PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Cross-domain Recommendation: Taxonomies, Methods, and Future Directions

Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, Jiadi Yu

arXiv:2108.03357v2cs.IRcs.LG

TL;DR

Traditional recommender systems face data sparsity and cold-start problems, while existing cross-domain surveys provide limited coverage of newer methods, scenarios, and tasks. This paper addresses the gap with a two-level taxonomy and structured synthesis of methods and datasets. It identifies 9 recommendation scenarios and 4 recommendation tasks, reports that deep-learning methods generally outperform traditional machine-learning methods in the reviewed setting, and highlights robustness and scalability as future challenges.

  • Problem

    Traditional recommendation systems are limited by data sparsity and cold-start problems, and existing cross-domain surveys have limited coverage of recent methods, recommendation scenarios, and tasks.

  • Method

    The paper develops a two-level taxonomy, categorizes methods under recommendation scenarios and tasks, organizes datasets, and discusses future research directions.

  • Results

    The survey identifies 9 recommendation scenarios and 4 recommendation tasks; reviewed deep-learning methods generally outperform traditional machine-learning methods in the user-fully-overlap and item-non-overlap scenario.

  • Takeaways & Limitations

    The taxonomy and structured organization provide a framework for comparing cross-domain recommendation research across scenarios, tasks, methods, and datasets.

  • Takeaways & Limitations

    Some approaches are limited by assumptions or requirements such as domain-consistent user preferences, overlapping entities, or available domain tags and external information.

Abstract

from arXiv · show

Traditional recommendation systems are faced with two long-standing obstacles, namely, data sparsity and cold-start problems, which promote the emergence and development of Cross-Domain Recommendation (CDR). The core idea of CDR is to leverage information collected from other domains to alleviate the two problems in one domain. Over the last decade, many efforts have been engaged for cross-domain recommendation. Recently, with the development of deep learning and neural networks, a large number of methods have emerged. However, there is a limited number of systematic surveys on CDR, especially regarding the latest proposed methods as well as the recommendation scenarios and recommendation tasks they address. In this survey paper, we first proposed a two-level taxonomy of cross-domain recommendation which classifies different recommendation scenarios and recommendation tasks. We then introduce and summarize existing cross-domain recommendation approaches under different recommendation scenarios in a structured manner. We also organize datasets commonly used. We conclude this survey by providing several potential research directions about this field.

1 INTRODUCTION

Cross-domain recommendation addresses data sparsity and cold-start problems by transferring information across domains, motivating increasingly varied methods and more refined classifications. This survey responds with a two-level taxonomy, structured method review, dataset organization, and future research directions.

  • Motivation: Data sparsity and cold-start problems limit traditional recommendation systems because interaction records are insufficient or absent for new users.These obstacles impede mining user interests and making recommendations.
  • Motivation: Cross-domain recommendation transfers useful information between domains, but its scenarios and tasks vary with user-item overlap and recommendation targets.Recommended items and users may belong to the same or different domains, with goals involving one or multiple domains.
  • Research gap: Earlier surveys used coarse classifications, including four overlap scenarios and several domain types, leaving room for finer organization of recent research.Prior work classified domains by system, data, or temporal distinctions and scenarios by user-item overlap.
  • Contributions: The survey proposes a two-level taxonomy identifying 9 recommendation scenarios from user-item overlap and 4 recommendation tasks.It compares the proposed scenario classification with previous surveys and organizes tasks separately.
  • Contributions: Existing cross-domain approaches are summarized and categorized by methods, recommendation scenarios, and recommendation tasks.The paper avoids mixing approaches across scenarios because methods proposed for one scenario may not apply to another.
  • Contributions: The survey also introduces commonly used datasets and outlines potential future research directions for cross-domain recommendation.Its organization covers both multi-domain and single-domain datasets.

2 TAXONOMY

The survey proposes a two-level taxonomy that organizes cross-domain recommendation by recommendation scenarios and tasks, then categorizes existing methods within applicable scenarios. It identifies nine scenarios and four tasks while noting uneven research coverage across task types.

