Over the past few years, session-based recommender systems (SBRSs) have garnered growing attention from both researchers and industry professionals. Diverging from conventional recommender systems, SBRSs rely on user-item interactions occurring within the current session to forecast recommendations for the next item. Academic literature has introduced numerous methods for developing SBRSs, including those that leverage Graph Neural Networks (GNNs). Due to their ability to handle intricate relationships such as in social network analysis, GNN has demonstrated their efectiveness in comparison to alternative approaches. This paper provides a first review of the prominent SBRSs based on GNNs and sheds some light on further research. This initiative seeks to support researchers by providing valuable insights into the latest advancements and prospects of GNNs, with the overarching goal of enhancing session-based recommender systems.
posed in the context of SBRS [2, 5]. Notably, there has Sequence-aware recommender systems (SARS) [1] are been a growing trend of exploring the integration of designed to consider the chronological sequence of user Graph Neural Networks (GNN) in the advancement of interactions or activities when generating personalized SBRS [2, 6, 7, 8]. recommendations. In contrast to conventional recom- GNN is a type of artificial neural network designed to mender systems that treat user-item interactions as sep- work with graph-structured data [9]. Graphs are comarate events, SARS analyse the temporal patterns and posed of nodes linked by edges, and GNNs are customsequences of user actions. These systems are particularly designed to process and analyse data represented in this relevant in scenarios where the order of user interactions structure [6, 7]. GNNs have gained significant popularity is crucial in understanding user preferences and intents. in various domains, including social network analysis,
As a sort of SARS, a Session-Based Recommender Sys- recommendation systems, molecular chemistry, and nattem (SBRS) [2, 3] focuses on providing personalized rec- ural language processing [10]. They demonstrate excepommendations based on short-term interactions within tional proficiency in tasks requiring the comprehension a user’s current session. Rather than relying on the en- of relationships and interdependencies among intercontirety of a user’s historical behaviour, SBRSs concentrate nected entities within a graph. on the user’s immediate context and preferences within In the absence of similar work, this paper aims to rea single session. view the main proposed approaches of SBRS based on
Session-based recommender systems are commonly GNN. However, its contributions can be summarized in used in various applications [4], including e-commerce, the following: news websites, and online advertising, where user interactions occur in real time, and personalized recommendations can greatly enhance user engagement and satisfaction. 1. Exhibit the use of diferent GNN architectures for
building SBRS, 2. Review of GNN-Based approaches for SBRS, 3. and, highlighting some open research issues for future directions on SBRS. approaches is given. Section 6 emphasizes certain emerging future directions on SBRS mainly with GNN. Finally, Section 7 concludes this review paper and unveils our nearest research.
Within the expanding landscape of recommender systems, session-based recommender systems (SBRS) are emerging as an intriguing trend [11]. SBRSs fall under the category of sequence-aware recommendation systems (SARS) [1] and specifically consider short-term user preferences or intentions. They make use of user interactions with previously viewed items during the current session to anticipate what item or items might be of interest next. This could involve predicting the next video to watch, the point of interest (POI) to visit, or the product to purchase. A crucial element of their recommendation process involves tracking the relationships and dependencies between time-stamped interactions of users with items during a session.
Currently, SBRSs are gaining considerable attention from both researchers and industry professionals [2]. They are becoming increasingly viable recommendation tools across a wide spectrum of real-world domains, including e-commerce, tourism, and leisure. For instance, as illustrated in Fig. 1, SBRSs can provide suggestions for future purchases based on a series of earlier user activities within the same online shopping session.
The primary attributes of SBRSs [2, 12, 3] are, (i) Identification of users is not feasible (because the majority of website trafic is from first-time or non-logged-in users), (ii) due to (i), Long-term user preferences information are unavailable, and (iii) User preferences must be deduced based on a limited set of consecutive interactions within the session.
Formally, a session is a non-empty bounded observations list of user-item interactions generated over a continuous timespan that may be linked with a particular user via some form of identification (e.g., user ID, session-ID, cookie). It can be represented as : = {1, 2, . . . , ||}.
In the literature, a multitude of methods and models have been proposed for the development of SBRSs [2]. Among them, Graph Neural Networks (GNNs) stand out as a remarkable method that ofers enhanced performance for SBRSs. GNNs have showcased their capacity to model intricate transitions occurring within and between sessions, treating them as graph-structured data and employing deep neural networks. The following section will provide an introduction to GNNs.
