andreiapostoae
commented
6 years ago Do not assume an accuracy until you implemented and tested your algorithm. I strongly believe that feature engineering is the key step you have to worry about and proved to have the most impact on accuracy. I wouldn't expect gradient boosting to achieve much higher accuracy than logistic regression as long as you use the same features as me. I think where you should look (and I'm thinking about this too at the moment) is how you use the pick order and how you model a team. Currently, the algorithm has no understanding that a team should only contain 5 heroes on each side and that Xth feature for radiant corresponds to Xth feature from dire. You also have to take into account that the current implementation uses both binary and continuous features, so you might want to take care of them separately. I feel like neural networks, properly modelled (convolutional + LSTM is what I would try), would lead to better results, but that is only a hunch.
dota2heqiuzhi
Hi, me and two of my friends are going to have this final project for our course in the university about dota-2 predictor using gradient boosting, which is another ML algorithm that we believed can be tested with this predictor and maybe have more than 60% accuracy. If our proposal will be accepted by our lecturers and teaching assistants, it will take us 1-2 months to finish the project.
Is there any suggestion or any tips that we can applied during later on? Many thanks