Session-based Recommendation with Graph Neural Networks

Shu WuYuyuan TangYanqiao ZhuLiang WangXing XieTieniu Tan

article2018AAAI1,955 citations

Proposes a session-based recommendation framework that models user click sequences as directed graphs and applies graph neural networks with an attention mechanism to capture complex item transitions and predict next-item interactions without requiring user profiles.

Listen

Modern online platforms frequently encounter anonymous users whose historical profiles and long-term preferences are entirely unavailable. In these scenarios, recommendation systems must rely solely on the limited sequence of clicks generated within an active browsing session. Conventional sequential models struggle in this context because they attempt to estimate explicit user profiles from minimal data and only evaluate simple transitions between consecutive items, neglecting complex, multi-step relationships across the entire session.

The article evaluates whether transforming individual browsing sessions into graph structures and applying graph neural networks can improve next-click prediction accuracy. The authors set out to demonstrate that this graph-based approach can better capture intricate item transitions while generating robust session representations without needing explicit user profiles.

To test this concept, the authors developed a model that converts click sequences into directed subgraphs, uses gated graph neural networks to learn item relationships, and applies an attention mechanism to combine immediate user interest with broader session intent. The framework was evaluated against leading conventional, sequential, and recurrent neural network baselines using two benchmark e-commerce transaction datasets, Yoochoose and Diginetica, encompassing millions of interaction clicks.

The analysis revealed three primary findings. First, the proposed graph neural network method consistently outperformed all baseline algorithms across both datasets, achieving top-20 precision of 70.57% to 71.36% on Yoochoose and 50.73% on Diginetica, while also improving recommendation ranking quality. Second, while prior state-of-the-art recurrent models suffered notable performance drops on longer sessions—falling by roughly 6 to 11 percentage points in precision—the proposed model maintained stable accuracy across both short and long browsing sequences. Third, ablation tests demonstrated that combining immediate last-click interest with an attention-weighted global session preference yielded superior results compared to using either factor in isolation or relying on simple averaging.

These findings indicate that shifting from linear sequence modeling to graph-based structures significantly enhances recommendation accuracy in anonymous environments. By capturing complex item transitions and filtering out noisy, drifting user clicks, platforms can deliver more relevant suggestions. For digital businesses, improved recommendation precision directly supports higher user engagement, better conversion rates, and reduced reliance on tracking personal user history.

Organizations operating e-commerce or content platforms with high shares of anonymous traffic should consider evaluating graph-based architectures within their recommendation pipelines. Implementation teams should adopt hybrid session embeddings that balance immediate user actions with overall session context. Because the source notes that short training steps prevent overfitting on brief sessions, engineering teams should tune training schedules accordingly.

Confidence in these findings is supported by consistent empirical gains across multiple real-world benchmark datasets and varied session lengths. However, decision-makers should note that the evaluation relied on historical e-commerce clickstream data where sessions were filtered to exclude single-click visits and rare items. Further testing in live operational environments and with auxiliary product metadata, such as item categories and descriptions, is recommended before full-scale deployment.

Center for Research on Intelligent Perception and Computing Institute of Automation, Chinese Academy of Sciences Microsoft Tongji University University of Chinese Academy of Sciences University of Science and Technology Beijing graph neural networks network architecture recommendation methods session-based recommendation graph neural networks session graphs

arXiv: 1811.00855 CRIPAC-DIG/SR-GNN

Paper: Session-based Recommendations with Recurrent Neural Networks, Balázs Hidasi et al. (2016). Read this earlier GRU-based session recommender first to understand the sequential-model baseline that the source’s graph approach is designed to improve upon.

Paper: Graph Neural Networks in Recommender Systems: A Survey, Shiwen Wu et al. (2020). This later survey places session-based graph recommendation within a broader taxonomy of GNN methods, extending the source’s focused approach into a wider research landscape.