Graph neural networks (GNNs) are a novel advanced model for employing deep learning techniques that operate on graph-structured data, as for social networks (relationships domain), citation networks (scientific domain), protein-protein interaction networks (biology domain), user-item interaction networks (recommendation domain), and others [10]. In contrast to traditional neural networks, which operate on structured input data like images or sequences, GNNs are designed to operate on non-Euclidean data, where data points are represented as nodes in a graph and the edges connecting nodes symbolize the connections and associations between them.
Over the past few years, GNNs have become popular because of their handling of complex and highly interconnected data structures. They use message-passing algorithms to propagate information through the graph, updating node representations based on the aggregated representations of their neighbouring nodes. GNNs can be used for a wide range of tasks, including node prediction, link prediction, and graph classification [13].
Recently, the recommendation field has benefited from GNNs for enhancing its systems [14, 6]. In the context of SBRSs, several approaches based on GNN have been proposed by incorporating various GNN architectures for efective proposals [ 2]. Using GNN have improved recommendation accuracy by modelling sequential interactions between users and items. Generally, graphs for SBRSs are directed in such a way that they maintain the order in which user-item interactions occur (see Fig. 2).
In the absence of a comprehensive survey paper on Session-Based Recommender Systems (SBRS) utilising Graph Neural Networks (GNNs), this work serves this purpose by conducting a review of the most significant contributions in this field.
From a technical perspective, this section ofers a reference framework which serves as a core model for any
Outgoin
The literature in the last years has witnessed an increasing adoption of GNN for developing SBRSs via diferent proposed frameworks and architectures. Fig. 3 depicts a reference architecture for SBRS based on GNN. This showcases three main steps: 1. Graph Construction. To apply GNN in the sequential recommendation, it is essential to convert sequential data into a sequential graph format in which nodes and edges can be featured with information vectors as their hidden states. 2. Information Propagation. To capture the transition patterns, a propagation mechanism should be adopted. Generally, each GNN layer is designed to execute a specific set of operations on each node within the graph, namely: • Message Passing: this is defined as the procedure of collecting the features of neighbouring nodes, transforming them, and then transmitting them to the source node.
In parallel, this process is repeated for every node in the graph, thereby ensuring a comprehensive examination of all neighbourhoods by the conclusion of this step. • Aggregation: neighbourhood aggregation involves the exchange of data between nodes within their respective neighbourhoods. A source node, equipped with its initial embeddings, receives input from its neighbours, and this information is transmitted through edge neural networks. • Update: upon receiving these messages, each node updates its features by considering both its current features and the aggregated information received from its neighbouring nodes.
From a practical standpoint, the session-based recommender systems development process with GNN can be generally carried out using the main following activities: 1. Data Representation: Arrange your data into sessions, where each session denotes a series of user engagements (e.g., clicks, views, purchases) with items. Create a graph structure in which items are depicted as nodes, and the connections between items, like co-occurrences, similarities, or pertinent associations, are represented as edges.
Encode both session-specific details and item attributes as features for the nodes in the graph. 2. Graph Neural Network Architecture: Select an appropriate Graph Neural Network (GNN) architecture for session-based recommendation (see Section 4). Customize the GNN to efectively manage sequential data and dynamic graph structures. 3. Session Representation: Leverage the GNN to acquire session representations through the aggregation of information from items within a session. Achieve this by executing message-passing across the graph while considering the relationships between items. Additionally, contemplate the integration of time decay or attention mechanisms to assign greater significance to recent interactions. 4. Item Embeddings: Derive item embeddings by employing the GNN individually to process each item within the graph, efectively capturing their interconnections with other items within the context of sessions. 5. Candidate Generation: Using the acquired session and item representations, produce a roster of potential items for recommendation. This can be accomplished by scoring items with a function that takes into account their pertinence to the context of the session. 6. Loss Function and Training: Establish a suitable loss function for your recommendation objective, such as the Bayesian Personalized Ranking (BPR) loss or Triplet loss, which incentivizes the model to prioritize positive items over negative ones in ranking. Proceed to train the GNN-based recommender system using session-level data and historical user-item interactions. 4 i → 1 i → 3 i → 2 i → 1 i h1 h2 h3 g n i s s a P e g a s s e M n o it a reggg A e t a d p U h'1 h'2 h'3
Multiple iterations Session
Graph construction
Input
GNN Model 7. Evaluation: Assess the recommendation system’s intensively used in the SBRS field [ 2] to handle complex performance by employing metrics such as Recall, user-item interactions and provide accurate next-item Precision, NDCG (Normalized Discounted Cumu- recommendations. Formally, this section will separately lative Gain), or Hit Rate. To gauge the model’s explain the utilization of each architectural design in the ability to generalize efectively, partition your SBRS context. data into training, validation, and test sets for evaluation purposes. 4.1. Gated Graph Neural Networks 8. Regularization and Hyper-parameter Tuning: Implement regularization methods like dropout or L2 regularization to counteract over-fitting. Conduct experiments to fine-tune hyper-parameters, including learning rate, the number of layers, hidden dimensions, and the number of epochs, to optimize the model’s performance.
