Source-linked AI summary
Neural Attentive Session-based Recommendation
Jing Li, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Jun Ma
TL;DR
Session-based recommendation must predict next clicks from short sessions without visible user profiles, while prior work emphasizes sequential behavior more than current purpose. NARM combines both through an attention-based hybrid encoder and bi-linear matching, and it outperforms state-of-the-art baselines on benchmark datasets. The authors also report stronger performance on long sessions, while noting limitations when sessions become too long and suggesting richer item attributes as future work.
Problem
Session-based recommendation predicts next items from current-session clicks without user profiles, but prior work does not emphasize the user’s main purpose.
Method
NARM uses an attention-based hybrid encoder to combine sequential behavior and session purpose, then scores candidate items with bi-linear matching.
Results
NARM outperforms state-of-the-art baselines on two benchmark datasets in recall and MRR, with better performance on long sessions.
Takeaways & Limitations
Modeling sequential behavior and main purpose together is associated with NARM’s stronger session-based recommendation performance.
Takeaways & Limitations
When sessions are too long, aimless clicks can prevent NARM’s local encoder from capturing the user’s main purpose; richer item attributes are proposed for future work.
Abstract
from arXiv · showhide
Given e-commerce scenarios that user profiles are invisible, session-based recommendation is proposed to generate recommendation results from short sessions. Previous work only considers the user's sequential behavior in the current session, whereas the user's main purpose in the current session is not emphasized. In this paper, we propose a novel neural networks framework, i.e., Neural Attentive Recommendation Machine (NARM), to tackle this problem. Specifically, we explore 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, which are combined as a unified session representation later. We then compute the recommendation scores for each candidate item with a bi-linear matching scheme based on this unified session representation. We train NARM by jointly learning the item and session representations as well as their matchings. We carried out extensive experiments on two benchmark datasets. Our experimental results show that NARM outperforms state-of-the-art baselines on both datasets. Furthermore, we also find that NARM achieves a significant improvement on long sessions, which demonstrates its advantages in modeling the user's sequential behavior and main purpose simultaneously.
1 INTRODUCTION
Session-based recommendation predicts the next item from current-session clicks when user profiles are unavailable, but existing approaches do not sufficiently capture the user’s main purpose. NARM combines sequential behavior and purpose through attention-based encoding and achieves stronger benchmark performance.
- Session-based recommendation predicts the next item from implicit feedback, such as clicks, within the current session.
- Existing sequential approaches can recommend items influenced by incidental clicks rather than the user’s main purpose.The motivating example contrasts accidental clicks on suit pants with repeated interest in shirts.
- NARM uses a hybrid encoder with attention to model sequential behavior and extract the user’s main purpose into a unified session representation.The attention mechanism assigns different importance to items in the session.
- NARM computes candidate-item recommendation scores through bi-linear matching and jointly learns item and session representations with their matchings.
- NARM outperforms state-of-the-art baselines in recall and MRR on two benchmark datasets and performs better on long sessions.
2 RELATED WORK
Session-based recommendation uses implicit sequential feedback, and related work spans traditional co-occurrence or Markov methods and neural approaches. NARM extends RNN-based session recommendation by combining sequential behavior with the user’s main purpose through attention.
- Session-based recommendation relies on implicit positive observations, such as clicks, rather than explicit preferences.
- Traditional methods: Traditional approaches include item-to-item similarity methods based on co-occurrence and Markov-chain recommenders that predict the next action.
- Deep learning based methods: Neural recommendation research includes collaborative filtering, auto-encoders, hierarchical encoder-decoder models, cross-domain representations, and RNNs for variable-length sequences.
- Deep learning based methods: NARM differs from existing session-based studies by combining sequential behavior and main purpose, applying attention to session-based recommendation.
3 METHOD
NARM uses a hybrid encoder with global sequence modeling and local attention to represent both a session’s sequential behavior and main purpose. It decodes this unified representation with bi-linear matching to rank candidate next items.
- 3 METHOD: Session-based recommendation predicts the next clicked item from a user’s current click sequence.The task uses implicit feedback from the current session rather than an enduring user profile.
- 3 METHOD: NARM converts the input click sequence into hidden representations and builds a current-session feature for producing item rankings.The encoder generates hidden states, while the session feature generator combines them with an attention signal before decoding.
- 3.3 Global Encoder in NARM: The global encoder uses a GRU to summarize the entire sequence and represent the user’s sequential behavior.Its final hidden state serves as the sequence representation, although a single vector can make precise current intention difficult to capture.