Cover for Session-based Recommendation with Graph Neural Networks

Abstract

The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embedding and take complex transitions of items into account, we propose a novel method, i.e. Session-based Recommendation with Graph Neural Networks, SR-GNN for brevity. In the proposed method, session sequences are modeled as graph-structured data. Based on the session graph, GNN can capture complex transitions of items, which are difficult to be revealed by previous conventional sequential methods. Each session is then represented as the composition of the global preference and the current interest of that session using an attention network. Extensive experiments conducted on two real datasets show that SR-GNN evidently outperforms the state-of-the-art session-based recommendation methods consistently.

Topics

graph neural networks network architecture recommendation methods real data sets

Keywords

session-based recommendation graph neural networks session graphs session embeddings graph-structured data collaborative filtering matrix factorization short-term attention priority model Yoochoose dataset

1 Introduction
2 Related Work
3 The Proposed Method
Notations
Constructing Session Graphs
Learning Item Embeddings on Session Graphs
Generating Session Embeddings
Making Recommendation and Model Training
4 Experiments and Analysis
Datasets
Baseline Algorithms
Evaluation Metrics
Parameter Setup
Comparison with Baseline Methods
Comparison with Variants of Connection Schemes
Comparison with Different Session Embeddings
Analysis on Session Sequence Lengths
5 Conclusions
References

Knowls

Knowl 1 — Problem Formulation and Session Graph Construction
model/method
In session-based recommendation, let $V = \{v_1, v_2, \dots, v_m\}$ denote the universe of $m$ unique items. An anonymous session $s$ is an ordered sequence of user item interactions $s = [v_{s,1}, v_{s,2}, \dots, v_{s,n}]$ , where $v_{s,i} \in V$ represents the item clicked at step $i$ .

Each session sequence $s$ is modeled as a directed graph $G_s = (V_s, E_s)$ , where the node set $V_s \subseteq V$ consists of all unique items appearing in $s$ . A directed edge $(v_{s,i-1}, v_{s,i}) \in E_s$ represents a sequential transition where a user clicks $v_{s,i}$ immediately after $v_{s,i-1}$ . When items appear multiple times, each edge is assigned a normalized weight equal to the count of that directed transition divided by the out-degree of the transition's source node.

The graph topology of session $s$ with $n$ unique items is encoded by a connection matrix $A_s \in \mathbb{R}^{n \times 2n}$ , defined as the horizontal concatenation of the outgoing adjacency matrix $A_s^{(\text{out})} \in \mathbb{R}^{n \times n}$ and the incoming adjacency matrix $A_s^{(\text{in})} \in \mathbb{R}^{n \times n}$ :

$A_s = \left[ A_s^{(\text{out})} \;\Vert\; A_s^{(\text{in})} \right]$

Every item $v \in V$ is mapped to a latent embedding vector $\mathbf{v} \in \mathbb{R}^d$ , where $d$ is the latent feature dimensionality.
Knowl 2 — Node Representation Learning via Gated Graph Neural Networks
model/method
To capture item transitions within a session graph $G_s$ containing $n$ nodes, SR-GNN updates latent node vectors $\mathbf{v}_i \in \mathbb{R}^d$ across propagation steps $t = 1, \dots, T$ using a Gated Graph Neural Network (GGNN) parameterized by trainable matrices $H \in \mathbb{R}^{d \times 2d}$ , $W_z, W_r, W_o \in \mathbb{R}^{d \times d}$ , $U_z, U_r, U_o \in \mathbb{R}^{d \times d}$ , and bias vector $\mathbf{b} \in \mathbb{R}^d$ .