It is worth emphasizing that the decision regarding the GNN architecture and hyper-parameters should be thoughtfully tailored to suit the unique characteristics of your dataset and the specific requirements of your problem domain. The optimal choices may vary based on factors such as data distribution, the nature of user interactions, and the desired level of recommendation accuracy. Therefore, a thorough understanding of your specific context is crucial in making these decisions.
including mainly gated graph neural networks (GGNNs) [15], graph convolutional networks (GCNs) [16], and graph attention networks (GATs) [17]. GGNNs use recurrent neural networks with GRU (Gated Recurrent Unit) cells to propagate information through the graph. While GCNs are one of the most commonly used GNN architectures and are based on a convolutional-like operation that aggregates information from neighbouring nodes. However, GATs use attention mechanisms to weigh the importance of each neighbouring node when updating a node’s representation. These three architectures are
tension of Scarselli et al.’s previous work on GNN [19]. They use gated recurrent units (GRUs) and compute gradients using the back-propagation through time (BPTT) algorithm.
Let = (, ) be a directed graph, with and representing the set of vertices and edges respectively. An edge is a tuple that contains the source and target nodes (cf. Fig. 2). The connections between the nodes are better stated in an adjacency matrix ∈ R| |× 2| |, which is a matrix structured with its rows and columns corresponding to the labelled vertices in the graph and a 0 or 1 in position (, ) indicating the nature of adjacency between and (Fig. 2).
Typically, an SBRS approach based on GGNN [2] first builds a directed graph from all the previously sorted items within a session, with the direction of each edge denoting the sequence of subsequent interactions. Then, GGNN processes the session graph successively to produce the embedding of node , specifically the embedding of the appropriate interaction . At the end of the process, the embeddings of all interactions are acquired, which are subsequently utilized to construct an embedding of the session context in order to develop recommendations for that session. A GRU is specifically utilised in GGNN to learn each node’s embedding by updating the embedding recurrently. In particular, the embedding (or hidden state) ℎ of node at step is updated (eq. 1) by the preceding hidden state of itself and its neighbourhood nodes, i.e., ℎ(− 1) and ℎ(− 1), ℎ = ︁( ℎ(− 1) , ℎ(− 1), ︁)
(1) ∑︁ ∈ () where () is the set of neighbourhood nodes of in the session graph, and is the adjacency matrix constructed based on the session graph. Upon reaching a stable equilibrium over several iterations, the hidden state at the final step of node is considered as its embedding .
themselves as one of the most widely adopted and highly efective GNN models. The convolution principle is borrowed from CNN [16] enabling GCNs to aggregate data based on local graph neighbourhoods [9].
GCN-based SBRSs primarily employ the pooling operation to incorporate information from node ’s neighbourhood node in the graph, and then to proceed to update the hidden state of as formulated by equations 2, 3: english ˆ
ℎ = ︁({︁ h(− 1), ∈ () }︁)︁ britishwhere () is the set of neighbourhood nodes of node . Various particular pooling techniques (such contingent on particular situations. Subsequently, the combined neighbourhood information can be integrated into the iterative update of the node’s hidden state : english ℎ = ℎ(− 1) + ℎˆ
ifnal hidden state of node british is adopted (eq. 3) as its embedding britishn. british
Graph ATtention Networks (GATs) utilize an attention mechanism to measure the significance of nearby nodes’ features while updating a node’s representation. The attention mechanism allows GATs to efectively model intricate relationships between nodes in a graph. Every node in a GAT has a hidden representation (features) that is updated depending on the neighbour representations using an attention mechanism [17].
(2) (3) as mean pooling and max pooling) can be employed, adopted as its embedding n.
