Structured Deep Factorization Machine: towards General-purpose Architectures

In spite of their great success, traditional factorization algorithms typically do not support features (e.g., Matrix Factorization), or their complexity scales quadratically with the number of features (e.g, Factorization Machine). On the other hand, neural methods allow large feature sets, but are often designed for a specific application. We propose novel deep factorization methods that allow efficient and flexible feature representation. For example, we enable describing items with natural language with complexity linear to the vocabulary size—this enables prediction for unseen items and avoids the cold start problem. We show that our architecture can generalize some previously published single-purpose neural architectures. Our experiments suggest improved training times and accuracy compared to shallow methods.