For node $v_{s,i}$ at propagation step $t$ , the information propagation and gated recurrent update equations are:

$\mathbf{a}_{s,i}^t = A_{s,i:} \left[ \mathbf{v}_1^{t-1}, \dots, \mathbf{v}_n^{t-1} \right]^\top H + \mathbf{b}$

$\mathbf{z}_{s,i}^t = \sigma\left( W_z \mathbf{a}_{s,i}^t + U_z \mathbf{v}_i^{t-1} \right)$

$\mathbf{r}_{s,i}^t = \sigma\left( W_r \mathbf{a}_{s,i}^t + U_r \mathbf{v}_i^{t-1} \right)$

$\tilde{\mathbf{v}}_i^t = \tanh\left( W_o \mathbf{a}_{s,i}^t + U_o \left( \mathbf{r}_{s,i}^t \odot \mathbf{v}_i^{t-1} \right) \right)$

$\mathbf{v}_i^t = (1 - \mathbf{z}_{s,i}^t) \odot \mathbf{v}_i^{t-1} + \mathbf{z}_{s,i}^t \odot \tilde{\mathbf{v}}_i^t$

where $A_{s,i:} \in \mathbb{R}^{1 \times 2n}$ is the $i$ -th row of the session connection matrix $A_s$ , corresponding to the outgoing and incoming edge weights of node $v_{s,i}$ ; $[\mathbf{v}_1^{t-1}, \dots, \mathbf{v}_n^{t-1}]^\top \in \mathbb{R}^{n \times d}$ is the stacked matrix of node vectors from the previous step; $\mathbf{a}_{s,i}^t \in \mathbb{R}^d$ represents the aggregated neighborhood context; $\mathbf{z}_{s,i}^t \in \mathbb{R}^d$ is the update gate; $\mathbf{r}_{s,i}^t \in \mathbb{R}^d$ is the reset gate; $\tilde{\mathbf{v}}_i^t \in \mathbb{R}^d$ is the candidate state; $\sigma(\cdot)$ is the element-wise sigmoid function; and $\odot$ denotes the Hadamard (element-wise) product.
Knowl 3 — Hybrid Session Embedding via Soft Attention
model/method
To represent an ongoing session $s = [v_{s,1}, v_{s,2}, \dots, v_{s,n}]$ , SR-GNN constructs a hybrid session embedding $\mathbf{s}_h \in \mathbb{R}^d$ by integrating the user's current local interest with their global long-term preference over the session.

The local session embedding $\mathbf{s}_l \in \mathbb{R}^d$ is defined as the latent vector of the last-clicked item in the sequence:

$\mathbf{s}_l = \mathbf{v}_n$

The global session embedding $\mathbf{s}_g \in \mathbb{R}^d$ aggregates all node vectors $\{\mathbf{v}_1, \dots, \mathbf{v}_n\}$ in the session graph using a soft-attention network conditioned on the last-clicked item $\mathbf{v}_n$ :

$\alpha_i = \mathbf{q}^\top \sigma\left( W_1 \mathbf{v}_n + W_2 \mathbf{v}_i + \mathbf{c} \right)$

$\mathbf{s}_g = \sum_{i=1}^n \alpha_i \mathbf{v}_i$

where $\mathbf{q}, \mathbf{c} \in \mathbb{R}^d$ and $W_1, W_2 \in \mathbb{R}^{d \times d}$ are trainable parameters, and $\alpha_i \in \mathbb{R}$ is the attention weight of item $v_{s,i}$ .

The hybrid session representation $\mathbf{s}_h$ is computed via a linear projection over the concatenation of the local and global embeddings:

$\mathbf{s}_h = W_3 \left[ \mathbf{s}_l \;\Vert\; \mathbf{s}_g \right]$

where $W_3 \in \mathbb{R}^{d \times 2d}$ projects the concatenated vector back to the $d$ -dimensional latent space.
Knowl 4 — Next-Item Recommendation Score and Training Objective
model/method
Given the hybrid session representation $\mathbf{s}_h \in \mathbb{R}^d$ , recommendation scores $\hat{z}_i$ for each candidate item $v_i \in V = \{v_1, \dots, v_m\}$ are calculated via inner product with the candidate item embedding $\mathbf{v}_i \in \mathbb{R}^d$ :

$\hat{z}_i = \mathbf{s}_h^\top \mathbf{v}_i$

The probability distribution $\hat{\mathbf{y}} = [\hat{y}_1, \dots, \hat{y}_m]^\top \in \mathbb{R}^m$ over all candidate items being the next click is computed using the softmax function:

$\hat{\mathbf{y}} = \text{softmax}(\hat{\mathbf{z}})$

$\hat{y}_i = \frac{\exp(\hat{z}_i)}{\sum_{j=1}^m \exp(\hat{z}_j)}$

The parameters of SR-GNN are trained end-to-end by minimizing the cross-entropy loss over the true next-clicked item:

$\mathcal{L}(\hat{\mathbf{y}}) = -\sum_{i=1}^m \left( y_i \log(\hat{y}_i) + (1 - y_i) \log(1 - \hat{y}_i) \right)$

where $\mathbf{y} \in \{0, 1\}^m$ is the one-hot ground-truth label vector indicating the actual next-clicked item. Optimization is performed using Back-Propagation Through Time (BPTT).
Knowl 5 — Experimental Setup and Evaluation Protocol
experimental setup
The evaluation utilizes two benchmark datasets:
1. Yoochoose (RecSys Challenge 2015): User click logs from an e-commerce platform over 6 months.
2. Diginetica (CIKM Cup 2016): Transactional click logs.
Preprocessing: Sessions of length 1 and items appearing fewer than 5 times are filtered out. For Yoochoose, the test set consists of sessions from the subsequent days. Two training splits are evaluated: the most recent $1/64$ fraction and $1/4$ fraction. For Diginetica, sessions from subsequent weeks form the test set. For an input sequence $s = [v_{s,1}, \dots, v_{s,n}]$ , training examples are generated incrementally as sequence-label pairs $([v_{s,1}], v_{s,2}), ([v_{s,1}, v_{s,2}], v_{s,3}), \dots, ([v_{s,1}, \dots, v_{s,n-1}], v_{s,n})$ .

Dataset Statistics:
- Yoochoose 1/64: 557,248 clicks, 369,859 training sessions, 55,898 test sessions, 16,766 items, average session length 6.16.
- Yoochoose 1/4: 8,326,407 clicks, 5,917,745 training sessions, 55,898 test sessions, 29,618 items, average session length 5.71.
- Diginetica: 982,961 clicks, 719,470 training sessions, 60,858 test sessions, 43,097 items, average session length 5.12.
Hyperparameters: Latent dimensionality $d = 100$ . Parameters are initialized from $\mathcal{N}(0, 0.1^2)$ . Optimization uses mini-batch Adam with batch size 100, initial learning rate 0.001 (decaying by a factor of 0.1 every 3 epochs), and $L_2$ regularization penalty $10^{-5}$ .

Evaluation Metrics:
- P@20 (Precision@20): Proportion of test instances where the ground-truth next item appears in the top-20 predicted items.
- MRR@20 (Mean Reciprocal Rank@20): Average reciprocal rank of the correctly recommended item within the top-20 list (set to 0 if rank $> 20$ ).

Knowl 6 — Recommendation Performance Comparison Across Datasets

data/table

The recommendation accuracy of SR-GNN was evaluated against traditional baseline methods (POP, S-POP, Item-KNN), factorization/Markov models (BPR-MF, FPMC), and deep sequential/attentive neural networks (GRU4REC, NARM, STAMP) on Yoochoose 1/64, Yoochoose 1/4, and Diginetica.

Method	Yoochoose 1/64		Yoochoose 1/4		Diginetica
Method	P@20	MRR@20	P@20	MRR@20	P@20	MRR@20
POP	6.71	1.65	1.33	0.30	0.89	0.20
S-POP	30.44	18.35	27.08	17.75	21.06	13.68
Item-KNN	51.60	21.81	52.31	21.70	35.75	11.57
BPR-MF	31.31	12.08	3.40	1.57	5.24	1.98
FPMC	45.62	15.01	–	–	26.53	6.95
GRU4REC	60.64	22.89	59.53	22.60	29.45	8.33
NARM	68.32	28.63	69.73	29.23	49.70	16.17
STAMP	68.74	29.67	70.44	30.00	45.64	14.32
SR-GNN	70.57	30.94	71.36	31.89	50.73	17.59

SR-GNN achieves the highest P@20 and MRR@20 across all three benchmark splits. Neural attentive architectures (NARM, STAMP) substantially outperform Markovian and factorization methods, but SR-GNN consistently exceeds them by modeling graph-structured item transitions within sessions rather than relying solely on sequential recurrent transitions.