The fundamental principle of GATs is to learn an atten- interests with long-term preferences to predict next betion coeficient for each nearby node that specifies how much weight to give to that node’s representation while updating the target node’s representation. The attention haviours of users. Another work in [21] proposed a graph contextualized self-attention network (GC-SAN) which begins by constructing dynamic directed graphs from coeficients are learned using a single-layer neural network that takes as input the features of the target node as well as the features of its neighbouring nodes. A softmax function is used to calculate the attention coeficients, ensuring that the weights given to each neighbouring node add up to one.
Moreover, the GAT computes numerous sets of weights for every node using multiple attention heads, which aids in capturing various facets of the graph structure. The multi-head attention enables the model to capture diverse patterns of relationships between nodes by using multiple parallel attention mechanisms, each with its weight matrix. Each attention head’s output is combined and sent through a non-linear activation function.
GAT in which the general module attention can be designated to diferent operations including self-attention, multi-head attention, etc.
english h = ︁({︁ h(− 1), ∈ () }︁)︁ (4) britishIn essence, the operations within can be partitioned into two stages: (1) computing the significance weights for each neighbouring node, and (2) aggregating the hidden states of neighbouring nodes based on their significance weights.
across multiple attention layers, the hidden state of each node within a session graph at the final layer is
presents the state-of-the-art methods for developing session-based recommendations that have used GNN with one or more mentioned architectures above. Three research categories can be distinguished, namely:
Initially, SBRS only used directed graphs to model each session sequence. SR-GNN [20] is the pioneer model that has utilized GNN in SBRS development for obtaining the latent vectors of nodes in each session graph. It is based on GGNN architecture including an attention mechanism layer. Aside from capturing the complex structure and transitions between items within a session, SR-GNN employs a strategy that combines current session sequences and then applies GGNN to pick up Just recently and innovatively, the authors in [27] local dependencies between items including contextual propose AutoGSR, a NAS (Neural Architecture Search) information of sessions. Next, a self-attention network framework for automatically searching suitable graph (SAN) has been used to capture global dependencies be- architectures to be adopted in diferent session-based rectween input and output sequences without regard to ommendation scenarios. To determine the optimal graph their distances. Finally, the local short-term and global neural network architecture, two novel GNN operations self-attended relationships are combined for the session (namely, Relational GGNN and Mixup which combines representation. the relational GGNN with the relational GAT) are added
In [22], the authors propose GACOforREC model for to build a complete and expressive search space, and a session-based recommendation in order to handle the diferentiable search algorithm is used. As part of Autolong-term and short-term preferences of users and pre- GSR, learning items are associated with meta-knowledge serve the hierarchy of potential preferences. This model that contains comprehensive session information. has used convolution operations of GCN to understand the sequence within the session and the spatial character- 5.2. Graph structure enrichment istics within the network for capturing the user’s shortterm preferences. For learning long-term preferences, it Other methods in the field of SBRSs aim to enhance the applied ConvLSTM, a variant of the LSTM network. In relationships and information within the session graph addition, GACOforREC proposes a new pair adaptive at- by incorporating data from other sessions or supplementention mechanism (Long-Attention and Short-Attention) tary sources. A-PGNN model [28] is proposed to capbased on GCNs to give consideration to the impact of ture firstly the complex dependencies between the sesdiferent propagation distances of GCNs. To improve the sion’s items, and at the same time, the user’s long-term model’s hierarchical learning of a variety of preferences, performance by modelling the efect of historical sesON-LSTM has been introduced, it is a network struc- sions on the ongoing session. In addition to the attention ture that focuses more on hierarchy and neuron ordering. mechanism and Transformer network, A-PGNN has used This ordering is crucial to the comprehensive perception GGNN with closed recurrent units (GRUs) to update inforof the model’s user preferences for accurate recommen- mation about each node’s hidden state. SGNN-HN [29] dations. Another GCN-based SBRS Model called AU- applies a star graph neural network (SGNN) to model TOMATE was presented in [23]. It integrates a graph the complex transition relationship between items (adjaconvolutional layer based on Auto-Regressive Moving cent and non-adjacent) in a current session to generate Average (ARMA) filters. It can capture complex trans- accurate item embeddings. To represent the propagaformations between items through sessions modelled as tion information, GGNN is used to feed satellite nodes graph-structured data. The core principle behind AUTO- in the star graph. In this work, the over-fitting probMATE revolves around leveraging the ARMAConv layer, lem of GNN in SBRS is tackled using highway networks which allows us to merge enduring user preferences with (HN) which can dynamically merge information from real-time session interests to generate the graph transfer item embeddings before and after multi-layer SGNNs. signal. Subsequently, the generated item embeddings are aggre