  • Taxonomy structure: The taxonomy classifies cross-domain recommendation using two levels: scenarios based on user/item-set relations and tasks based on recommendation structure.The first level concerns recommendation scenarios; the second concerns recommendation tasks.
  • Recommendation scenarios: The scenario taxonomy uses user-set overlap and item-set overlap as its two classification dimensions.User overlap distinguishes non-overlap, partial overlap, and full overlap; the corresponding item dimension is used alongside it.
  • Recommendation scenarios: The survey extends the earlier four-category scenario classification to nine categories.Three scenario combinations currently have no corresponding studies, and symmetric cases can often be handled by exchanging users and items.
  • Recommendation tasks: The task taxonomy distinguishes intra-domain from inter-domain recommendation and single-target from multi-target recommendation.Inter-domain recommendation recommends items from a different domain than the user and is often treated as cold-start user recommendation.
  • Research coverage: In the surveyed literature, intra-domain multi-target recommendation is most widely studied, followed by intra-domain single-target recommendation.Inter-domain multi-task recommendation is least studied, with only two corresponding papers.
  • Method-based categorization: The survey organizes existing approaches by method and recommendation scenario rather than mixing methods across scenarios with different applicability.It reviews three studied non-equivalent scenario groups and excludes a scenario equivalent to single-domain recommendation from further method-based classification.

3 SCENARIO 1: USER NON-OVERLAP & ITEM NON-OVERLAP

When neither users nor items overlap, CDR methods transfer shared structure indirectly through cluster-level rating patterns, tags, or latent factors. The survey organizes this scenario's approaches and notes trade-offs involving transfer granularity and available cross-domain information.

  • This scenario arises when user and item correspondences across domains are unavailable, motivating indirect knowledge transfer.
  • Extracting Cluster-Level Rating Patterns: Cluster-level methods extract shared rating patterns from one domain and transfer them to another through user and item cluster indicators.A compact codebook B encodes cluster-level ratings and supports target-domain prediction.
  • Extracting Cluster-Level Rating Patterns: Variants extend cluster-level transfer to multiple source domains, common and domain-specific patterns, or shared user-item-tag tensors.TALMUD weights source-domain codebooks by relevance, while FUSE factorizes a shared three-dimensional tensor.
  • Capturing Tag Correlations: Other approaches exploit cross-domain tag co-occurrences or item-session relations to connect domains without overlapping users or items.Tag-based methods assume shared tags can encode related preferences or item properties across domains.
  • Discussion: Cluster-level transfer is universal without overlapping entities but coarse-grained, whereas tag-based methods need additional cross-domain information such as tags or comments.Active-learning alternatives require budget and depend on the quantity and quality of identified entity pairs.

4 SCENARIO 2: USER PARTIAL OVERLAP & ITEM NON-OVERLAP

With partial user overlap but no item overlap, methods use shared users to constrain factorization, combine representations, map latent spaces, or propagate information through graphs. Their suitability depends on overlap coverage, non-overlapping-user prevalence, and computational resources.

  • This scenario contains users active in both domains alongside users confined to one domain, and includes five method classes.The classes are collective matrix factorization, representation combination, embedding and mapping, graph neural networks, and aspect-correlation methods.
  • Collective Matrix Factorization: Collective matrix factorization factorizes each domain's ratings while using cross-domain knowledge as constraints within the factorization process.
  • Representation Combination of Overlapping Users: Representation-combination methods merge embeddings of overlapping users before separately training prediction models for each domain.Embedding layers use domain data, combination layers unify overlapping-user representations, and prediction layers train domain-specific recommenders.
  • Embedding and Mapping: Embedding-and-mapping methods learn latent factors in each domain, train a mapping function, and generate target-domain factors for source-only users.EMCDR uses MF or BPR for latent factors and linear or MLP mappings.
  • Discussion: Graph neural network approaches often achieve optimal performance but may have limited scalability because they demand substantial computing resources.Representation-combination methods generally perform well with sufficient overlap but deteriorate when non-overlapping users dominate.

5 SCENARIO 3: USER FULL OVERLAP & ITEM NON-OVERLAP

For fully overlapping users and disjoint items, methods transfer information across domains through six families, spanning shallow factorization models and deep learning approaches.

