Unsupervised Graph Embeddings for Session-based Recommendation with Item Features

Andreas Peintner, Marta Moscati, Emilia Parada-Cabaleiro, Markus Schedl, Eva Zangerle · 2025

Read Unsupervised Graph Embeddings for Session-based Recommendation with Item Features by Andreas Peintner, Marta Moscati, Emilia Parada-Cabaleiro, Markus Schedl, Eva Zangerle (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-13763.txt.

Introduction

Peintner and four co-authors boost session-based recommendations by adding item features into graph embeddings, in a paper from twenty twenty-five. Session systems predict next clicks from a user's current browsing run. Top methods either graph the session flow or use item features for similarity. Their method, GCNext, fuses both strengths in one graph.

GCNext builds a rich graph of items that co-occur in sessions. Graph convolutions learn embeddings from that graph with no labels needed. Plugged into standard recommenders, it lifts both neighbor and neural methods. Gains reach near thirteen percent on a key ranking score across three data sets. The extension is flexible and easy to add to existing systems.

Catalog record

CREATORAndreas Peintner, Marta Moscati, Emilia Parada-Cabaleiro, Markus Schedl, Eva Zangerle
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-13763

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Unsupervised Graph Embeddings for Session-based Recommendation with Item Features Andreas Peintner, Marta Moscati, Emilia Parada-Cabaleiro, Markus Schedl, Eva Zangerle Abstract . In session-based recommender systems, predictions are based on the user’s preceding behavior in the session. State-of-the-art sequential recommendation algorithms either use graph neural networks to model sessions in a graph or leverage the similarity of sessions by exploiting item features. In this paper, we combine these two approaches and propose a novel method, G raph C onvolutional N etwork Ext ension (GCNext) , which incorporates item features directly into the graph representation via graph convolutional networks. GCNext creates a feature-rich item co-occurrence graph and learns the corresponding item embeddings in an unsupervised manner. We show on three datasets that integrating GCNext into sequential recommendation algorithms significantly boosts the performance of nearest–neighbor methods as well as neural network models. Our flexible extension is easy to incorporate in state-of-the-art methods and increases the MRR@20 by up to 12.79 % 12.79\% . Conference: ; September 23rd, 2022; Seattle, WA, USA CARS: Workshop on Context-Aware Recommender Systems (RecSys ’22), September 23rd, 2022, Seattle, WA, USA 1. Introduction Recommender systems (RecSys) traditionally leverage the users’ rich interaction data with the system. However, in some cases, such data are not available. Session-based recommender systems, in contrast, aim to predict the next item the user will interact with (e. g., click on, purchase, or listen to) only based on the preceding interactions in the current session. The task of session-based recommendation can be defined as follows. Consider the set X of all items in the catalog, x\in X being an individual item, and m=|X| being the number of items in the catalog. Given an interaction session [x_{1},x_{2},...,x_{n}] (ordered by timestamp), the goal is to predict a ranked list [y_{1},y_{2},...,y_{m}] of items with corresponding relevance scores to continue the session. The top- k values of the ranked list are chosen as recommendation candidates. As opposed to session-aware or sequential recommendations, the inputs to session-based RecSys are only items of the current session and their features; users are anonymous and no inter-session data is available. Current approaches for session-based recommendation leverage Recurrent Neural Networks (RNNs) ( Hidasi et al., 2016a ; Tan et al., 2016 ; Ren et al., 2019 ) , attention networks ( Li et al., 2017 ; Liu et al., 2018 ) , Graph Neural Networks (GNNs) ( Pan et al., 2021 ; Wu et al., 2019 ; Xu et al., 2019 ) , or transformer architectures ( Kang and McAuley, 2018 ; Sun et al., 2019 ; de Souza Pereira Moreira et al., 2021 ) . Also, classical nearest–neighbor methods have been used ( Ludewig et al., 2021 ; Ludewig et al., 2018 ; Jannach and Ludewig, 2017 ; Garg et al., 2019 ) . Most current methods focus on the sequential nature of sessions; RNNs and nearest–neighbor methods have dominated research in the past few years. Extensions to these models use additional item features to enrich the item representations. Item features capture contextual information (e. g., item