This paper proposes a novel neural networks framework, Neural Attentive Recommendation Machine (NARM), for session-based recommendation. Session-based recommendation is used in e-commerce scenarios where user profiles are not available, and it generates recommendations based on short sessions. Previous methods only consider the user's sequential behavior in the current session, but not the user's main purpose. NARM addresses this by using a hybrid encoder with an attention mechanism to model the user's sequential behavior and capture the user's main purpose in the current session. These are combined into a unified session representation. The recommendation scores for each candidate item are computed using a bi-linear matching scheme based on this unified session representation. NARM is trained by jointly learning the item and session representations as well as their matchings. Extensive experiments on two benchmark datasets show that NARM outperforms state-of-the-art baselines on both datasets. Furthermore, NARM achieves significant improvements on long sessions, demonstrating its advantages in modeling the user's sequential behavior and main purpose simultaneously. The main contributions of this work are: (1) proposing a novel NARM model that considers both the user's sequential behavior and main purpose in the current session and computes recommendation scores using a bi-linear matching scheme; (2) applying an attention mechanism to extract the user's main purpose in the current session; and (3) conducting extensive experiments on two benchmark datasets, showing that NARM outperforms state-of-the-art baselines in terms of recall and MRR on both datasets. Moreover, NARM achieves better performance on long sessions, demonstrating its advantages in modeling the user's sequential behavior and main purpose simultaneously.This paper proposes a novel neural networks framework, Neural Attentive Recommendation Machine (NARM), for session-based recommendation. Session-based recommendation is used in e-commerce scenarios where user profiles are not available, and it generates recommendations based on short sessions. Previous methods only consider the user's sequential behavior in the current session, but not the user's main purpose. NARM addresses this by using a hybrid encoder with an attention mechanism to model the user's sequential behavior and capture the user's main purpose in the current session. These are combined into a unified session representation. The recommendation scores for each candidate item are computed using a bi-linear matching scheme based on this unified session representation. NARM is trained by jointly learning the item and session representations as well as their matchings. Extensive experiments on two benchmark datasets show that NARM outperforms state-of-the-art baselines on both datasets. Furthermore, NARM achieves significant improvements on long sessions, demonstrating its advantages in modeling the user's sequential behavior and main purpose simultaneously. The main contributions of this work are: (1) proposing a novel NARM model that considers both the user's sequential behavior and main purpose in the current session and computes recommendation scores using a bi-linear matching scheme; (2) applying an attention mechanism to extract the user's main purpose in the current session; and (3) conducting extensive experiments on two benchmark datasets, showing that NARM outperforms state-of-the-art baselines in terms of recall and MRR on both datasets. Moreover, NARM achieves better performance on long sessions, demonstrating its advantages in modeling the user's sequential behavior and main purpose simultaneously.