D2DraftNet — Embedding-Based Model for Dota 2 Draft Prediction

Overview

D2DraftNet is a neural network model that predicts the outcome of a Dota 2 match using only the hero draft. The core idea is to represent each hero pick not as a one-hot vector, but as a learned point in a low-dimensional vector space. This allows the model to capture complex, non-obvious relationships between heroes — such as synergies, counters, or general playstyle tendencies.

---

TODO

• Modify steam_api.py so that it creates a dataset that the embedding_model.py loader can use.

---

Draft Representation

Hero Embedding

Each hero is mapped to a learned vector in $R^d$, where (d) is a tunable hyperparameter (typically small, e.g., (d=3)).

Formally:

$$ h_i \mapsto \mathbf{v}_i \in \mathbb{R}^d $$

where:

• $h_i$ is the hero ID
• $\mathbf{v}_i$ is its embedding vector (learned during training)

---

Team Embedding

A team's draft (5 heroes) is represented by the average of its hero embeddings:

$$ v_{team} = \frac{1}{5} \sum_{i=1}^{5} v_i $$

This gives a single vector summarizing the composition and identity of the team in a continuous space.

---

Model Architecture

The model takes the two team vectors — Radiant and Dire — and concatenates them into a single vector:

$$ x = [v_{radiant}; v_{dire}] $$

This is passed through a simple fully-connected neural network:

Embedding Layer → Mean Pooling → Concatenation → Linear Layers → Sigmoid Output

• Linear layers perform affine transformations.
• Non-linearities (ReLU) allow for complex decision boundaries.
• The final output is a single probability: (P(\text{Radiant Win}))

---

Why Embeddings?

One-hot encoding treats every hero as equally distant from every other hero — which is unrealistic in Dota 2.

Embedding the draft into a learned vector space allows the model to:

• Group similar heroes together.
• Capture interactions and matchups.
• Compress drafts into dense, informative vectors.
• Learn synergy and counterplay patterns from data — not from hand-crafted features.

---

Training

Component	Method
Loss Function	Binary Cross Entropy (BCE)
Optimizer	Adam
Input	Radiant and Dire drafts
Output	Probability of Radiant Victory