Knowl 7 — Impact of Graph Connection Schemes on Session Recommendations
empirical result
To evaluate how session graph connectivity affects representation learning, three connection schemes were evaluated within the SR-GNN framework:
1. SR-GNN (Default): Models directed transitions between consecutive items within individual session graphs.
2. SR-GNN-NGC (Normalized Global Connections): Replaces session-level connection weights with normalized edge transition probabilities computed from a global directed item graph aggregated across all historical sessions.
3. SR-GNN-FC (Full Connections): Explicitly connects all high-order item pairs within a session with boolean weights, appending this full adjacency matrix to the consecutive connection matrix.
Empirical Findings:
- All three graph variants outperform or match state-of-the-art sequential baselines (STAMP and NARM), supporting the efficacy of graph-based session modeling.
- SR-GNN-NGC performs worse than SR-GNN because incorporating global cross-session edge weights dilutes the influence and integrity of the specific contextual transitions present in the ongoing session.
- SR-GNN-FC performs slightly worse than SR-GNN. Direct higher-order edges bypass intermediate sequential steps (e.g., jumping directly from $A$ to $C$ in $A \to B \to C$ ), indicating that intermediate transitions remain essential for accurate next-item prediction.
Knowl 8 — Impact of Session Representation Aggregation Mechanisms
empirical result
An ablation study evaluated four strategies for pooling node vectors into session embeddings:
1. SR-GNN-L (Local Only): Represents the session solely by the latent vector of the last-clicked item, $\mathbf{s} = \mathbf{v}_n$ .
2. SR-GNN-AVG (Average Pooling): Computes the session embedding by uniform mean pooling over all node vectors in the session, $\mathbf{s} = \frac{1}{n} \sum_{i=1}^n \mathbf{v}_i$ .
3. SR-GNN-ATT (Attention Only): Uses the soft-attention weighted global representation $\mathbf{s} = \mathbf{s}_g$ without explicitly combining the local embedding.
4. SR-GNN (Hybrid): Combines local and soft-attention global embeddings via linear projection, $\mathbf{s}_h = W_3 [\mathbf{s}_l; \mathbf{s}_g]$ .
Empirical Findings:
- The hybrid model (SR-GNN) achieves the highest P@20 and MRR@20 across Yoochoose 1/64, Yoochoose 1/4, and Diginetica, confirming the utility of combining current focus with global context.
- SR-GNN-ATT outperforms SR-GNN-AVG, demonstrating that soft attention successfully weights informative actions while downweighting noisy or exploratory clicks.
- SR-GNN-L achieves near-optimal performance (closely trailing SR-GNN-ATT and SR-GNN), because node vectors learned via GNN message passing already implicitly encode high-order graph neighborhood context into the final item embedding $\mathbf{v}_n$ .

Knowl 9 — Model Performance Across Different Session Sequence Lengths

data/table

To evaluate model stability with respect to sequence length, sessions from Yoochoose 1/64 (70.1% Short, 29.9% Long) and Diginetica (76.4% Short, 23.6% Long) were partitioned into "Short" (length $\le 5$ ) and "Long" (length $> 5$ ).

Method	Yoochoose 1/64		Diginetica
Method	Short	Long	Short	Long
NARM	71.44	60.79	51.22	45.75
STAMP	70.69	64.73	47.26	40.39
SR-GNN-L	70.11	69.73	49.04	50.97
SR-GNN-ATT	70.31	70.64	50.35	51.05
SR-GNN	70.47	70.70	50.49	51.27

Recurrent and attentive baselines (NARM and STAMP) degrade substantially on long sessions (e.g., NARM drops by 10.65% on Yoochoose 1/64 and 5.47% on Diginetica), demonstrating the vulnerability of sequential RNN architectures to long sequences and interest drift. In contrast, SR-GNN and its embedding variants (SR-GNN-L, SR-GNN-ATT) maintain stable performance across both short and long sessions.

