Machine Learning-based prediction of travel destinations based on multivariate parameters

Travellers rarely struggle to recognise a destination that appeals to them. They struggle to describe it. A conventional search interface demands precisely what an undecided traveller cannot supply, namely explicit criteria, and the result is a filtered list that reflects the vocabulary of the query rather than the intent behind it. This thesis approaches the problem from the opposite end. The search form is replaced by a sequence of paired travel images, and preference is inferred from the choices a user makes.

A convolutional neural network assigns each image to one of nine travel categories derived from point-of-interest data. As the comparisons accumulate, so does an interest profile, which is matched against 29 destinations and resolved into three ranked suggestions, each reported with the degree to which it corresponds to the profile. The intention is not to answer the question for the traveller, but to give the subconscious preferences behind it something concrete to react to.

Results

Nine participants completed three runs each, giving 27 trials. In 19 of them at least one of the three proposed destinations was judged attractive, an approval rate of 70.37%. That figure invites a reference point: drawing three destinations at random from a pool of 29 would surface a relevant one in 10.34% of cases.

Rome proved the most frequent first-ranked suggestion at 25.93% of runs, with Istanbul close behind at 22.22%. Measured across all three positions rather than the first alone, Istanbul leads at 16.05%. Seven destinations never surfaced at any point.

The pattern is instructive. Destinations whose category values are evenly distributed dominate the results, which points to an asymmetry in the source data rather than to any intended behaviour of the model.

Approach

The development follows the CRISP-DM process model across five stages.

1. Point-of-interest data is acquired for 29 global destinations, among them Rome, Istanbul and New York, with TripAdvisor serving as the source.
2. The resulting properties are condensed from 45 to nine categories in KNIME, with entropy used to establish how much each category genuinely discriminates between destinations.
3. A convolutional neural network is trained on roughly 1,000 images per category, so that a photograph of a museum resolves to “Culture & Education”.
4. The user is presented with pairs of generic travel images. The categories behind each selection accumulate into an interest profile.
5. Destinations are scored against that profile and the three strongest matches are returned, after which the user rates the quality of the prediction.

Model and tooling

Component	Implementation
Framework	Keras
Input	300 x 300 px images, 80/20 train/test split
Architecture	4 x (Conv2D, MaxPooling2D, Dropout 0.22), Flatten, 3 Dense layers (512 units first)
Activation	ReLU
Optimiser / loss	Adam / binary cross-entropy, accuracy as metric
Hardware	GeForce RTX 2080 SUPER, 60-90 s per epoch, approximately 35 min per training run
Data processing	KNIME

A constraint satisfaction formulation was weighed as an alternative methodology. The machine learning route prevailed on grounds of extensibility: introducing further destinations or categories requires only another training run, whereas a CSP equivalent would demand that its variables and constraints be redefined.

Abstract

When choosing a travel destination, fundamental decisions must always be made. For example, which destinations are interesting? Which activities might be interesting to do? Especially an indecisive user may find it difficult to formulate these conditions precisely. How can a relevant travel destination still be found and how can a user be supported in his decision-making process by use of computerized means? To solve this problem, this thesis devises a dual image comparison, in which a user is presented with a series of generic travel images. These images are classified by CNN into several possible travel categories, which leads to a user implicitly deciding whether e.g. he is interested in sightseeing tours or wellness trips, or which destinations are of interest to him in general. The destinations themselves are evaluated based on information of POIs, which allow conclusions to be drawn about the characteristics and distribution of possible travel activities. The user input thus creates an interest profile, which is compared with possible destinations. Based on the calculated match rate, destinations are suggested to the user. The user approval rate of the displayed results is measured at 70.37%.

@mastersthesis{golombeck2020travel,
  author   = {Golombeck, Marius},
  title    = {{Machine Learning-based Prediction of Travel Destinations Based
              on Multivariate Parameters: Conception and Development}},
  school   = {University of Applied Sciences and Arts Dortmund},
  address  = {Dortmund, Germany},
  type     = {Master's Thesis},
  year     = {2020},
  month    = nov,
  language = {english},
  doi      = {10.13140/RG.2.2.23262.25923/1},
  url      = {https://www.researchgate.net/publication/345431946}
}