PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning based Recommender System: A Survey and New Perspectives

Shuai Zhang, Lina Yao, Aixin Sun, Yi Tay

arXiv:1707.07435v7cs.IR

TL;DR

Recommender systems help manage information overload, but deep learning recommendation lacked a comprehensive, organized review of its growing research. This paper surveys the field through a taxonomy and state-of-the-art synthesis, finding broad academic and industrial adoption while identifying limitations and open problems.

  • Problem

    Deep learning recommendation had few systematic reviews that comprehensively organized existing work, summarized current efforts, and detailed open problems.

  • Method

    The paper conducts an extensive review, proposes a classification scheme, and synthesizes notable deep learning recommender models, challenges, and future directions.

  • Results

    Deep learning has become pervasive in recommender systems research and industry, with cited applications showing improvement over traditional models in online testing.

  • Takeaways & Limitations

    The survey provides researchers, practitioners, and educators with an organized overview for understanding and applying deep learning recommendation models.

  • Takeaways & Limitations

    Deep learning recommendation remains constrained by interpretability challenges and the need for sufficient data to support rich parameterization.

Abstract

from arXiv · show

With the ever-growing volume of online information, recommender systems have been an effective strategy to overcome such information overload. The utility of recommender systems cannot be overstated, given its widespread adoption in many web applications, along with its potential impact to ameliorate many problems related to over-choice. In recent years, deep learning has garnered considerable interest in many research fields such as computer vision and natural language processing, owing not only to stellar performance but also the attractive property of learning feature representations from scratch. The influence of deep learning is also pervasive, recently demonstrating its effectiveness when applied to information retrieval and recommender systems research. Evidently, the field of deep learning in recommender system is flourishing. This article aims to provide a comprehensive review of recent research efforts on deep learning based recommender systems. More concretely, we provide and devise a taxonomy of deep learning based recommendation models, along with providing a comprehensive summary of the state-of-the-art. Finally, we expand on current trends and provide new perspectives pertaining to this new exciting development of the field.

1 INTRODUCTION

Recommender systems address information overload and over-choice, while deep learning has rapidly expanded recommendation research and practice. This survey organizes the field, reviews its state of the art, and identifies open problems and future directions.

  • Motivation: Recommender systems personalize access to products, media, and other information amid expanding online choice.They support user experience, business, and decision-making across e-commerce and media websites.
  • Motivation: Deep learning has rapidly influenced recommender architectures because of its success across application domains and its potential to improve recommendation performance.The paper describes growing research activity and regular workshops dedicated to deep learning for recommender systems.
  • Industrial adoption: Deep learning recommender systems are pervasive in industry, with 80 percent of Netflix movies watched and 60 percent of YouTube video clicks attributed to recommendations.The cited industrial systems reportedly passed online testing and improved over traditional models.
  • Research gap: Few systematic reviews had comprehensively organized deep learning recommendation research, summarized current efforts, and detailed open problems.Earlier related surveys were short or covered limited numbers of models and papers.
  • Survey scope: The survey collects over a hundred papers using Google Scholar, Web of Science, major conferences, and keywords spanning recommendation and deep learning.Its contributions include a classification scheme, a state-of-the-art overview, and discussion of challenges and future directions.

2 OVERVIEW OF RECOMMENDER SYSTEMS AND DEEP LEARNING

Recommender systems estimate user preferences from interactions, content, and auxiliary information, while deep learning learns representations and nonlinear functions for recommendation. The overview presents neural architectures, their compositional advantages, and recognized limitations including interpretability and data requirements.

  • Recommender systems: Recommender systems estimate user preferences and recommend items using collaborative filtering, content-based, or hybrid strategies.Inputs can include user-item interactions, item content, and auxiliary information such as text, images, videos, temporal data, and spatial data.
  • Deep learning: Deep learning learns multiple levels of representations and abstractions through differentiable neural architectures optimized with stochastic gradient descent.The overview includes architectures such as multilayer perceptrons, autoencoders, recurrent networks, and deep reinforcement learning.
  • Why deep learning: For interaction-only recommendation, deep neural networks are primarily justified by high task complexity or a large number of training instances.The survey notes that traditional methods such as BPR, MF, and CML can also perform reasonably well on interaction-only data.
  • Why deep learning: Neural networks support modular, composite recommendation models that combine different structures and train end-to-end on multimodal content.This is especially relevant when recommendation uses textual, image, or other web data.
  • Potential limitations: Deep learning recommendation faces limitations in interpretability and data requirements, although attention models can improve interpretability and recommendation often has abundant data.Hidden weights and activations remain difficult to interpret, while rich parameterization requires sufficient data.