category) which is relevant to the task of session-based recommendation, which itself can be considered a special case of context-aware recommender systems ( Quadrana et al., 2018 ) . However, recently, GNN models have been shown to outperform RNN- and nearest-neighbor–based methods ( Ludewig et al., 2021 ; Wu et al., 2019 ; Gwadabe and Liu, 2022 ) . Yet, to the best of our knowledge, no approach combines auxiliary item features and GNNs to learn informative embeddings for sequential models. In this paper, we therefore propose G raph C onvolutional N etwork Ext ension (GCNext) , which extracts node embeddings from a feature-rich item co-occurrence graph via unsupervised learning with Graph Convolutional Networks (GCNs). We then use these pre-trained item embeddings as auxiliary features describing items and their structural dependence. One major advantage and novelty of GCNext is that it can flexibly be plugged into any current sequential recommendation method. Particularly, we (1) use the computed item embeddings to initialize sequential neural network models, and (2) extend (non-neural) nearest–neighborhood methods with pre-trained item graph embeddings to refine the search of candidate sessions for recommendation. Our main contributions are as follows: (1) We present GCNext, a novel method for session-based recommendation based on a item co-occurence graph for sessions. GCNext combines the topological representation power of GCNs with the session representation generated by neural network sequential models without modifying their architecture; (2) GCNext can easily extend nearest-neighbor methods as well as neural network models in a plug-in fashion to further enhance the performance of these models; (3) We perform a large-scale evaluation of graph-based item embeddings and their impact on a diverse set of sequential models. We find that adding GCNext is not only able to boost the performance of current methods, but also shows significant performance improvements over current state-of-the-art sequential models. 2. Related Work In the following, we briefly present related works in the field of graph and node embedding. We subsequently discuss approaches for sequential and session-based recommendation, which incorporate side information or GNNs. 2.1. Graph and Node Embeddings Graph embedding aims to generate low-dimensional vector representations of the graph’s nodes which preserve topology and leverage node features. Non-deep learning methods are mainly based on random walks to explore node neighborhoods ( Perozzi et al., 2014 ; Tang et al., 2015 ; Grover and Leskovec, 2016 ) . With GCNs ( Kipf and Welling, 2017 ; Velickovic et al., 2018 ) , more sophisticated graph embedding methods were introduced: To scale GCNs to large graphs, the layer sampling algorithm ( Hamilton et al., 2017 ) generates embeddings from a fixed node neighborhood. Current state-of-the-art methods in unsupervised learning of representations rely on contrastive methods which base their loss on the difference between positive and negative samples. Deep Graph Infomax (DGI) ( Velickovic et al., 2019 ) contrasts node and graph encodings by maximizing the mutual information between them. Hassani and Khasahmadi ( Hassani and Ahmadi, 2020 ) propose multi-view representation learning by contrasting first-order neighbor encodings with a general graph diffusion. Contrastive learning methods usually require a large number of negative examples and are, therefore, not scalable for large graphs. The approach by Thakoor et al. ( Thakoor et al., 2021 ) learns by predicting substitute augmentations of the input and circumventing the need of contrasting with negative samples. 2.2. Sequential Recommendation Non-neural sequential recommendation approaches focus on the similarity of sessions to extract potential next items. Several works extend the session-based nearest-neighbors method with additional factors such as positions, recency, and popularity ( Ludewig et al., 2021 ; Ludewig et al., 2018 ; Jannach and Ludewig, 2017 ; Garg et al., 2019 ) . Other works ( Tan et al., 2016 ; Li et al., 2017 ; Liu et al., 2018 ; Kang and McAuley, 2018 ) model item-to-item transitions using neural networks, possibly incorporating item features ( Hidasi et al., 2016b ; Zhang et al., 2019 ; de Souza Pereira Moreira et al., 2021 ) . Recent works exploit the graph-based representation of sessions for improved recommendations. Current state-of-the-art use GNNs—in combination with attention or self-attention modules—to capture complex transitions and rich local dependencies ( Wu et al., 2019 ; Xu

Read the complete text