Newwit AI model inputs consists of three parts.
Player embedding This is a mathematical matrix describing a player’s features, including age, occupation, education, number of predictions, overall accuracy, topic accuracy etc.
Question embedding This is a mathematical matrix describing a question’s features, including title, associated topics, associated subtopics, number of answers etc.
Interaction embedding This is a mathematical matrix describing how the player interacted with the question, including player’s prediction, player’s comments, number of upvotes received from the comment etc.
Model target With the above 3 input embeddings, the model will compute a prediction power score per user using a deep neural network. During training, the AI model will push the prediction power to 1 if the player’s prediction was correct and -1 if the player’s prediction was incorrect. When the model converges, it will know how much to trust a player when it comes to a particular question. The level to trust is the measure of prediction power.
Note that prediction power scores of 1 and -1 are equally important to the model. Player with -1 prediction power score contributes to the model by saying that most likely his/her prediction is incorrect. Prediction power score of 0 will mean the player’s prediction performance is not consistent and is considered noise statistically.
The prediction power of each player in each question is different.
Final Newwit Prediction
Finally, the prediction power scores are summed to compute the final prediction.