  • This scenario has interactions from all users in every domain, while items are disjoint across domains.
  • Existing approaches comprise collective matrix factorization, tensor factorization, factorization machines, deep sharing user representations, deep dual knowledge transfer, and deep integration of source-domain information.The first three classes are generally shallow, whereas the last three are deep learning-based.
  • 5.1 Collective Matrix Factorization: Collective matrix factorization jointly factorizes two domains’ rating matrices while constraining their user representations to be shared.CMF uses U_A = U_B; SoRec similarly shares user feature matrices across rating and social-network data.
  • 5.1 Collective Matrix Factorization: Nonlinear mappings and shared-plus-specific latent factors relax exact representation sharing by modeling domain relationships and domain-specific characteristics.The nonlinear mappings enforce f(U_A) = U_B and g(U_B) = U_A, while LSCD uses U_A = U_S + H_A and U_B = U_S + H_B.

5.2 Tensor Factorization

Tensor factorization generalizes matrix factorization to higher-order domain, time, or aspect information, producing predictions from multiple related representations.

  • Tensor factorization extends matrix factorization by incorporating domain, time, and aspect factors alongside user and item factors.Multiplying user representations with other factor representations yields domain-, time-, or aspect-specific representations.
  • The final prediction is derived from the product of multiple related representations, with CP/PARAFAC identified as the most widely used tensor factorization algorithm.
  • CDTF models user-item-domain triadic relations using shared user factors, domain-specific item factors, and a domain-factor matrix.The domain factors express traits of each domain and support recommendations from triadic interactions.
  • RB-JTF jointly factorizes rating information into user, item, and aspect latent factors while transferring knowledge through shared user and aspect factors.
  • Hyper-structure transfer captures nonlinear correlations between domains rather than transferring only linearly correlated knowledge.It constructs a more complex structure before transferring rating-related information.

5.3 Factorization Machines

Factorization Machines model feature combinations in sparse recommendation data and must be extended to incorporate cross-domain interaction patterns.

  • Factorization Machines address feature combination in large-scale sparse data and can be extended from second-order to multi-order models.
  • 5.3.1 The basic paradigm: The second-order FM predicts y from a bias, individual feature terms, and pairwise feature-combination terms.The feature vector dimension is n, and w_ij represents the importance of a feature combination.
  • Cross-domain recommendation requires FM extensions that incorporate user interaction patterns from different domains.
  • FM-MCMC expands target-domain feature vectors with source-domain, user-specific interaction patterns controlled by a domain-dependent transfer function.The expanded vectors are then provided to a general FM model.

5.4 Deep Sharing User Representations

Deep sharing user-representation methods learn user and item embeddings through neural modules while sharing the user-processing component across domains. Later variants incorporate content, graph propagation, attention, and domain-shared or domain-specific preferences.

  • Deep sharing methods transform initial user and item embeddings through separate neural modules, with the user module shared across domains.The modules learn latent representations for users and items in domains A and B.
  • MVDNN generates embeddings from user features and item attributes, using the same user view for two domains and separate item views.
  • CCCFNet and GCBAN combine collaborative-filtering and content-based factors when generating initial user and item embeddings.
  • PPGN models cross-domain interactions with a preference matrix and propagates high-order user preferences through graph convolution and propagation layers.
  • DeepAPF, EATNN, and MSDCR jointly model domain-shared and domain-specific preferences, using attention or fine-grained aspect-level mechanisms.These methods share only domain-shared representations while preserving domain-specific information.

5.5 Deep Dual Knowledge Transfer

Deep Dual Knowledge Transfer methods transfer information bidirectionally between domains through deep model connections, with variants using richer feature representations, preference mappings, sequential encoders, or graph convolution.

  • Deep dual knowledge transfer fuses each domain’s previous-layer output with the other domain’s output as input to the next layer.The approach uses domain-specific and domain-shared parameters within multilayer architectures.
  • CoNet introduces cross-connections between two domain networks using concatenated one-hot embeddings for a shared user and domain-specific items.Domain A receives [x_u, x_i], while domain B receives [x_u, x_j].
  • ACDN extends CoNet by extracting item-image aesthetic features with a pre-trained convolutional network before deep cross-domain transfer.These aesthetic features are combined with user and item one-hot embeddings.
  • DDTCDR improves on CoNet with pre-trained autoencoders, a latent orthogonal mapping for user preferences, and joint within-domain and cross-domain preference modeling.
  • Later variants apply the framework to sequential or graph-based recommendation using separate RNNs, shared-account transfer units, or graph convolutional networks.π-Net encodes behavior sequences, while BiTGCF propagates user and item features over domain-specific bipartite graphs.