Coverage note — None was omitted; all core architectural formulations, loss functions, experimental settings, baseline comparisons, connection scheme ablations, session embedding ablations, and sequence length analyses have been fully captured.

References

1.Cho, K.; Van Merriënboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. EMNLP 1724–1734.
2.Duvenaud, D.; Maclaurin, D.; Aguilera-Iparraguirre, J.; Gómez-Bombarelli, R.; Hirzel, T.; Aspuru-Guzik, A.; and Adams, R. P. 2015. Convolutional networks on graphs for learning molecular fingerprints. In NIPS.
3.Gori, M.; Monfardini, G.; and Scarselli, F. 2005. A new model for learning in graph domains. In IJCNN, volume 2, 729–734.
4.Grover, A., and Leskovec, J. 2016. Node2vec: Scalable feature learning for networks. In KDD, 855–864.
5.Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2016a. Session-based recommendations with recurrent neural networks. In ICLR.
6.Hidasi, B.; Quadrana, M.; Karatzoglou, A.; and Tikk, D. 2016b. Parallel recurrent neural network architectures for feature-rich session-based recommendations. In RecSys, 241–248.
7.Jannach, D., and Ludewig, M. 2017. When recurrent neural networks meet the neighborhood for session-based recommendation. In RecSys, 306–310.
8.Kipf, T. N., and Welling, M. 2016. Semi-supervised classification with graph convolutional networks. In ICLR.
9.Koren, Y., and Bell, R. 2011. Advances in collaborative filtering. In Recommender Systems Handbook. Springer. 145–186.
10.Koren, Y.; Bell, R.; and Volinsky, C. 2009. Matrix factorization techniques for recommender systems. Computer 42(8):30–37.
11.Li, Y.; Tarlow, D.; Brockschmidt, M.; and Zemel, R. S. 2015. Gated graph sequence neural networks. In ICLR.
12.Li, J.; Ren, P.; Chen, Z.; Ren, Z.; Lian, T.; and Ma, J. 2017a. Neural attentive session-based recommendation. In CIKM, 1419–1428.
13.Li, R.; Tapaswi, M.; Liao, R.; Jia, J.; Urtasun, R.; and Fidler, S. 2017b. Situation recognition with graph neural networks. In ICCV, 4183–4192.
14.Li, Z.; Ding, X.; and Liu, T. 2018. Constructing narrative event evolutionary graph for script event prediction.
15.Liu, Q.; Wu, S.; Wang, L.; and Tan, T. 2016. Predicting the next location: A recurrent model with spatial and temporal contexts. In AAAI, 194–200.
16.Liu, Q.; Zeng, Y.; Mokhosi, R.; and Zhang, H. 2018. Stamp: Short-term attention/memory priority model for session-based recommendation. In KDD, 1831–1839.
17.Mao, J.; Xu, W.; Yang, Y.; Wang, J.; Huang, Z.; and Yuille, A. 2015. Deep captioning with multimodal recurrent neural networks (m-rnn). ICLR.
18.Marino, K.; Salakhutdinov, R.; and Gupta, A. 2017. The more you know: Using knowledge graphs for image classification. In CVPR, 20–28.
19.Mikolov, T.; Karafiát, M.; Burget, L.; Černocký, J.; and Khudanpur, S. 2010. Recurrent neural network based language model. In INTERSPEECH, volume 2, 3.
20.Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In NIPS, 3111–3119.
21.Mnih, A., and Salakhutdinov, R. 2007. Probabilistic matrix factorization. In NIPS, 1257–1264.
22.Perozzi, B.; Al-Rfou, R.; and Skiena, S. 2014. Deepwalk: Online learning of social representations. In KDD, 701–710.
23.Rendle, S.; Freudenthaler, C.; Gantner, Z.; and Schmidt-Thieme, L. 2009. Bpr: Bayesian personalized ranking from implicit feedback. In UAI, 452–461.
24.Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L. 2010. Factorizing personalized markov chains for next-basket recommendation. In WWW, 811–820. ACM.
25.Sarwar, B.; Karypis, G.; Konstan, J.; and Riedl, J. 2001. Item-based collaborative filtering recommendation algorithms. In WWW.
26.Scarselli, F.; Gori, M.; Tsoi, A. C.; Hagenbuchner, M.; and Monfardini, G. 2009. The graph neural network model. TNN 20(1):61–80.
27.Serban, I. V.; Sordoni, A.; Bengio, Y.; Courville, A.; and Pineau, J. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In AAAI, 3776–3784.
28.Shani, G.; Brafman, R. I.; and Heckerman, D. 2002. An mdp-based recommender system. In UAI, 453–460.
29.Tan, Y. K.; Xu, X.; and Liu, Y. 2016. Improved recurrent neural networks for session-based recommendations. In DLRS, 17–22.
30.Tang, J.; Qu, M.; Wang, M.; Zhang, M.; Yan, J.; and Mei, Q. 2015. Line: Large-scale information network embedding. In WWW, 1067–1077.
31.Tuan, T. X., and Phuong, T. M. 2017. 3d convolutional networks for session-based recommendation with content features. In RecSys, 138–146.
32.Wu, C., and Yan, M. 2017. Session-aware information embedding for e-commerce product recommendation. In CIKM, 2379–2382.
33.Yu, F.; Liu, Q.; Wu, S.; Wang, L.; and Tan, T. 2016. A dynamic recurrent basket recommendation model. In SIGIR. ACM.
34.Zhang, Y.; Dai, H.; Xu, C.; Feng, J.; Wang, T.; Bian, J.; Wang, B.; and Liu, T.-Y. 2014. Sequential click prediction for sponsored search with recurrent neural networks. In AAAI, 1369–1376.

