Source-linked AI summary
Heterogeneous Global Graph Neural Networks for Personalized Session-based Recommendation
Yitong Pang, Lingfei Wu, Qi Shen, Yiming Zhang, Zhihua Wei, Fangli Xu, Ethan Chang, Bo Long, Jian Pei
TL;DR
Session-based recommenders often overlook a user’s historical sessions, while personalized approaches may omit transition patterns from other users’ histories. HG-GNN uses a heterogeneous global graph and graph-augmented preference encoders to combine these sources. The paper reports that it outperforms state-of-the-art session-based models in experiments.
Problem
Session-based methods often overlook users’ historical sessions, and personalized methods may ignore item-transition patterns from other users’ histories.
Source: §1 · p. 2Method
HG-GNN uses a heterogeneous global graph of session transitions, user-item interactions, and global co-occurrences, with graph-augmented encoders to model preference.
Source: §1 · p. 2Results
Takeaways & Limitations
The authors conclude that HG-GNN is effective for personalized session-based recommendation.
Source: §6 · p. 8
Abstract
from arXiv · showhide
Predicting the next interaction of a short-term interaction session is a challenging task in session-based recommendation. Almost all existing works rely on item transition patterns, and neglect the impact of user historical sessions while modeling user preference, which often leads to non-personalized recommendation. Additionally, existing personalized session-based recommenders capture user preference only based on the sessions of the current user, but ignore the useful item-transition patterns from other user's historical sessions. To address these issues, we propose a novel Heterogeneous Global Graph Neural Networks (HG-GNN) to exploit the item transitions over all sessions in a subtle manner for better inferring user preference from the current and historical sessions. To effectively exploit the item transitions over all sessions from users, we propose a novel heterogeneous global graph that contains item transitions of sessions, user-item interactions and global co-occurrence items. Moreover, to capture user preference from sessions comprehensively, we propose to learn two levels of user representations from the global graph via two graph augmented preference encoders. Specifically, we design a novel heterogeneous graph neural network (HGNN) on the heterogeneous global graph to learn the long-term user preference and item representations with rich semantics. Based on the HGNN, we propose the Current Preference Encoder and the Historical Preference Encoder to capture the different levels of user preference from the current and historical sessions, respectively. To achieve personalized recommendation, we integrate the representations of the user current preference and historical interests to generate the final user preference representation. Extensive experimental results on three real-world datasets show that our model outperforms other state-of-the-art methods.
1 INTRODUCTION
Session-based recommenders focus on the active session, but often omit the current user’s history and cross-user item-transition patterns. HG-GNN combines information from sessions, user-item interactions, and global item co-occurrences to model preference.
1 INTRODUCTION
- Many session-based methods model sequential item transitions in the active session, while traditional recommenders learn preferences from long-term interactions and may not suit session-focused applications.Session-based recommendation instead generates recommendations mainly from interactions in the active session.Source: §1 · p. 1
- Methods that treat users as anonymous can overlook their historical sessions, so users with similar current sessions may receive the same candidate items despite different histories.The introduction illustrates this with two users whose similar item sequences lead to the same recommendations.Source: §1 · p. 2
- Existing personalized session recommenders use the current user’s sessions but ignore item-transition patterns from other users’ historical sessions.The paper suggests these other-session transitions might help model user preference.Source: §1 · p. 2
- HG-GNN combines session transitions, user-item interactions, and globally co-occurring items in a heterogeneous graph, then uses graph-augmented preference encoders for personalized session recommendation.The paper reports that experiments on three datasets show superiority over state-of-the-art models.Source: §1 · p. 2
2 RELATED WORKS
Prior work on session-based recommendation spans co-occurrence and Markov methods, deep sequence models, and graph-based methods. Personalized approaches use a user’s own history, while the paper identifies cross-user historical item transitions as an additional source of potentially useful patterns.
Traditional Methods
- Markov-chain methods model next actions from the last action, but usually capture only first-order transitions and cannot represent more complex sequential patterns.Earlier session-based systems also used co-occurrence information following collaborative filtering.Source: §2 · p. 2
Deep Learning-based Methods
- Deep learning methods for session-based recommendation include RNN-, CNN-, and attention-based approaches that learn session representations or sequential patterns.Examples include GRU4Rec, NARM, and Caser.Source: §2 · p. 2
GNN-based Methods
- GNN-based methods learn item representations from session or global graphs, but the cited approaches mainly focus on current-session sequences and may produce non-personalized recommendations.Examples include SR-GNN, GC-SAN, and GCE-GNN.Source: §2 · p. 2
Personalized Session-based Recommendation
- Personalized session methods use the current user’s historical sessions, while item-transition patterns from other users’ histories remain an additional potentially useful source.The text describes H-RNN and A-PGNN as methods centered on the current user’s history and notes that other users’ transitions might help model preference.Source: §2 · p. 3
3 PRELIMINARY
The paper frames personalized session-based recommendation as predicting a user’s next item from the current session and that user’s preceding sessions. Its formulation assigns a prediction score to each candidate item.