- 3.4 Local Encoder in NARM: The local encoder applies item-level attention so predictions can emphasize important inputs and capture the user’s main purpose.Attention weights model alignment between previous clicked items and the current hidden state, enabling adaptive focus during prediction.
- 3.5 NARM Model: NARM concatenates global and local representations into a unified session feature that models sequential behavior and session purpose together.The global encoder represents the entire sequence, while the local encoder computes attention using previous hidden states.
- 3.5 NARM Model: The bi-linear decoder computes similarity scores between the session representation and candidate items, reducing parameters from |N| ∗|H| to |D| ∗|H|.The scores enter a softmax layer to obtain next-item probabilities, and experiments report improved NARM performance with this decoder.
4 EXPERIMENTAL SETUP
NARM is evaluated on two transaction datasets against traditional and RNN-based baselines using top-20 recommendation metrics. Experiments compare decoders and show advantages for the bi-linear decoder and NARM, especially on DIGINETICA.
- Datasets: Experiments use YOOCHOOSE and DIGINETICA transaction datasets after filtering short sessions and infrequent items.YOOCHOOSE contains 7,981,580 sessions and 37,483 items; DIGINETICA contains 204,771 sessions and 43,097 items.
- Preprocessing: Training examples are generated from session prefixes, with each next click serving as the prediction label.For an input session, prefixes from the first item through the penultimate item are paired with their subsequent clicks.
- Methods and metrics: NARM is compared with five traditional methods and two RNN-based models using Recall@20 and MRR@20.Recall@20 measures whether the desired item appears among the top 20, while MRR@20 accounts for its rank.
- Decoder comparison: Bi-linear decoding improves Recall@20 by approximately 0.65%, 0.24%, and 4.74% across the three evaluated datasets.MRR@20 is slightly worse on both YOOCHOOSE splits but remains clearly better than fully connected decoding on DIGINETICA.
- Decoder selection: The experiments adopt bi-linear decoding because Recall@20 is considered more important than MRR@20 when recommending 20 items at once.The actual item selected by a user is expected to be among the recommended list of 20 items.
- Baseline comparison: NARM outperforms all baselines in Recall@20 across three datasets and most baselines in MRR@20.On DIGINETICA, NARM improves over Improved GRU-Rec by about 7.98% in Recall@20 and 9.70% in MRR@20.
5 ANALYSIS
The analysis evaluates NARM’s feature combinations, session-length behavior, and attention weights. Results show that combining sequential behavior with user purpose is beneficial, while performance gains decline for excessively long sessions.
- 5.1 Influence of Using Different Features: NARM_global and NARM_local perform similarly and worse than NARM_hybrid, indicating that either feature alone is insufficient.NARM_hybrid outperforms both single-feature variants on Recall@20 and MRR@20 across datasets and hidden-state dimensions.
- 5.1 Influence of Using Different Features: 3.52% and 5.09% are NARM_hybrid’s relative Recall@20 improvements over NARM_global and NARM_local on DIGINETICA with hidden dimension 50.
- 5.2 Influence of Different Session Lengths: NARM is designed for sessions where click behavior revolves around a main purpose, but that purpose is difficult to capture after only a few clicks.
- 5.2 Influence of Different Session Lengths: When sessions become too long, NARM’s performance improvements decline because aimless clicks can prevent the local encoder from capturing the user’s purpose.
- 5.3 Visualize the Attention Weights: Attention weights indicate that important items are often near the session’s end, but can also occur at the beginning or middle.The analysis reports that important items are generally continuous, suggesting localized user intentions.
- 5.3 Visualize the Attention Weights: The attention mechanism can emphasize important items regardless of their position, supporting NARM’s modeling of localized intentions.
6 CONCLUSION & FUTURE WORK
The paper concludes that NARM addresses session-based recommendation by combining sequential behavior and session purpose through attention. It reports stronger performance than state-of-the-art methods and identifies additional item attributes and neighboring sessions as future directions.
- 6 CONCLUSION & FUTURE WORK: NARM uses an encoder-decoder architecture with attention to capture both sequential behavior and the user’s main purpose in a session.
- 6 CONCLUSION & FUTURE WORK: Experiments on two benchmark datasets show that NARM outperforms state-of-the-art methods across different evaluation metrics.
- 6 CONCLUSION & FUTURE WORK: User-click analysis finds that intentions are localized in most sessions, supporting the model’s attention-based design.
- 6 CONCLUSION & FUTURE WORK: Future work may add item attributes such as prices and categories, exploit nearest-neighbor sessions, and examine attribute importance with attention.