Personalizing Session-based Recommendations with Hierarchical Recurrent Neural Networks

• Sessions group user interactions within a time frame and typically serve a goal such as finding a restaurant or listening to music.Source: §1 · p. 1
• RNNs recently improved session-based recommendation over item-based methods by 15% to 30% in ranking metrics.Source: §1 · p. 1
• When user identifiers are available, behavior from past sessions may provide valuable information for recommendations in a subsequent session.Source: §1 · p. 1
• Concatenating past and current sessions is a simple personalization strategy, but it does not yield the best results experimentally.Source: §1 · p. 1
• The proposed Hierarchical RNN transfers information across sessions while supporting both identifiable-user and no-identifier recommendation settings.Source: §1 · p. 1
• On two industry datasets, Hierarchical RNNs outperform plain session-based RNNs and item-based collaborative filtering by a healthy margin.Source: §1 · p. 1

Session-based recommendations

• Classical collaborative filtering can fail when session-based settings provide no user profile, motivating item-to-item recommendation from co-clicked session items.Source: §2 · p. 1
• Item-to-item methods precompute item similarities from session data and use frequently co-clicked items to generate recommendations.Source: §2 · p. 1

Recurrent Neural Models

• RNNs are widely used for sequential data, with LSTMs adding gates that regulate input and hidden-state resets.Source: §2 · p. 2
• Earlier session RNNs used clicked item IDs from the current session to recommend the next item, whereas this work models behavior across sessions.Source: §2 · p. 2

3.1 Session-based Recurrent Neural Network

• The baseline session RNN uses one GRU to model within-session interactions and output a score for every catalog item as the next-item likelihood.Source: §3.1 · p. 2
• The session RNN processes current item IDs sequentially, initializes its hidden state with the null vector, and generates session-level representations.Source: §3.1 · p. 2
• TOP1 loss is used because it outperformed the other tested ranking losses; it regularizes the relative rank of the relevant item using sampled irrelevant items.Source: §3.1 · p. 2

3.2 Hierarchical RNN

• HRNN adds a user-level GRU that tracks cross-session activity and user-interest evolution, alongside user-parallel mini-batches for efficient training.Source: §3.2 · p. 2, §3.2.1 · p. 2

3.2.1 Architecture

• At session end, the user representation is updated from the session-level state; at the next session, it initializes the session GRU and may also enter each input.Source: §3.2.1 · pp. 2–3
• The user-level GRU learns session evolution rather than simply copying the previous session state, producing a user profile that personalizes recommendations.Source: §3.2.1 · p. 3
• HRNN Init uses the user representation only to initialize the next session, while HRNN All additionally propagates it at every next-session step.Source: §3.2.1 · p. 3

3.2.2 Learning

• User-parallel batches group and time-order sessions by user, enabling training across users with different session counts and lengths.Source: §3.2.2 · pp. 3–4

Experimental setup

• The evaluation uses sequential next-item prediction, bootstrapping HRNN and RNN Concat models with preceding user sessions in original order.Metrics assess the rank of the selected next item after each event.Source: §4.3 · p. 5

Results on XING

• XING exhibits strong activity repetitiveness, making personalized recommendation difficult and making personalized popularity highly competitive.The session-based RNN is significantly worse than PPOP across network capacities.Source: §4.3 · p. 5
• On XING, HRNNs significantly outperform RNN Concat in Recall and Precision and improve large-network MRR, while exceeding PPOP by approximately 11% in Recall and Precision with comparable MRR.HRNN gains over RNN Concat reach +3%/+1% in Recall for small/large networks and +5.4% MRR with HRNN All in large networks.Source: §4.3 · p. 5
• On XING, HRNN All and HRNN Init have no significant differences in Recall or Precision, while HRNN All has a small approximately 2% MRR advantage.The comparison motivates further analysis on VIDEO.Source: §4.3 · p. 5

Results on VIDEO

• On VIDEO, session-based RNN outperforms both Item-KNN and PPOP, while HRNN Init significantly outperforms all baselines and RNN Concat.In the reported large-network setting, HRNN Init is up to 6.5% better in Recall and 2.3% in MRR than RNN Concat.Source: §4.3 · pp. 5–6

Analysis on user history length

• Longer user histories improve Recall for all VIDEO methods, but improve MRR only for RNN Concat and HRNN Init; HRNN Init’s advantage over session-based RNN grows with history length.The Recall@5/MRR@5 gain grows from 5%/12% for short histories to 7%/19% for long histories.Source: §4.3.1 · p. 6

Analysis within sessions

• Within VIDEO sessions, HRNN Init’s early-session advantage reaches 10% in Recall and 25% in MRR, while later session dynamics reduce Recall gains but not MRR gains.Personalization remains associated with higher ranking quality throughout the session.Source: §4.3.2 · p. 7
• On the small VIDEO dataset, HRNN Init outperforms session-based RNN by approximately 28% in Recall@5 and 41% in MRR@5.RNN Concat performs similarly to the session-based RNN in this evaluation.Source: §4.3.3 · p. 8

• Hierarchical RNNs transfer long-term user-interest knowledge to session-level representations for personalized recommendations to returning users.The model adds a GRU level that models user activity across sessions and interest evolution over time.Source: §5 · p. 8
• HRNNs significantly outperform state-of-the-art session-based RNNs and basic personalization strategies on two real-world datasets.Source: §5 · p. 8
• HRNN Init, which initializes the session-level representation with the evolving user representation, gives the best results.Source: §5 · p. 8
• The proposed approach sets new state-of-the-art performance for personalized session-based recommendation.Source: §5 · p. 8
• Future work will examine attention models, item and user features, and personalized session-based models in domains such as music, e-commerce, and online advertising.Source: §5 · p. 8