AFFM: Auto feature engineering in field-aware factorization machines for predictive analytics

Abstract

User identification and prediction is one typical problem with the cross-device connection. User identification is useful for the recommendation engine, online advertising, and user experiences. Extreme sparse and large-scale data make user identification a challenging problem. To achieve better performance and accuracy for identification a better model with short turnaround time, and able to handle extremely sparse and large-scale data is the key. In this paper, we proposed a novel efficient machine learning approach to deal with such problem. We have adapted Field-aware Factorization Machine's approach using auto feature engineering techniques. Our model has the capacity to handle multiple features within the same field. The model provides an efficient way to handle the fields in the matrix. It counts the unique fields in the matrix and divides both the matrix with that value, which provide an efficient and scalable technique in term of time complexity. The accuracy of the model is 0.864845, when tested with Drawbridge datasets released in the context of the ICDM 2015 Cross-Device Connections Challenge.