Problem Formulation
- Each user’s interaction records comprise historical session sequences ordered chronologically.The formulation denotes the universal sets of users and items as U and V, respectively.Source: §3 · p. 3
- The task is to predict the item a user is most likely to click next, given that user’s historical sessions and current session.The next item is indexed as the item following the current session sequence.Source: §3 · p. 3
- The model generates a prediction score for each candidate item in the item set.The formulation denotes the score for item v_i as ŷ_i.Source: §3 · p. 3
4.1 Overview
HG-GNN organizes historical sessions and global information in a heterogeneous graph, then derives session representations by combining graph-learned user preference with current-session items.
4.1 Overview
- The heterogeneous global graph connects users and items using historical user-item interactions, item co-occurrences, and pairwise transitions within session sequences.The graph is intended to organize historical sessions and extract global information.Source: §4.1 · p. 3
- A heterogeneous graph neural network learns long-term user preference and item representations, while a Personalized Session Encoder combines general user preference with current-session items.The combination produces a personalized session representation.Source: §4.1 · p. 3
- The resulting personalized session representation is intended to capture user preference comprehensively.The overview describes it as integrating general preference with items from the current session.Source: §4.1 · p. 3
4.2 Heterogeneous Global Graph Construction
HG-GNN builds a directed heterogeneous global graph over users and items, combining session transitions, historical user-item interactions, and global item co-occurrences. Its item edges capture both adjacent transitions and frequent co-occurrences, while user-item edges encode interaction history.
4.2.1 Item-to-Item.
- Unlike methods using only adjacent session transitions [29] [30], the graph also captures frequent within-session co-occurrences, which complement adjacent interactions.Transitions are represented by paired directed incoming and outgoing edges, weighted by their frequency across historical sessions and restricted to each item's top-S edges.Source: §4.2.1 · pp. 3–4
- The graph adds directed similar-item edges for frequently co-occurring items, limiting their number per item to avoid introducing too much noise.When an incoming transition edge and a similar-item edge connect the same items, the transition edge is retained; the similar-edge limit depends on the item's adjacent interaction-item count.Source: §4.2.1 · p. 4
4.2.2 Item-to-User.
- User-item historical interactions become directed edges from items to users and from users to items, representing the interaction behavior.The item-to-user meta-path is described as revealing users' implicit long-term preferences.Source: §4.2.2 · p. 4
- The global graph has user and item nodes and combines session item transitions, historical user-item interactions, and global co-occurrence information.It supports unified learning of user and item representations, item correlations, and long-term user preference.Source: §4.2 · p. 3, §4.2.2 · p. 4
4.3 Heterogeneous Global Graph Neural Network
HGNN propagates information across the directed heterogeneous global graph to encode users and items. It aggregates relation-specific neighbor messages, updates node representations, and combines outputs across layers into global-level embeddings.
4.3 Heterogeneous Global Graph Neural Network
- HGNN initializes user and item nodes with d-dimensional embeddings and encodes both node types on the directed heterogeneous global graph.The resulting representations are intended to capture long-term user preference and refined item semantics.Source: §4.3 · p. 4
- For each edge type, HGNN aggregates messages from the corresponding neighbors; item nodes receive messages through incoming, outgoing, similar-item, and user-interaction relations.The model uses mean aggregation in practice to combine messages when updating item representations.Source: §4.3 · p. 4
- HGNN updates user representations using an aggregation operation analogous to the item update.The passage describes user-node aggregation as similar to the item operation.Source: §4.3 · p. 4
- After K layers, the model combines the embeddings from each layer to form final global-level user and item representations.Each layer has a nonnegative importance weight; the experiment sets these weights uniformly to 1/(K + 1).Source: §4.3 · p. 4
4.4 Personalized Session Encoder
The personalized session encoder combines a current-session preference with a general preference informed by the user's long-term interest. A gating mechanism integrates these representations into the final session preference.
