Source-linked AI summary
A Survey on Accuracy-oriented Neural Recommendation: From Collaborative Filtering to Information-rich Recommendation
Le Wu, Xiangnan He, Xiang Wang, Kun Zhang, Meng Wang
TL;DR
Neural recommender research has expanded rapidly, but the field needs a systematic accuracy-oriented synthesis. This survey organizes neural recommendation models by data usage, reviews representative work, and discusses promising directions and reproducibility concerns.
Problem
The rapid growth of neural recommender models creates a need for a systematic accuracy-oriented review of this developing field.
Method
The survey categorizes neural recommender models into collaborative filtering, content enriched recommendation, and temporal/sequential recommendation based on data usage.
Results
The survey reviews influential work in each category and identifies directions spanning foundations, modeling, evaluation, and reproducibility.
Takeaways & Limitations
Neural recommendation remains an active area with opportunities for further research in graph reasoning, self-supervised learning, evaluation, and reproducibility.
Takeaways & Limitations
Published improvements in neural recommendation raise reproducibility concerns because models are difficult to tune and results vary with datasets and experimental settings.
Abstract
from arXiv · showhide
Influenced by the great success of deep learning in computer vision and language understanding, research in recommendation has shifted to inventing new recommender models based on neural networks. In recent years, we have witnessed significant progress in developing neural recommender models, which generalize and surpass traditional recommender models owing to the strong representation power of neural networks. In this survey paper, we conduct a systematic review on neural recommender models from the perspective of recommendation modeling with the accuracy goal, aiming to summarize this field to facilitate researchers and practitioners working on recommender systems. Specifically, based on the data usage during recommendation modeling, we divide the work into collaborative filtering and information-rich recommendation: 1) collaborative filtering, which leverages the key source of user-item interaction data; 2) content enriched recommendation, which additionally utilizes the side information associated with users and items, like user profile and item knowledge graph; and 3) temporal/sequential recommendation, which accounts for the contextual information associated with an interaction, such as time, location, and the past interactions. After reviewing representative work for each type, we finally discuss some promising directions in this field.
1 INTRODUCTION
This survey reviews neural recommender models from an accuracy-oriented recommendation-modeling perspective, organizing work by the data used: collaborative interactions, side information, and contextual information. It contrasts this organization with surveys centered on specific topics or deep-learning techniques.
- Neural recommender models use neural networks to review and advance recommendation modeling, motivated by their expressiveness for large data with complicated patterns.
- The survey is organized around recommendation modeling with an accuracy goal rather than solely around individual deep-learning architectures.
- The survey reviews existing methods, discusses future directions and open issues, and frames recommendation as learning a function that predicts whether a user favors an item under context.
- Collaborative filtering: Collaborative filtering forms the basis of personalized recommendation and uses user or item IDs and interaction histories while ignoring context data.
- Content-enriched recommendation: Content-enriched models extend collaborative filtering by integrating user and item side information, including profiles, social networks, attributes, and knowledge graphs.
- Temporal/sequential recommendation: Context-aware models incorporate interaction-associated information such as time, location, and past interaction sequences alongside user and item data.
2 COLLABORATIVE FILTERING MODELS
Collaborative filtering models learn user and item representations from interaction data, addressing sparse behavior through history aggregation, autoencoders, and graph learning. They then estimate user-item preference with inner-product, distance-based, or neural interaction functions.
- Collaborative filtering predicts target-user behavior from the collective behaviors of users, evolving from memory-based similarity methods to matrix factorization and neural models.
- Representation Learning: CF representation learning seeks user and item embeddings from the interaction matrix, but sparse user behavior makes accurate embedding learning difficult.
- History Behavior Aggregation: History-based models improve user representations by pooling interacted items, while attention models assign item-specific importance and can condition contributions on the target item.
- Autoencoder-based Models: Autoencoder models reconstruct incomplete user-item records through encoder-decoder networks to learn latent user representations and predicted preferences.
- Graph-based Representation Learning: Graph models propagate embeddings over the user-item bipartite graph to encode higher-order connectivity, with simplified variants showing superior practical performance without carefully chosen activations.
- Interaction Modeling: Interaction modeling estimates preference from user and item embeddings using inner products, distances, or neural networks; inner products remain efficient but can miss complex relationships.
3 CONTENT-ENRICHED RECOMMENDATION
Content-enriched recommendation extends collaborative filtering by incorporating user, item, and contextual auxiliary information. Neural models learn richer feature, textual, social, and knowledge-graph representations, while integration choices and practical scalability remain important.
- Overview: Auxiliary content and context enrich collaborative-filtering representations beyond user-item interactions, including user or item features, text, and social information.These sources target semantic relatedness and additional behavioral or contextual signals.
