Context:
I was looking at using an LSTM model to forecast the amount of gold gained for each of 10 heroes in a game of Dota 2, a MOBA game, as a base model in some type of model architecture. The game is a 5v5 with a main feature of the game to amass gold. The challenge is how to include categorical feature variables for the hero and a feature variable for the team the hero is a part of into the model.
1. Target Variable: Gold
Here is an example of the data without conforming to the typical 3D forms of TF/Keras and PyTorch:
match_id team hero_id gold_t
7517376613 radiant 26 [0, 170, 260, 350, 440, 827, 1051, 1141, 1278,...
7517376613 radiant 84 [0, 247, 580, 826, 1235, 1481, 1876, 2279, 265...
7517376613 dire 25 [0, 357, 715, 1130, 1578, 1968, 2261, 2861, 33...
2. Categorical Variable: hero_id & team
I have thousands/hundreds-of-thousands of matches of time-series data for each hero in the game I can train on. Is there a way to incorporate the categorical variable of the hero chosen to develop a model that forecasts specifically for that hero? This seems to be answered by a source I posted at the bottom.
Is there a way to incorporate interactions between the time series as features to predict for each? For instance, if I was to train a model on a time-series forecast for hero_id=26, can the progression of hero_id=25 time-series be treated as a feature while incorporating whether they were on the same team? I understand multi-variate time-series, the question is whether the time-series could be uniquely identified as belonging to hero_id 25 and capture its unique interaction with hero_id=26. I am familiar with using time-lags/multi-variate data as features, but not in combination with a categorical variable.
If no options to include all of these is available, that's fine. There is a route to simplification:
Simplify this to train a different model for each hero_id (lots of processing), using each game as a separate time-series (is that even possible?). After training, each model would need to be called separately per the hero_id.
3. What I've Read Elsewhere (only one thing seemed relevant):
I have referenced and played with two items:
In summary, it suggests initializing RNN states with a learned representation of the categorical variables. Could this be accomplished with multiple categories?
Even if that is possible, I did not know how to incorporate using separate matches as training data. In most examples I've seen, the time-series is one long instance and treats each time-step as a sample. In my situation the time-series would not last more than 45 steps, but I have access to 100,000+ time-series' of 45 steps. Is there a meaningful way these 100,000+ matches can be used as training data or will the model work perfectly fine on a single time-series?
