MOBA games are currently one the most popular online game genres. In their basic gameplay, two teams of multiple players compete against each other to destroy the enemy's base, controlling a powerful unit known as "hero". Each hero has different abilities, roles and strengths. Thus, choosing a good combination of heroes is fundamental for the success in the game. In this paper we propose a recommendation system for selecting heroes in a MOBA game. We develop a mechanism based on association rules that suggests the more suitable heroes for composing a team, using data collected from a large number of DOTA 2 matches. For evaluating the efficacy of the line-up, we trained a neural network capable of predicting the winner team with a 88.63% accuracy. The results of the recommendation system were very satisfactory with up to 74.9% success rate.