bibradar

Inspiration

Like many other students, we quite frequently found ourselves in an annoying situation: Taking a 30-minute train ride to a university library just to realize it is fully occupied. This problem is especially critical during the exam phase. Since this is a real hassle for many students, we wanted to find a solution which makes life easier for them!

So we built a simple website which offers easy but meaningful occupancy information on all libraries. It helps students by calculating which library is most likely to still have free seats by the time they arrive there.

What it does

We kept it as frictionless as possible: The website allows you to provide your location (it doesn't have to be exact) and receive an immediate suggestion for the best library. The result is based on your travel time to the library, as well as a prediction of the occupancy of the library at the time you arrive. Alternative options are provided to you ranked from best to worst.

How we built it

The first step was collecting occupancy information of the libraries. We used the MWN WLAN Statistics, as an indicator for occupation, since everybody has a phone in their pocket nowadays, which automatically connects to the eduroam or BayernWaln Wi-Fi. After some research on which access point belongs to which library, we were able to collect one year of occupancy data on most of the major university libraries in Munich.

We then continued by applying general data science approaches to formulate a mathematical representation of the problem. With the insights gained from that analysis, we decided to train a transformer-based model, based on a paper published by the University of Oxford and Google Cloud AI research in 2021, to predict future occupation data.

This model fitted our use case because it can capture time-varying relationships over different time intervals and had a satisfying long-term forecasting accuracy.

The predictions of the model are combined with other factors, like the distance of the user to the library, to determine which library would be best suited for the user. This scoring is especially focused on making sure that the library still has free spots when the user arrives there.

Challenges we ran into

The first issue we had to overcome was the unintuitive MWN WLAN API. Figuring out which access points are located in which library took a significant amount of time.

After that, we spent a lot of time on the machine learning part of the project. We had to try out different approaches and models before finding a suitable solution.

Another challenge was the training of a separate model for every library due to limitations in GPU infrastructure and time. For that reason, the model was trained solely for the library ‘Philologicum’ for now, but can be easily rolled out for all other libraries.

Accomplishments that we're proud of

After some long hours, we were delighted that we were able to assemble an easy-to-use aggregation of all the MWN WLAN access point usage information for the libraries.

Also, training our own transformer-model was a major achievement for our team! We were pleased that it worked out, and we could deploy it.