- Modeling General Feature Interactions: Feature-enhanced collaborative filtering progresses from Factorization Machines to neural models that implicitly or explicitly learn higher-order interactions.The surveyed neural approaches include MLP, cross-network, and tree-enhanced modeling.
- Modeling General Feature Interactions: Tree-enhanced modeling combines decision-tree-derived cross features with attention, providing strong representation ability and explainability.TEM uses trees to extract high-order feature interactions before attentive prediction.
- Modeling Textual Content: Neural language models and autoencoders learn semantic item or user representations from textual content, including reviews and raw item features.Autoencoder methods jointly optimize rating prediction and content reconstruction, while word-embedding methods integrate CNNs with matrix factorization.
- Modeling Textual Content: Review-based models exploit user and item review histories, while generation-based models can jointly predict ratings and generate review explanations.The target user-item review is reported as especially predictive, but is unavailable at test time.
- Modeling Social Network and Knowledge Graph: Social networks and knowledge graphs supply relational signals: heterogeneous GNNs jointly model social and interest diffusion, while KG methods propagate neighbor information.Auxiliary data can improve representation learning and recommendation performance, but large knowledge graphs may require costly path construction and transfer.
4 TEMPORAL/SEQUENTIAL MODELS
Temporal and sequential recommendation models account for evolving preferences, item-transition patterns, or both. The survey organizes them into temporal, session-based, and combined temporal-session approaches using recurrent, memory, attention, translation, and graph-based techniques.
- Temporal based recommendation: Temporal recommendation models interaction timestamps to capture users’ evolving preferences over time.Many approaches use RNNs because they are effective at modeling temporal patterns.
- Session based recommendation: Session-based recommendation targets anonymous users by mining sequential item-transition patterns from short sessions.GRU4REC applies an RNN-based framework to session records without user IDs.
- Temporal and session based recommendation: Combined temporal-session models capture both users’ temporal evolution and sequential item patterns in transactions represented by user, session, and time.These methods model either long- and short-term preferences separately or a unified user representation.
- Temporal based recommendation: RRN learns user and item latent-vector evolution with two RNNs and can incorporate social, metadata, and multimedia context.Its dynamic embeddings are updated through LSTM-based autoregressive architectures and contextual embeddings.
- Temporal based recommendation: Memory networks store historical states in memory slots and update them through read and write operations to represent longer-term interests.Attention with soft addressing is used to capture stable long-term and short-term temporal interests.
- Session based recommendation: Session models also use translation, self-attention, and GNNs to represent item correlations and transitions beyond consecutive interactions.SR-GNN constructs item-transition edges from consecutive session orders and learns item embeddings with a GNN.
5 DISCUSSION AND FUTURE DIRECTIONS
The survey identifies unresolved challenges in benchmarking, model development, evaluation, and reproducibility. It points toward graph reasoning, self-supervised learning, multi-objective recommendation, and stronger experimental standards as future directions.
- Discussion and future directions: Although neural recommendation models have achieved superior quality, current solutions remain far from satisfactory and leave substantial research opportunities.The survey frames future work from basis, modeling, evaluation, and reproducibility perspectives.
- Basis: Recommendation Benchmarking: Recommendation benchmarking remains difficult because scenarios, datasets, inputs, and modeling assumptions vary across systems.The survey calls for large, continually updated datasets that track state-of-the-art performance across recommendation problems.
- Models: Graph Reasoning & Self-supervised Learning: Graph reasoning and self-supervised learning are promising directions for exploiting recommendation structure and auxiliary supervision from limited interactions.Recommendation data can be represented as user-item, attributed, heterogeneous, or knowledge graphs.
- Evaluation: Multi-Objective Goals for Social Good Recommendation: Accuracy-focused evaluation does not capture goals such as diversity, explanations, stakeholder balance, fairness, and social good.The survey notes that accuracy-oriented data-driven decisions may introduce algorithmic bias and disadvantage long-tailed items.
- Discussion: Reproducibility: Reproducibility is challenged by difficult neural-network tuning and sensitivity to datasets, preprocessing, sampling, objectives, optimization, and metrics.These factors make fair performance comparisons difficult across recommendation experiments.
6 CONCLUSION
The survey systematically organizes neural recommender research by data usage and discusses promising directions, including reproducibility. It aims to help readers quickly understand recommendation modeling’s development and key aspects.
- The survey categorizes neural recommender models into collaborative filtering, content enriched, and temporal/sequential models based on data usage.
- It summarizes influential research in each category and concludes the corresponding main contributions and opinions.
- The survey discusses promising directions spanning basics, modeling, evaluation, and reproducibility in recommender systems.
- The survey aims to help readers quickly understand recommendation modeling’s development and key aspects and inspire future studies.