4.4 Personalized Session Encoder
- The encoder models dynamic preference from the ongoing session alongside general preference informed by the user's long-term stable interest.The current preference module models session item transitions, while the general preference module uses attention.Source: §4.4 · p. 4
4.4.1 Current Preference Learning.
- Current Preference Learning uses item-level attention to combine session-item information, with item position informing each item's contribution to the session representation.It concatenates HGNN-learned item representations with reversed position embeddings and applies soft attention to obtain current preference C_u.Source: §4.4.1 · pp. 4–5
4.4.2 General Preference Learning.
- General Preference Learning uses attention to relate items in the current session to the HGNN user embedding, which contains long-term stable preference.The module linearly combines item representations to obtain the session's general preference.Source: §4.4 · p. 4, §4.4.2 · p. 5
- A gating mechanism integrates general preference O_u and current preference C_u into the final session preference S_u.The general preference reflects long-term user interest, while the current preference uses basic item information from the current session.Source: §4.4.2 · p. 5
4.5 Prediction and Training
The model derives recommendation probabilities for candidate items from the session preference and their initial embeddings, then trains with cross-entropy loss.
4.5 Prediction and Training
- The model computes candidate-item recommendation probabilities from the session preference representation and initial candidate-item embeddings.Each probability denotes the likelihood that the user will click that item in the current session.Source: §4.5 · p. 5
- Training uses a cross-entropy loss objective.Source: §4.5 · p. 5
5 EXPERIMENTS
Experiments compare HG-GNN with session-based and personalized recommenders across three datasets, then assess its components, graph design, hyperparameters, runtime, and personalization. HG-GNN leads state-of-the-art methods on most metrics across datasets, while ablations and a case study examine its design and recommendations.
5.1 Experimental Setup
- Experiments use Last.fm, Xing, and Reddit, comparing GNN-based and personalized baselines with HR@{5, 10} and MRR@{5, 10}.Sessions with fewer than three interactions and users with fewer than five sessions are filtered out; the last 20% of sessions form the test set.Source: §5.1.1 · pp. 5–6, §5.1.2 · p. 6, §5.1.3 · p. 6
5.2 Model Comparison (RQ1)
- HG-GNN outperforms state-of-the-art models on most metrics across all datasets, including A-PGNN by 4.03% in HR@5 and 5.72% in MRR@5 on Reddit.Table 2 reports HR@{5, 10} and MRR@{5, 10}; its dagger marks significant gains over the best baseline using a paired t-test (p<0.01).Source: §5.2 · p. 6, Table 2 · p. 7
- HG-GNN is more efficient than A-PGNN and GCE-GNN in measured network runtime per training epoch.The timing excludes data loading, testing, and graph creation.Source: §5.2 · pp. 6–7
5.3 Ablation and Effectiveness Analyses (RQ2)
- Removing HGNN, current preference learning, or general preference learning worsens performance, supporting the effectiveness of these components.Table 4 also compares HG-GNN layer configurations.Source: §5.3.1 · p. 7, Table 4 · p. 8
5.4.1 Effect of the sampling size.
- Sampling size S=8 gives the highest reported performance on Last.fm and Xing, while sampling that is too large or too small reduces performance.The authors relate the optimum to balancing useful information against irrelevant information.Source: Figure 3 · p. 7, §5.4.1 · p. 8
5.3.2 Impact of different Global Graph Design.
- Removing user nodes causes the largest performance drop; adjacent-interaction edges matter more than similar-item edges, though co-occurrence edges also help.The study reports that adjacent-interaction and global co-occurrence relationships complement each other.Source: §5.3.2 · p. 7, Table 5 · p. 8
5.4 Hyper-parameters Study (RQ3)
- The optimal number of similar-item edges varies across datasets, making this hyperparameter sensitive to dataset characteristics.Similar items are selected using global co-occurrence information.Source: Figure 4 · p. 8, §5.4.1 · p. 8
5.5 Case Study (RQ4)
- For two users with similar current sessions but different listening histories, HG-GNN produces different recommendations from the non-personalized SR-GNN baseline.The case study illustrates recommendations reflecting both current-session relevance and other user preferences.Source: Figure 5 · p. 8, §5.5 · p. 8
6 CONCLUSION
HG-GNN combines historical user-item interactions, item transitions, and global co-occurrence information in a heterogeneous graph, then uses a graph-augmented hybrid encoder to represent user preference. The paper reports that it outperforms other state-of-the-art session-based models.
6 CONCLUSION
- HG-GNN outperforms other state-of-the-art session-based models in the reported experiments.The conclusion describes this as evidence of the model's effectiveness.Source: §6 · p. 8
- Its heterogeneous global graph combines historical user-item interactions, item transitions, and global co-occurrence information.Source: §6 · p. 8
- A graph-augmented hybrid encoder combines a heterogeneous graph neural network with a personalized session encoder to capture user preference comprehensively.Source: §6 · p. 8