Thereafter, TAGNN [24] captures rich item transitions gated through an attention mechanism to represent a in sessions and learns the node vectors using GGNN. It user’s final preference which is then combined with her extends SR-GNN by proposing a novel target-aware atten- recent interest expressed (i.e., last clicked items in the tion mechanism that learns diferent user interests with session) for next-items prediction. In [30], the authors respect to varied target items. In [25], the authors model propose a position-aware gated graph attention network user-item sessions using GAT in what is called PSR-GAT (PA-GGAN) model to assign the position information for Personalized Session-based Recommendation using of items in the session sequences into session graphs. Graph Attention Networks. The PSR-GAT model com- The enhanced GGNN that underpins this model has been bines a user’s past preferences at several scales in addition supplied with a self-attention mechanism for aggregatto the available information in item transitions. Further- ing features from nodes. Similarly, [31] proposes MKMmore, the new model of Knowledge-enhanced Graph SR which incorporates user Micro-behaviours and item Attention Network for Session-based Recommendation Knowledge simultaneously into Multi-task learning for (KGAT-SR) is proposed in [26]. It exploits the knowledge Session-based Recommendation. Instead of a sequence about items via a knowledge graph attention network to of items, a session in MKM-SR is modelled on the microgenerate a knowledge-enhanced session graph (KESG). behaviour level accompanied by a sequence of operations The latter is aggregated via weighted graph attention. on each item to suficiently capture the transition patWhile the node features and graph topology in the graph tern in the session. GGNN and GRU are used to learn are used to generate appropriate session embedding for item and operation embeddings, and the multi-task learnrecommending the next item. ing processes involve learning knowledge embeddings to promote the major task of SBRS. Otherwise, to add the independence between each pair of factors. As each other information sources (such as social information) in item is represented with independent factors, an attensession-based recommendation processes, the authors in tion mechanism is designed to determine the user’s intent [32] propose a novel session-based social recommenda- regarding the diferent factors of each item. Subsequently, tion model named GNNRec, in which a gated graph neu- the session embedding is computed by aggregating all ral network (GGNN) is first used to represent the current the items that have been assigned factor-level attention. session information of the user. Next, a GAT is utilized to The user intents at the factor level are taken into account aggregate social information on users and their friends on when determining the purpose of a session. As a solution social networks for modelling user’s interests. However, for the over-smoothing issue associated with sessionto make session-based social recommendations more ef- based recommendation using GNN (i.e., all nodes reach ifciently, the research work in [ 33] has proposed a model the same value), SR-HGNN [37] is proposed. It is based called Social-aware Eficient Recommender (SERec) that a hybrid-order GGNN, where insignificant patterns are implements the SEFrame framework in which a hetero- avoided and complex interactions between items are capgeneous knowledge graph is constructed from the social tured. Furthermore, an attention mechanism is adopted network and historical user behaviours. In this study, a to learn diferent weights of orders in the propagation. GGNN is utilized to derive a contextualized feature vec- Very recently, GPAN or Graph Positional Attention tor by integrating information from neighbouring nodes Network has been presented in [38]. It is based on posiand the initial feature vectors. tion attention in response to the use of the user’s higher
Recently, and based on GNN and attention networks, order features and to address the impact of item posia session-enhanced graph neural network (SE-GNNRM) tion information on the current session, enhancing premodel has been proposed in [34]. During the encoding dictions in SBRSs. Still with the use of hyper-graphs, phase of this model, the intricate transitional relation- the study in [39] has introduced HyperS2Rec, where it ships between items and item features are captured in takes into account both item consistency and sequential GNN and SAN. Then, the attention mechanism is em- item dependence simultaneously. This model leverages ployed to combine short-term and long-term preferences hypergraph-structured data through HGCN and captures to construct a global session graph. Furthermore, a GAT sequential information using GRU to collectively model is devoted to recognising features between similar ses- user preferences. In this proposal, the reversed position sions and integrating similarity information between ses- embedding mechanism and soft attention mechanism are sions in which GGNNs are used to extract node features. combined to derive session representations.
