Just a few days after officially starting our classes for the 1st term in the master program, we received an email about a gaming data hackathon being organized by SocialPoint – a big name in mobile gaming. The company is spread across 7 floors of a 10-storey building in the heart of Barcelona with amazing views of the city. Having just finished the brush-ups, we were not quite aware of the kind of data challenges that lay ahead in such competitions but it was surely going to be a learning experience so we went ahead and registered a team.
We arrived at the venue on the morning of 15 Oct to learn about the challenge at hand. We were provided with gameplay data from the company’s most successful game Monster Legends. For our readers with some time at their disposal (not quite the case with most of us doing the master), you should try this game but I shall warn you it can be addictive! The basic aim of the game is to collect the strongest monsters and build up your team of monsters to win the maximum battles against other players. And this brings me to the problem we had to solve: Given the list of monsters currently owned by a player, predict the monster they don’t already have and are most likely to choose next.
We were given:
- A training dataset of 70,000 players and all the monsters they owned.
- A test dataset of 30,000 players with one random monster removed from their list of monsters – This removed item was to be predicted
- Basic information about each of the 100,000 players (country, sex, platform, days_played, payer_dummy etc)
- Features of each monster available in the game
- A baseline Item Based Collaborative Filtering (IBCF) model in R which gave an accuracy of 15.5%
There were 2 tracks on which each team would be judged:
- Objective evaluation of the precision of the predictions which was accessible through a live leaderboard – it evaluated the submissions on 20% of the solutions to avoid overfitting.
- Subjective evaluation of a Business Insights presentation which was supposed to be delivered the next morning with suggestions to SocialPoint about how to possibly increase revenue from the game, given our analysis on the provided data.
There were cash prizes for both tracks – 1st and 2nd prize of €750 and €500 for the accuracy track and a unique prize of €500 for the business insights track.
We would be presenting our results to – Horacio Martos (CEO, SocialPoint), Sharon Biggar (Head of Analytics, SocialPoint), Manuel Bruscas (Director of Analytics, eDreams) and our program co-director Christian Fons-Rosen.
With all this information at our disposal – the countdown timer for 24 hours started and we put on our thinking caps to brainstorm. We began with visualizing as much data as possible to get a general sense of the data provided. Apparently our brains were waiting for some calorie intake as only after the lunch (which was a fancy courtesy of the organizers), we came up with a basic idea of modeling the game as chronological event and predicting the monster for a player based on the player’s current level.
2 of us set on to implement this approach while the other 2 continued tweaking the baseline model in hopes of achieving a better precision. However, we failed to cross the 15% mark after several submissions coupled with hours of stackoverflowing and reading about various techniques.
Much to our disappointment, our chronological approach also didn’t give very optimistic results and DeepYellow (our team) was still lingering around 17% by the time we left the venue on Saturday evening. Some of us stayed up all night trying to implement different ideas – clustering players on features other than level and other techniques like user-based filtering. I believe the competitiveness of the event with regular submissions by all teams even at 3AM and the idea of trying new models kept us going.
After all our efforts, we were still at 19% at the end of 24 hours and clearly not the winning the team – the winners managed to achieve 37% accuracy. Given our unsuccessful attempts at analyzing the data – we were not really motivated about the presentation and we gave a very basic one which was not really very insightful. However, we learned a lot from the presentations of the other teams.
At the end of the event, we were given a tour of the offices and we got back to our lecture slides albeit not with cash prizes but with ideas on practically facing a data problem and hopefully some acquired knowledge. Hackathons are always a fun way of squeezing out the creative side in you in a limited amount of time coupled with the bonus of meeting many like-minded new people.
Our team consisted of Akhil Lohia, Nandan Rao, Robert Lange and Roman Kopso (an exchange student at UPC who was included in our team to assure all teams had 4 members).
About the author (Akhil Lohia):
I graduated with a bachelor’s degree in Economics from the Indian Institute of Technology Kanpur in June 2015. I have been involved in economics research projects in the capacity of a research assistant. My deep fascination towards technology and the ever increasing involvement of data to augment and provide more useful services to users in their context led my decision in choosing the Data Science program at BGSE in a smart city like Barcelona.