TransFM: Integrating Translation-based Recommendation and Factorization Machines for Sequential Recommendation

The paper "TransFM" (RecSys 18) by Rajiv Pasricha and Julian McAuley extends the translation‑based recommendation model (TransRec) by integrating factorization machines (FM) to better capture high‑order feature interactions in sequential recommendation tasks.

Traditional recommender models such as matrix factorization ignore temporal dynamics, while sequence‑aware models like TransRec embed items in a translation space and model user behavior as translation vectors. FM, on the other hand, models pairwise feature interactions via inner products and can be applied to any prediction task.

TransFM replaces FM’s inner‑product with squared Euclidean distance and represents each feature vector as the sum of an embedding vector and a translation vector. This allows the model to capture transitive relationships among features and improves generalization.

The model’s objective is optimized using Sequential Bayesian Personalized Ranking (S‑BPR) with L2 regularization, and training is performed in TensorFlow using mini‑batch gradient descent with the Adam optimizer.

Experiments on Amazon and Google Local datasets show that TransFM consistently outperforms baselines such as PRME, HRM, and pure FM in terms of AUC, while maintaining linear computational complexity O(nk) with respect to feature dimension n and embedding size k.

Key take‑aways: (1) TransFM effectively merges the strengths of TransRec and FM for sparse, large‑scale data; (2) using squared Euclidean distance instead of inner product enhances model expressiveness; (3) the framework easily incorporates additional contextual features (time, geography, demographics) without altering the model architecture.