between items in real-world scenarios, the authors in [35] have proposed DHCN (Dual Channel Hyper-graph Based on the literature review conducted in the preceding Convolutional Networks). This model is based on a hyper- section, we can initiate a discussion on specific aspects graph using convolution operations and integrating self- and identify potential avenues for future research. supervised learning to generate high-quality sessionbased recommendations. 6.1. Discussion
In order to acquire more refined item representations, GNN-based models can extract additional information Table 1 displays the primary GNN architecture employed from high-order neighbours over the graph structure. in each reviewed study, notwithstanding the incorporaNew methods aim to enhance the recommendation by tion of additional layer types, such as attention mechmodelling the high-order relations in session data. Very anisms. In addition, Fig 4 shows an overview of the recently, Disen-GNN [36] brought up the problem of distribution of the usage of GNN architectures among anonymous user purpose in a session and demonstrated the summarized approaches. the promise of using disengaged learning to solve this It is clear that GGNN architecture has been more approblem. This model captures the session’s intent by plied than GCN and GAT to address various issues raised considering factor-level attention for each item in the in the cited research works. Evenly, attention mechasession, and it employs a disentangled learning technique nisms represent another common point between most of to transform item embeddings into multiple-factor em- the mentioned works. beddings. The embedding of each factor is learned sepa- Furthermore, to harness the capabilities of GNN arrately via GGNN based on the item adjacent similarity chitectures, some research work have undertaken the matrix computed for each factor. Additionally, the dis- approach of amalgamating two or more GNN architectance correlation is employed as a means of enhancing tures.
Also, these techniques may not adequately capture intricate dependencies and still have poor ability against the cold start problem.
Graph Neural Networks (GNNs) serve as a potent instrument for examining and characterizing intricate data structures. Nonetheless, they continue to pose certain dificulties, such as issues related to scalability, and adaptability to novel graph instances. Additionally to what is mentioned in [2], we outline the main potential directions for future research and development in sessionbased recommender systems, including when developing them with advanced GNN-based models: • Contextual Embeddings: SBRSs may benefit from incorporating additional contextual information such as user demographics, device information, or contextual data related to the session. This requires investigation of advanced methods for grasping and leveraging session context more effectively, including the application of contextual embeddings like transformer-based models (such as BERT and GPT) to gain a deeper comprehension of user intentions and preferences. • Sequential Attention Mechanisms: Develop enhanced attention mechanisms that can handle long-term dependencies within sessions more effectively, potentially by incorporating memoryaugmented neural networks or adaptive attention mechanisms. • Hybrid Models: Explore hybrid recommender systems that merge session-based techniques with conventional collaborative filtering or contentbased strategies, efectively harnessing both short-term and long-term user preferences. • Transfer Learning: Examine transfer learning methodologies to utilize insights gained from one
The field of session-based recommender systems (SBRS) research is thriving, marked by a continuous stream of innovative techniques and emerging approaches. Among them, graph neural networks (GNN) have emerged as a powerful deep learning technique to perform inference on non-Euclidean data described by graphs. In this paper, we present contemporary GNN architectures and conduct a comprehensive review of prominent approaches that employ GNNs for advancing Session-Based Recommender Systems (SBRSs). Furthermore, we have elucidated potential research directions in this evolving field for future exploration.
Our forthcoming research will concentrate on comparing and evaluating the performance of these reviewed approaches.
ntic55069.2022.10100590. [20] S. Wu, Y. Tang, Y. Zhu, L. Wang, X. Xie, T. Tan, [11] D. Jannach, B. Mobasher, S. Berkovsky, Re- Session-based recommendation with graph neusearch directions in session-based and sequen- ral networks, in: Proceedings of the Thirty-Third tial recommendation, User Modeling and User- AAAI Conference on Artificial Intelligence and Adapted Interaction 30 (2020) 609–616. URL: Thirty-First Innovative Applications of Artificial Inhttps://doi.org/10.1007/s11257-020-09274-4. doi:10. telligence Conference and Ninth AAAI Symposium 1007/s11257-020-09274-4. on Educational Advances in Artificial Intelligence, [12] W. Song, S. Wang, Y. Wang, S. Wang, Next-item AAAI’19/IAAI’19/EAAI’19, AAAI Press, 2019. URL: recommendations in short sessions, in: Proceed- https://doi.org/10.1609/aaai.v33i01.3301346. doi:10. ings of the 15th ACM Conference on Recommender 1609/aaai.v33i01.3301346.
Systems, RecSys ’21, Association for Computing [21] C. Xu, P. Zhao, Y. Liu, V. S. Sheng, J. Xu, F. Zhuang, Machinery, New York, NY, USA, 2021, pp. 282– J. Fang, X. Zhou, Graph contextualized self291. URL: https://doi.org/10.1145/3460231.3474238. attention network for session-based recommendadoi:10.1145/3460231.3474238. tion, in: International Joint Conference on Artificial [13] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, Intelligence, 2019.