3 DEEP LEARNING BASED RECOMMENDATION: STATE-OF-THE-ART

The survey organizes deep learning recommendation models by neural techniques and reviews representative architectures, application tasks, and state-of-the-art developments. It covers neural building blocks, hybrid models, and extensions of traditional recommendation methods.

  • Model taxonomy: The survey classifies deep learning recommendation models by employed techniques and divides them into neural building blocks and deep hybrid models.The first category includes MLP, autoencoder, CNN, RNN, RBM, NADE, attention, adversarial, and deep reinforcement learning recommenders; the second combines multiple neural techniques.
  • MLP-based recommendation: MLPs extend recommendation models by learning nonlinear transformations and modeling user-item interactions from side information or identifiers.Neural collaborative filtering combines matrix-factorization linearity with MLP nonlinearity, while traditional matrix factorization is a special case of NCF.
  • MLP-based recommendation: DeepFM integrates factorization machines and MLPs to model low-order and high-order feature interactions in one end-to-end model.The factorization-machine component captures linear and pairwise interactions, while the MLP models high-order interactions; subsequent work jointly models explicit and implicit interactions.
  • MLP-based recommendation: Wide & Deep Learning combines memorization of direct historical features with generalization through abstract multilayer representations.The model uses a single-layer wide component and a multilayer perceptron, and its final recommendation list is generated from predicted scores.
  • MLP-based recommendation: Wide & Deep Learning requires manually selecting features for its wide and deep components and manually designing cross-product transformations.These preprocessing choices can substantially influence model utility, while deep factorization models can reduce the feature-engineering burden.

3.3 Autoencoder based Recommendation

Autoencoder-based recommendation models reconstruct interaction data or learn lower-dimensional representations, with variants extending collaborative filtering through denoising, side information, and deep probabilistic integration. The surveyed models target rating prediction and ranking, while later extensions improve robustness, scalability, or performance.

  • General framework: Autoencoders support recommendation by learning bottleneck representations or directly reconstructing missing interaction-matrix entries.These approaches include denoising, variational, contractive, and marginalized autoencoders.
  • AutoRec: AutoRec reconstructs user- or item-level partial vectors, with item-based AutoRec generally outperforming its user-based variant.Performance also depends on activation functions and hidden-layer size; adding layers can yield slight improvement.
  • Collaborative filtering extensions: CFN extends AutoRec with denoising and side information, injecting user profiles or item descriptions to address sparsity and cold-start effects.Its corruption options include Gaussian, masking, and salt-and-pepper noise.
  • Collaborative denoising autoencoder: CDAE corrupts implicit-feedback preference vectors with Gaussian noise and learns user-specific reconstruction parameters for ranking prediction.Negative sampling reduces training time without degrading ranking quality according to the surveyed authors.
  • Variational autoencoders: Multi-VAE and Multi-DAE report better performance than CDAE on implicit recommendation data, using Bayesian parameter estimation.The surveyed work describes the Bayesian approach as favorable compared with commonly used likelihood functions.
  • Deep collaborative models: CDL integrates a stacked denoising autoencoder with probabilistic matrix factorization, whereas CDR uses a pairwise top-n ranking framework and outperforms CDL on ranking prediction.mDA-CF marginalizes corrupted inputs for greater scalability than CDL and incorporates both item and user content information.

3.4 Convolutional Neural Networks based Recommendation

CNN-based recommender systems use convolutional architectures mainly to extract features from multimedia or behavioral data, and they also model collaborative interactions and structured recommendation inputs. Applications span audio, video, images, text, context, and graph-based data.

  • Overview: CNN-based recommendation models primarily use convolution and pooling operations for feature extraction from unstructured multimedia data.The extracted representations can support content-based, collaborative, or hybrid recommendation models.
  • Image and context features: CNNs extract image features for POI recommendation by modeling interactions between visual content and latent user or location factors.Context-aware tag recommendation combines CNN image features with neural context representations before softmax prediction.
  • Text features: DeepCoNN uses parallel CNNs over review text to model user behavior and item properties, addressing sparsity through semantic representations and preserving word order.The resulting user and item representations are combined in a factorization-machine prediction layer.
  • Audio and video: Audio CNNs extract music-signal features across multiple timescales and can alleviate cold start when music has not yet been consumed.Other work uses ResNet to extract audio features within collaborative metric learning.
  • Collaborative filtering: ConvNCF replaces the user-item dot product with an outer product and applies CNNs to capture higher-order correlations among embedding dimensions.CNNs are also applied to sequential recommendation with user identifiers.
  • Graph recommendation: Graph CNN recommenders treat user-item interactions as structured bipartite-graph data and frame recommendation as link prediction.This connects recommendation with graph-based modeling for non-Euclidean data.