Citation

MLA

Wu, S., et al. “Session-Based Recommendation with Graph Neural Networks”. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 346–53, https://doi.org/10.1609/aaai.v33i01.3301346.

APA

Wu, S., Tang, Y., Zhu, Y., Wang, L., Xie, X., & Tan, T. (2019). Session-Based Recommendation with Graph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 346–353. https://doi.org/10.1609/aaai.v33i01.3301346

Chicago

Wu, S., Y. Tang, Y. Zhu, L. Wang, X. Xie, and T. Tan. 2019. “Session-Based Recommendation with Graph Neural Networks”. Proceedings of the AAAI Conference on Artificial Intelligence 33 (01): 346–53. https://doi.org/10.1609/aaai.v33i01.3301346.

Harvard

Wu, S. et al. (2019) “Session-Based Recommendation with Graph Neural Networks”, Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), pp. 346–353. Available at: https://doi.org/10.1609/aaai.v33i01.3301346.

Vancouver

1. Wu S, Tang Y, Zhu Y, Wang L, Xie X, Tan T (2019) Session-Based Recommendation with Graph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence 33:346–353

BibTeX

@article{Wu_2019, title={Session-Based Recommendation with Graph Neural Networks}, volume={33}, ISSN={2159-5399}, url={http://dx.doi.org/10.1609/aaai.v33i01.3301346}, DOI={10.1609/aaai.v33i01.3301346}, number={01}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, publisher={Association for the Advancement of Artificial Intelligence (AAAI)}, author={Wu, Shu and Tang, Yuyuan and Zhu, Yanqiao and Wang, Liang and Xie, Xing and Tan, Tieniu}, year={2019}, month=July, pages={346–353} }

Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Session-based Recommendation with Graph Neural Networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Problem Formulation and Session Graph Construction

Knowl 2 — Node Representation Learning via Gated Graph Neural Networks

Knowl 3 — Hybrid Session Embedding via Soft Attention

Knowl 4 — Next-Item Recommendation Score and Training Objective

Knowl 5 — Experimental Setup and Evaluation Protocol