L. Wang, C. Li, M. Sun, Graph neural networks: [22] M. Zhang, Z. Yang, GACOforRec: Session-based A review of methods and applications, AI Open graph convolutional neural networks recommen1 (2020) 57–81. URL: https://www.sciencedirect. dation model, IEEE Access 7 (2019) 114077– com/science/article/pii/S2666651021000012. 114085. URL: https://doi.org/10.1109%2Faccess.2019. doi:https://doi.org/10.1016/j.aiopen. 2936461. doi:10.1109/access.2019.2936461. 2021.01.001. [23] H. Wang, G. Xiao, N. Han, H. Chen, Session-based [14] C. Gao, Y. Zheng, N. Li, Y. Li, Y. Qin, J. Piao, Y. Quan, graph convolutional arma filter recommendation J. Chang, D. Jin, X. He, Y. Li, A survey of graph neu- model, IEEE Access 8 (2020) 62053–62064. doi:10. ral networks for recommender systems: Challenges, 1109/ACCESS.2020.2984039. methods, and directions, ACM Trans. Recomm. [24] F. Yu, Y. Zhu, Q. Liu, S. Wu, L. Wang, T. Tan, Syst. 1 (2023). URL: https://doi.org/10.1145/3568022. TAGNN: target attentive graph neural networks doi:10.1145/3568022. for session-based recommendation, CoRR [15] Y. Li, D. Tarlow, M. Brockschmidt, R. S. Zemel, abs/2005.02844 (2020). URL: https://arxiv.org/abs/ Gated graph sequence neural networks, in: Y. Ben- 2005.02844. arXiv:2005.02844. gio, Y. LeCun (Eds.), 4th International Conference [25] Y. Xie, Z. Li, T. Qin, F. Tseng, K. Johannes, S. Qiu, on Learning Representations, ICLR 2016, San Juan, Y. L. Murphey, Personalized session-based recomPuerto Rico, May 2-4, 2016, Conference Track mendation using graph attention networks, in: Proceedings, 2016. URL: http://arxiv.org/abs/1511. 2021 International Joint Conference on Neural Net05493. works (IJCNN), IEEE, 2021. URL: https://doi.org/ [16] S. Zhang, H. Tong, J. Xu, R. Maciejewski, Graph 10.1109%2Fijcnn52387.2021.9533533. doi:10.1109/ convolutional networks: a comprehensive review, ijcnn52387.2021.9533533.
Computational Social Networks 6 (2019) 11. URL: [26] Q. Zhang, Z. Xu, H. Liu, Y. Tang, KGAT-SR: https://doi.org/10.1186/s40649-019-0069-y. doi:10. Knowledge-enhanced graph attention network for 1186/s40649-019-0069-y. session-based recommendation, in: 2021 IEEE [17] P. Velickovic, G. Cucurull, A. Casanova, A. Romero, 33rd International Conference on Tools with ArP. Lio, Y. Bengio, Graph attention networks, in: tificial Intelligence (ICTAI), IEEE, 2021. URL: https: International Conference on Learning Representa- //doi.org/10.1109%2Fictai52525.2021.00164. doi:10. tions, 2018. URL: https://openreview.net/forum?id= 1109/ictai52525.2021.00164. rJXMpikCZ. [27] J. Chen, G. Zhu, H. Hou, C. Yuan, Y. Huang, Autogsr: [18] Y. Li, D. Tarlow, M. Brockschmidt, R. S. Zemel, Neural architecture search for graph-based session Gated graph sequence neural networks, in: Y. Ben- recommendation, in: Proceedings of the 45th Ingio, Y. LeCun (Eds.), 4th International Conference ternational ACM SIGIR Conference on Research on Learning Representations, ICLR 2016, San Juan, and Development in Information Retrieval, SIGIR Puerto Rico, May 2-4, 2016, Conference Track ’22, Association for Computing Machinery, New Proceedings, 2016. URL: http://arxiv.org/abs/1511. York, NY, USA, 2022, pp. 1694–1704. URL: https: 05493. //doi.org/10.1145/3477495.3531940. doi:10.1145/ [19] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, 3477495.3531940.