3.5 Recurrent Neural Networks based Recommendation

RNN-based recommender systems model temporal interaction patterns, especially when users lack persistent identifiers, and can incorporate sequential side information. The surveyed models span session recommendation, dynamic user-item states, and hybrid tasks, while evaluation evidence highlights the importance of strong neighbourhood baselines.

  • RNNs suit recommendation settings involving temporal dynamics, sequential user behaviour, and sequential side information such as text or audio.
  • Session-based Recommendation without User Identifier: Session-based recommendation addresses settings without user identifiers by learning short-term preferences from recent interaction sequences.
  • GRU4Rec: GRU4Rec encodes the current session with 1-of-N item vectors and predicts the likelihood that each item will be next.
  • Session models incorporating identity, image, text, or context information outperform models based solely on historical interactions in the reported experiments.
  • Simple neighbourhood methods can match GRU4Rec accuracy, and combining neighbourhood methods with RNNs can yield the best performance.
  • Sequential Recommendation with User Identifier: Recurrent recommender networks use LSTMs for dynamic user and item states while combining them with stationary latent attributes learned through matrix factorization.

3.6 Restricted Boltzmann Machine based Recommendation

RBM-based recommendation represents ratings through neural-network visible units and learns shared hidden representations for collaborative filtering. The survey describes extensions for implicit feedback, item-based and hybrid formulations, while noting that RBM training is computationally intractable directly.

  • RBM-based collaborative filtering represents each rating with a one-hot visible vector because RBM visible units are restricted to binary values.
  • Each user has an RBM with shared parameters, and hidden units model the structure of the user’s rated-item matrix.
  • RBM parameters are learned with Contrastive Divergence because the model is not tractable.
  • Conditional RBMs extend the approach to incorporate implicit feedback.
  • RBM-CF can be formulated as user-based, item-based, unified user-item, or hybrid models incorporating item features.

3.7 Neural Atention based Recommendation

Neural attention mechanisms select informative portions of recommendation inputs and are organized by how attention scores are learned. The survey covers vanilla attention and co-attention across latent-factor, recurrent, convolutional, and sequential recommendation models.

  • Attention mechanisms filter uninformative features and reduce the effects of noisy input data in recommender systems.
  • The survey classifies neural attention models into vanilla attention, which uses a parameterized context vector, and co-attention, which learns weights from two sequences.
  • Recommendation with Vanilla Attention: An attentive collaborative filtering model uses item-level attention to select representative items and component-level attention to select informative multimedia features.
  • Attention-based LSTMs combine sequential modelling with recognition of informative words in microblog posts for hashtag recommendation.
  • Attention-based CNN recommendation uses separate global and local attention channels for multi-label hashtag classification.
  • Recommendation with Co-Attention: Self-attention can learn short-term user intents from recent interactions while metric learning produces expressive user and item embeddings.

3.8 Neural AutoRegressive based Recommendation

Neural autoregressive recommendation models estimate users’ rating distributions by decomposing joint probabilities into ordered conditional probabilities. CF-NADE provides a tractable alternative to RBM-CF and can be extended with implicit feedback and user-item co-autoregressive modelling.

  • CF-NADE models the joint probability of a user’s rating vector with the chain rule over conditional rating probabilities.
  • CF-NADE offers a tractable alternative to RBM-CF, whose likelihood-gradient approximation relies on Contrastive Divergence.
  • The autoregressive procedure predicts each rating conditioned on the ratings preceding it in a chosen item ordering.
  • Although timestamps are an ideal ordering, random ordering also achieves good empirical performance.
  • Extensions incorporate implicit feedback and user-item co-autoregressive modelling for rating estimation and personalized ranking.

3.9 Deep Reinforcement Learning for Recommendation

Deep reinforcement learning (DRL) addresses recommendation as a sequential, personalized decision problem by incorporating feedback and additional information into real-time strategies.

  • DRL has attracted attention for personalized recommendation because conventional models often treat recommendation as static and miss temporal intentions.The survey identifies sequential interaction settings as a key application context.
  • DEERS uses both negative and positive feedback in a sequential interaction setting for recommendation.
  • Deep neural networks make reinforcement learning practical for modeling extra information in real-time recommendation strategies.

