ASWI - Pokročilé softwarové inženýrství » ASWI 2021 » OpenData vlastní téma (KIV) - Tři Mušketýři

h1. Application predicting class attendance based on weather

h2. WARNING! THIS INFORMATION IS NOT UP TO DATE, FOR INFORMATION ABOUT CURRENTLY DEVELOPED APPLICATION CONSULT [[Product Vision Statement]]

h2. Data

The application uses datasets containing historical information about the weather on campus and attendance based on JIS card verifications and historical timetable information

* JIS data http://opendata.zcu.cz/Snimace-JIS.html

* Timetable data http://opendata.zcu.cz/Obsazeni-mistnosti.html

* Weather data http://opendata.zcu.cz/Energeticky-dispecink.html

* Authentication system http://opendata.zcu.cz/Autentizacni-system.html

The timetable provides the baseline attendance, JIS card data can be used to find out the real attendance (and therefore fluctuation in attendance). The authentication data could be used to find out how many people have logged in during classes.

h3. Weather

Weather data for model training contains the following information: date, temperature, wind, rain, light in k lux

User can input weather values manually, by selecting options from a form

* values: temperature, wind, rain, sunny/overcast/partly cloudy

There also is an option for data to be automatically downloaded from a server upon request from the user.

Current options for data sources

* RSS from yr.no http://www.kanonbra.com/rss/yr_forecast_rss.php?language=en_US&location=%C4%8Ceahkka/Plze%C5%88sk%C3%BD_Kraj/Plze%C5%88

* JSON from wttr.in http://wttr.in/Plzen,czechia?format=j1

App provides the ability to ask for data from today or days in the future, and give hourly time specifics (this time will be then approximated into intervals of 3h as is data provided by the website)

Again the app will use the following values: temperature, wind, rain, sunny/overcast/partly cloudy

Notes:

Values sunny/partly cloudy/overcast/dark are approximated from data input in lux values.

The temperature will be evaluated with certain tolerance (probably determined experimentally)

h3. Datasets

Datasets are updated every once in a while with new data. How often update?

-> has to trigger re-training the prediction model, probably do not want to update too often.

Detect when there's new data, check every day? every x days?

New OpenData is added monthly. How complex will be the training? Can it be done on the target device? Will there be a server from which a new model could be downloaded?

Notes:

Should exclude data from distance learning periods.

h2. Design

h3. The output of the Algorithm

We need to determine what should be the actual prediction:

* A specific number of students in a class,

* information about relative occupancy rate.

h2. Motivation

Attendance prediction could be used to deliver crucial course info and material. The lecturers could use this information ahead of seeing the real attendance to adjust the material or schedule important events. Attendees could be motivated with bonus attendance points on slow days, announced whenever low attendance is expected. Alternatively, interesting connections regarding attendance could be discovered in the making of this project.

h2. Implementation

The application is a mobile app, possibly WebGL app

It is developed in unity which provides more possibilities for export with modifications for given platforms

* Mobile app

Android support 5.0 and higher, use Unity UI tools to provide touch support.

Deliver an apk of the application to download outside of the Playstore or optionally, create a developer account (perhaps at a later date once the project is finished).

* WebGL app

Test options to embed a Unity Canvas app in a simple website. Possible issues with support in various browsers, but unlike the mobile app, the university could provide hosting.

Additionally, we have to figure out the Unity collaboration system and how to use GitLab with it.

h3. Prediction

* Naive Bayes classifier

* Neural network

Could classify days into high-medium-low attendance days, or more granularly (very high, very low...).

Risks:

We have very little experience with neural networks. Bayesian classification might provide good results with much less effort.

Further research on technologies:

* https://docs.microsoft.com/en-us/archive/msdn-magazine/2017/august/test-run-deep-neural-network-io-using-csharp

* https://nugetmusthaves.com/Tag/neural

* https://www.codeproject.com/Articles/5286497/Starting-with-Keras-NET-in-Csharp-Train-Your-First

A neural network library could be used with C#, test integration with Unity.