDCG@10 is 0.3261 ± 0.0072; Tokyo is 0.3436 ± 0.0021. The ± is the across 20 runs: how much retraining varies. It is not a maximum error or uncertainty around the paper’s number. Short lines in the chart mark medians, which differ from the means reported here.
We also used a to ask how the difference from the simple rule might change with the test-user mix. Model minus personal-frequency NDCG is about −0.0621 in NYC, with interval [−0.0689, −0.0553]; in Tokyo it is about −0.0006, with interval [−0.0063, +0.0058]. Under this protocol, NYC trails clearly. Tokyo’s small gap spans zero, so superiority is not established.
An removes a component for comparison: context, attention or a training task. A drop can suggest that component helps, but comparable settings and repeated runs matter. Some simpler versions performed better here, so we cannot claim every component is essential.
A good ranking still needs a feasible route
You may like three places that lie on opposite sides of the city. Ranking favourites first does not guarantee you can visit them in an afternoon. Route planning asks: can you afford it, fit it in and travel between the stops?
Greedy selection tries places in score order and adds them when feasible. Combinatorial optimization compares subsets and orders to maximize total score within the limits. Here it searches only the 12 supplied candidates, not the whole city; equal-score routes are not necessarily the shortest.
Why analyze hotel data separately?
Hotel data ask whether an existing booking will be cancelled—not whether someone will buy or where they will go next. We cannot join these datasets by arbitrary user numbers. We ran a separate cancellation task: the random forest’s mean AUC across 20 runs was 0.7795 ± 0.0042.
describes how well the model ranks a randomly selected cancelled booking above a non-cancelled one, with ties half-credited. 0.7795 does not mean 77.95% of bookings are classified correctly, nor that the model can recommend hotels.
Return to the paper with questions in mind
You do not have to struggle through the paper from start to finish. Follow this route and look for one answer at a time. Try your answer before opening the hint. Page references use the original 31-page PDF.
Stop 1 | Abstract, page 1: what is promised?
Find four jobs: learn spatiotemporal sequences, predict behaviour, personalize rankings and construct constrained routes. Mark the reported improvement as a claim to examine, not a fact proved by the abstract alone.
Stop 2 | Section 3.1 and Figures 1–2: do the inputs exist?
Check place, category, time, location, price and action type. Public check-ins support the first four. Linked real prices, clicks, cart events and purchases were not recovered. An input appearing in a diagram does not make it present in the dataset.
Stop 3 | Section 3.2 and Figure 3: how does history become a prediction?
Explain representation → sequential memory → attention → prediction heads. Then ask what labels train the visit/booking head. With sampled negatives, its output cannot simply be read as a real booking probability.
Stop 4 | Section 3.3 and Figure 4: why is ranking not an itinerary?
Look for constraints beyond preference scores, especially whether travel time enters the total. Our reconstruction always includes it. Having enough time at individual stops does not mean having enough time for the whole route.
Stop 5 | Section 4 and Table 5: were the scores calculated in the same way?
Check users, places, splits, candidates, relevant labels per example, metric definitions and baselines. If any is missing, a bigger number may be answering a different question.
Stop 6 | Figures 13–15 and the conclusion: how broad a claim is justified?
Separate repeated runs, correlations and survey satisfaction. Means and error bars do not uniquely determine run-level results or individual survey answers. We preserved readable published numbers without presenting them as newly collected observations.
Read the original paper or the publisher page. Title: Tourism consumer behavior prediction and personalized product recommendation based on deep learning and big data; Discover Artificial Intelligence, 2026.
How do data become charts?
1 | Preserve the source, then clean
Cross-check source versions. Removing exact duplicate check-ins leaves 800,304 events. Record user and event counts separately before and after cleaning.
2 | Split in time, then learn the catalogue
Split each user from past to future. Build the place catalogue and feature mappings from training data only, rather than collecting answers from the future.
3 | Train, select an epoch, keep the test separate
Training updates the model; validation selects the epoch. At the final test, record each target’s rank, evaluating each target once.
4 | Convert ranks into metrics and compare
For each target, check top-ten membership and rank, then average over targets. Plot all 20 runs, calculate means and variation, and show cold-start coverage and simple baselines too.
What can—and cannot—we conclude?
| Supported here | Not established |
|---|---|
| Public check-ins were recovered and the independent experiments can be recomputed. | The authors’ original experiments, private data and survey have been fully recovered. |
| The reconstruction learns some mobility patterns; the simple baseline is strong. | TourRec-Stack has been proved superior to all methods under matched conditions—or definitively disproved by this reconstruction. |
| Routes can satisfy explicit assumed budgets and time limits. | Routes are ready for deployment or already improve real traveller satisfaction. |
Recovering a figure differs from recovering an experiment. A mean over 20 runs does not reveal each run; average satisfaction from 1,000 people does not reveal individual answers. Approximate, clearly labelled points can redraw a curve, but cannot reconstruct unique raw data.
What should the next study improve first?
- Obtain the processed-data description, label definitions and evaluation method before claiming a matched comparison.
- Distinguish commuters from tourists, and separately test discovery of places a user has not visited.
- Use a common calendar cutoff for a more realistic future test, and repeat the main comparison models too.
- Use real roads, opening hours and costs. To study satisfaction, conduct a real user study and preserve the responses.
Can you explain the study to a classmate?
You do not need to recite terminology. Hide the page and complete four sentences: We want to solve…; Our data are…; We test fairly by…; We found… but cannot yet claim….
Example explanation | Rewrite it in your own words
This project predicts the next stop from past check-ins, then arranges candidate places into a route within time and budget. We recovered public NYC and Tokyo check-ins and split each person’s events into learning, selection and testing periods. The reconstruction learns some habits, but does not consistently beat suggesting frequently visited places, and is weaker on places new to that user. Missing original data and evaluation code prevent us from claiming recovery of the original experiment or improved traveller satisfaction.
Practise a defence: answer these four questions directly
Is this an original model or a reproduction?
It is an independent reconstruction of the paper’s ideas using recovered public sources. Added implementation choices are explicit; newly written code is not the authors’ original code.
Why are the scores so different from the paper?
No single cause is established. User counts, preprocessing, candidates, relevance definitions and implementation may differ. These must be aligned before a matched comparison.
Why is a simple baseline necessary?
Without one, we cannot tell whether the model adds useful patterns or mainly captures repeated visits.
What evidence would you seek next?
First obtain checkable data provenance and evaluation definitions. If tourism recommendation is the goal, next test discovery, real route feasibility and traveller feedback.
Return here whenever a term is unfamiliar
Hover or tap the marked terms in the lesson. The glossary below is the same reference; there is no need to memorize it all.
Where do the numbers on this page come from?
Measured charts come from the completed independent reproduction; every training-run point corresponds to an actual run. Formula and timeline demonstrations are labelled teaching examples, and itinerary times and costs are labelled assumptions. This page was written with AI assistance and does not replace the student’s own reading, understanding and checking of the paper.