3.10 Adversarial Network based Recommendation

Adversarial recommendation models combine generative and discriminative retrieval through minimax training, with applications including item recommendation and ranking.

  • IRGAN applies generative adversarial networks to information retrieval tasks including web search, item recommendation, and question answering.
  • IRGAN combines generative retrieval and discriminative retrieval in a unified minimax game between generator and discriminator.The generator produces relevant documents to fool the discriminative retrieval model.
  • The generative model approximates the user relevance distribution, while the discriminator distinguishes relevant from non-relevant documents.
  • Generator and discriminator parameters are learned alternately with gradient descent.
  • For ranking tasks, the pointwise objective can be adapted to a pairwise paradigm to generate higher-quality ranking lists.
  • In item recommendation, matrix factorization formulates the model score, but factorization machines or neural networks can replace it.
  • Adversarial personalized ranking extends Bayesian personalized ranking with an adversary that adds noise or permutations to maximize BPR loss.

3.11 Deep Hybrid Models for Recommendation

Deep hybrid recommender models combine neural building blocks to process heterogeneous content, sequential information, and interaction signals for task-specific recommendation.

  • Hybrid models integrate neural building blocks such as CNNs, RNNs, autoencoders, and deep reinforcement learning for specific recommendation tasks.The survey emphasizes that combinations should be designed carefully for the target application.
  • CNNs and Autoencoder: CKE combines CNNs and an autoencoder to learn structural, textual, and visual item representations.It extends collaborative deep learning beyond item text information.
  • CNNs and RNNs: CNNs extract local or visual features while RNNs or LSTMs process sequential text in quote, hashtag, citation, and video recommendation.These models support multimodal or sequential recommendation inputs.
  • RNNs and Autoencoder: CRAE integrates robust recurrent networks with denoising autoencoders to model sequential item content information.Wildcard denoising and beta-pooling are designed to reduce overfitting.
  • RNNs with DRL: A supervised deep reinforcement learning framework combined with RNNs can learn prescription policies from indicator and evaluation signals.The reported system inferred and discovered optimal treatments automatically.

4 FUTURE RESEARCH DIRECTIONS AND OPEN ISSUES

The survey identifies open issues spanning multimodal representation, explainability, model depth, reasoning, cross-domain transfer, multitask learning, and evaluation practices.

  • Joint Representation Learning from User and Item Content Information: Existing models use profiles, implicit feedback, context, reviews, and other side information, but rarely integrate these sources comprehensively.The survey also notes limited use of user footprints from social media and the physical world.
  • Joint Representation Learning from User and Item Content Information: Feature engineering remains insufficiently studied despite its importance in industrial recommendation, where manual feature construction is time-consuming.Deep learning is presented as a promising route toward automatic feature crafting.
  • Joint Representation Learning from User and Item Content Information: Future architectures may learn joint multimodal user-item representations from text, images, and interactions in an end-to-end fashion.
  • Explainable Recommendation with Deep Learning: Attention models can improve interpretability by exposing weights that provide insight into model operation and support explainable results.Interpretability depends on the application domain and the content modality used.
  • Explainable Recommendation with Deep Learning: Future attention mechanisms may progress toward conversational or generative explanations based on evidence contributing to decisions.
  • Going Deeper for Recommendation: Most neural collaborative-filtering models plateau at three to four layers, leaving the value and training of deeper architectures unclear.Suggested directions include auxiliary losses, layer-wise learning rates, and residual strategies.
  • Machine Reasoning for Recommendation: Machine reasoning could support multi-step inference over social profiles, purchases, and multiple modalities for recommendation.The survey identifies reasoning architectures as a possible next frontier.
  • Cross Domain Recommendation with Deep Neural Networks: Cross-domain recommendation can use transferred knowledge to address sparsity and cold-start problems across diversified products and services.The survey describes deep learning as suited to learning abstractions across domains.

5 CONCLUSION

The survey organizes and reviews deep learning–based recommender-system research, evaluates the field’s advantages and disadvantages, and identifies open problems and future extensions.

  • The article extensively reviews notable research on deep learning–based recommender systems.
  • It proposes a classification scheme to organize and cluster existing publications while highlighting influential research prototypes.
  • The survey discusses both advantages and disadvantages of applying deep learning to recommendation tasks.
  • It identifies pressing open problems and promising future extensions for this ongoing research area.The authors note that new techniques and emerging models continue to appear each year.
Loading 1707.07435v7…