G. Monfardini, The graph neural network model, [28] M. Zhang, S. Wu, M. Gao, X. Jiang, K. Xu, L. Wang, IEEE Transactions on Neural Networks 20 (2009) Personalized graph neural networks with atten61–80. doi:10.1109/TNN.2008.2005605. tion mechanism for session-aware recommendation, IEEE Transactions on Knowledge and Data En- order gated graph neural network for session-based gineering 34 (2022) 3946–3957. doi:10.1109/TKDE. recommendation, IEEE Transactions on Industrial 2020.3031329. Informatics 18 (2022) 1458–1467. doi:10.1109/ [29] Z. Pan, F. Cai, W. Chen, H. Chen, M. de Rijke, Star TII.2021.3091435.
graph neural networks for session-based recom- [38] L. Dong, G. Zhu, Y. Wang, Y. Li, J. Duan, M. Sun, mendation, in: Proceedings of the 29th ACM In- A graph positional attention network for sessionternational Conference on Information and Knowl- based recommendation, IEEE Access 11 (2023) 7564– edge Management, CIKM ’20, Association for Com- 7573. URL: https://doi.org/10.1109%2Faccess.2023. puting Machinery, New York, NY, USA, 2020, pp. 3235353. doi:10.1109/access.2023.3235353. 1195–1204. URL: https://doi.org/10.1145/3340531. [39] C. Ding, Z. Zhao, et al., Session-based recommenda3412014. doi:10.1145/3340531.3412014. tion with hypergraph convolutional networks and [30] J. Wang, Q. Xu, J. Lei, C. Lin, B. Xiao, Pa-ggan: sequential information embeddings, Expert SysSession-based recommendation with position- tems with Applications 223 (2023) 119875. doi:10. aware gated graph attention network, in: 2020 IEEE 1016/j.eswa.2023.119875.
International Conference on Multimedia and Expo [40] C. Sun, C. Li, X. Lin, T. Zheng, F. Meng, X. Rui, (ICME), 2020, pp. 1–6. doi:10.1109/ICME46284. Z. Wang, Attention-based graph neural networks: 2020.9102758. a survey, Artificial Intelligence Review (2023). URL: [31] W. Meng, D. Yang, Y. Xiao, Incorporating user https://doi.org/10.1007%2Fs10462-023-10577-2. micro-behaviors and item knowledge into multi- doi:10.1007/s10462-023-10577-2. task learning for session-based recommendation, in: [41] M. Lin, K. Wen, X. Zhu, H. Zhao, X. Sun, Graph Proceedings of the 43rd International ACM SIGIR autoencoder with preserving node attribute simiConference on Research and Development in Infor- larity, Entropy 25 (2023) 567. URL: https://doi.org/ mation Retrieval, SIGIR ’20, Association for Com- 10.3390%2Fe25040567. doi:10.3390/e25040567. puting Machinery, New York, NY, USA, 2020, pp. [42] L. Müller, M. Galkin, C. Morris, L. Ram1091–1100. URL: https://doi.org/10.1145/3397271. pasek, Attending to graph transformers, 2023. 3401098. doi:10.1145/3397271.3401098. arXiv:2302.04181. [32] C. Liu, Y. Li, H. Lin, C. Zhang, Gnnrec: gated graph neural network for sessionbased social recommendation model, Journal of Intelligent Information Systems (2022). URL: https://doi.org/10.1007/s10844-022-00733-5. doi:10.
1007/s10844-022-00733-5. [33] T. Chen, R. C.-W. Wong, An eficient and effective framework for session-based social recommendation, in: Proceedings of the 14th ACM International Conference on Web Search and Data Mining, WSDM ’21, Association for Computing Machinery, New York, NY, USA, 2021, pp. 400– 408. URL: https://doi.org/10.1145/3437963.3441792.
doi:10.1145/3437963.3441792. [34] L. Yin, P. Chen, G. Zheng, Session-enhanced graph neural network recommendation model (se-gnnrm), Applied Sciences 12 (2022). URL: https://www.mdpi.com/2076-3417/12/9/4314.
doi:10.3390/app12094314. [35] X. Xia, H. Yin, J. Yu, Q. Wang, L. Cui, X. Zhang, Selfsupervised hypergraph convolutional networks for session-based recommendation, Proceedings of the AAAI Conference on Artificial Intelligence 35 (2021) 4503–4511. URL: https://doi.org/10.1609%2Faaai.
v35i5.16578. doi:10.1609/aaai.v35i5.16578. [36] A. Li, Z. Cheng, F. Liu, Z. Gao, W. Guan, Y. Peng, Disentangled graph neural networks for session-based recommendation, ArXiv abs/2201.03482 (2022). [37] Y.-H. Chen, L. Huang, C.-D. Wang, J.-H. Lai, Hybrid