The technology behind the solutions: inside Droplabs AI systems
This post is part of a larger article: AI in the service of heritage: Droplabs Dynamic Pricing and attendance prediction in shaping sustainable tourist behavior in Tatra National Park. Part one describes the business side of attendance forecasting - read it here!
TPN uses the AI-based Attendance Forecasting engine provided by Droplabs as part of its online sales system with Dynamic Pricing. Attendance forecasting improves TPN's management efficiency: it helps plan Park Service (patrol) schedules and the optimal deployment and preparation of staff at individual parking lots. It also helps TPN tailor information messages for visitors on an ongoing basis.
As outlined in part one, the TPN parking complex, the gateway to Morskie Oko, is a multi-element infrastructure. It includes the main parking lot at Palenica Białczańska (with a capacity of about 600 vehicles), a complementary roadside lane between Palenica Białczańska and Łysa Polana, and a paved parking lot at Łysa Polana.
A look at the technology
It is worth stressing that the AI mechanism developed by Droplabs is based on machine learning models, the foundation of artificial intelligence. It does not use the mechanisms most commonly associated with AI today, such as the large language models (LLMs) known from chatbots (e.g., Gemini or ChatGPT).
To look under the hood, let's start with the stage where work on a machine learning model begins: data mining. Among other tools, the scatter plots mentioned in part one are useful here.
Scatter plots allow us to refute hypotheses about simple relationships between weather and attendance.
Below are further examples showing attendance at the Morskie Oko parking lot in relation to individual weather features. The black lines on the charts show hypothetical linear trends, whose correlation is measured by the Pearson coefficient. Plotting them does not, of course, mean that customers' actual behavior is linear; this analysis disregards other types of correlations and dependencies, which are examined with other methods in later steps. It is also worth remembering that these are only selected visualizations: the full analysis takes into account features from many different data sources, not just weather forecasts.
Just as scatter plots help refute hypotheses about simple relationships between weather and attendance while revealing the characteristics of the data, another method used in the early stages of building predictive models, cluster analysis, helps capture different customer behaviors, non-obvious relationships and similarities between dates, and unusual data points.
“Groups 1 and 3 shown in the chart, which correspond to planned and unplanned trips in the summer season respectively, are the days with the highest attendance of the year. However, their purchasing patterns differ significantly: on planned days, about 80% of tickets are bought at least one day in advance, while on unplanned days this figure falls below 60%.
The analysis shows that this difference is driven mainly by weather conditions. Customers tend to plan visits on days with a favorable forecast. Interestingly, on days with a less favorable forecast, attendance does not drop significantly, but customer behavior changes fundamentally: people become more spontaneous and respond to current conditions (e.g., when the expected rain turns out to be just morning drizzle).”
- adds Michał Piłat, Data Scientist at Droplabs
Model architecture and training
Once the characteristics of the data are understood, the next stage is choosing the model and the optimal architecture for a specific time horizon. At this stage, many different machine learning models and data configurations are tested for each prediction horizon. Model performance is evaluated using a so-called confusion matrix.
Based on the performance analysis, an architecture from the deep neural network family was chosen as the final solution.
Deep learning models are not the standard approach for this type of data. Research and process optimization by the Droplabs team led to the choice of a custom architecture that combines deep learning with ensemble learning, an intelligent combination of many modules. In this way, Droplabs creates a “team of experts” in which each model specializes in a different area, and the final decision is based on the conclusions of all individual components.
This architecture gives Droplabs a universal solution that can be adapted to the individual needs of a given site, as identified during data mining.
Once the model architecture is chosen, the next important stage begins: training. It is important to understand that predicting attendance for the next day is a different analytical problem from forecasting 3 or 7 days ahead. Different situations come with different sets of key data. That is why each prediction horizon is treated as a separate model that must be optimized independently. The first task is to determine the optimal parameters and select features. To do this, Droplabs carries out a detailed sensitivity analysis. This analysis precisely determines how strongly individual features influence the final prediction. It can be performed both for specific cases and for the entire data set. The result is a list of features ranked by their influence on the prediction, which allows each model to be tuned optimally.
Another critical step is selecting and calibrating the so-called hyperparameters. These are structural parameters of the model that are set before training begins and do not change during it, but fundamentally affect its performance. They include settings closely tied to the model architecture, such as the number of neural network blocks or the size of the input layer. To achieve maximum performance for a specific prediction horizon, hundreds of experiments are carried out. This iterative optimization process aims to find the configuration that maximizes the model's performance.
Once the hyperparameters are selected and the model is trained, it must be validated, that is, checked to confirm that it achieves the expected level of precision and accuracy. This process is visualized with a chart of actual vs. predicted values.
Monitoring and maintenance (Machine Learning Engineering)
It may seem that the work ends here, but in fact this is the beginning of an endless cycle designed to deliver high-quality predictions every day. After training, the model goes through an automated process of validation, versioning, and deployment to the production environment, where the solution becomes available to customers. The model is packaged as a service (e.g., a REST API), which allows it to communicate with other parts of the Droplabs system, such as the sales system. This process automatically runs predictions, saves logs, and monitors model performance. As one of the final results, the client receives attendance predictions as a visualization in a clear web panel.
The panel shows the predicted attendance ranges (purple zone) and the actual number of visitors (blue dots) for past dates. This allows TPN to verify the system's historical accuracy.
After deployment comes continuous monitoring and maintenance. This includes technical monitoring (server performance) and monitoring of prediction quality. It is essential to track both data drift and model drift.
“Changes such as a modified ticket offer, new marketing campaigns, or shifts in tourist preferences mean that, to keep working optimally, the model needs periodic retraining so that its quality always stays above a satisfactory threshold.”
- adds Piotr Pyznarski, Droplabs founder responsible for product and technology (CTO)
Summary
Effective attendance prediction requires AI engineering tailored to the specifics of the attractions and tourist destinations industry. A breakthrough for this industry is Droplabs' use of an architecture based on deep neural networks that works as a team of models (ensemble learning).
“This approach enables Droplabs to create solutions that adapt flexibly to specific tourist attractions, including Tatra National Park. Thanks to in-depth data mining specific to each site, we are able to deliver accurate attendance predictions.”
- notes Monika Karpierz-Nałęcz, Head of AI at Droplabs
At Tatra National Park, the expected attendance range for the next day is correct in about 90% of cases.
The process of creating, monitoring, and continuously improving models is complex and requires proprietary solutions. A team of Droplabs specialists from various fields, including Data Science, Machine Learning Engineering, DevOps, Data Analytics, and Backend Engineering, makes sure that high-quality predictions appear in the TPN panel every day.
This post is part of a larger article: AI in the service of heritage: Droplabs Dynamic Pricing and attendance prediction in shaping sustainable tourist behavior in Tatra National Park. The next part shows how Droplabs Dynamic Pricing helps shape visitor behavior.