5.6 Deep Integration of Source Domain Information

Deep Integration of Source Domain Information treats one domain as the target and incorporates other domains as auxiliary information to improve target-domain recommendation.

  • These approaches use asymmetric models for single-target recommendation, with source-domain information incorporated into a dominant target domain.The source-domain information is denoted R^A and serves as additional auxiliary information.
  • MTNet combines target-domain review representations, transferred source-domain knowledge, and user-item interaction representations for final recommendation.A memory network generates context-text representations, while a transfer network represents source knowledge.
  • NATR shares item rather than user embeddings to reduce user-privacy leakage during cross-domain transfer.Attention units aggregate embeddings from overlapping source-domain items and fuse them with the target-domain user embedding.

5.7 Discussion

The discussion contrasts six method classes, finding that deep learning methods generally outperform traditional shallow methods while identifying limitations tied to domain similarity, noise transfer, and target-domain asymmetry.

  • Deep learning methods generally outperform traditional machine learning methods because deep networks better model complex nonlinear user-item interactions.The comparison covers six classes: three traditional methods and three deep learning-based methods.
  • Deep Sharing User Representations can ignore domain differences, while Deep Dual Knowledge Transfer may introduce irrelevant factors as noise through indiscriminate transfer.The former relies on shared representations, whereas the latter transfers knowledge bidirectionally between domains.
  • Deep sharing and dual-transfer approaches generally perform well when domains are highly similar and their data sparsity is relatively close.
  • Deep Integration of Source Domain Information focuses on the target domain, does not fully exploit complementary cross-domain knowledge, and requires retraining to recommend in a source domain.Its customized target-domain structure and parameters support target performance but limit direct source-domain recommendation.

6 DATASETS FOR CROSS-DOMAIN RECOMMENDATION

The survey organizes cross-domain recommendation datasets by domain coverage and application area, emphasizing multi-domain resources and commonly used movie, book, music, social, POI, and mobile datasets.

  • Datasets with Multiple Domains: Amazon, Douban, and Epinions are representative multi-domain datasets widely used because shared user identities support cross-domain task construction.Amazon spans 24 item domains and includes explicit ratings plus review text.
  • Datasets with Multiple Domains: Pairs of domains can form tasks with overlapping users and non-overlapping items, while domains across datasets can support non-overlapping-user scenarios.The survey gives Amazon, Douban, and cross-dataset examples of these task constructions.
  • Single-Domain Datasets: MovieLens, Netflix, and EachMovie are widely used movie datasets, with shared movies enabling both non-overlapping-item and overlapping-item recommendation scenarios.MovieLens provides benchmark variants including MovieLens-100K, MovieLens-1M, and MovieLens-20M.
  • Single-Domain Datasets: Book-Crossing and LibraryThing support book recommendation, including scenarios matching books across datasets through unique ISBN identifiers.Book-Crossing contains 278,858 users, 1,149,780 ratings, and 271,379 books.
  • Single-Domain Datasets: Yahoo! Music and Last.FM represent music datasets, while Twitter, YouTube, Weibo, and DBLP commonly provide auxiliary social interaction information.
  • Single-Domain Datasets: Yelp, Brightkite, and Foursquare provide point-of-interest behavior data, and Cheetah Mobile naturally supports overlapping-user, non-overlapping-item tasks across app installation and news browsing.POI resources include ratings, reviews, check-ins, timestamps, locations, and venue or friendship information.

7 FUTURE DIRECTIONS AND CHALLENGES

The survey identifies open challenges for cross-domain recommendation involving unexplored overlap scenarios, knowledge transfer, robustness, scalability, and explainability. It also points to traditional recommendation methods and related technologies as sources for future model development.

  • Future Research Opportunities: Three of nine cross-domain recommendation scenarios involving overlapping users and items have not yet been studied.These are user partial overlap with item partial overlap, user partial overlap with item full overlap, and user full overlap with item partial overlap.
  • Methodological Opportunities: Traditional recommendation methods and technologies from related fields can be directly referred to when developing cross-domain recommendation methods.The survey specifically frames knowledge extraction within each domain as an area where existing approaches can be reused.
  • Robustness: Robust cross-domain models must transfer useful knowledge while avoiding negative transfer and handling the greater noise sensitivity of complex architectures.The survey describes adversarial perturbations and methods targeting negative transfer as existing responses to these challenges.
  • Scalability: Scalable cross-domain recommendation must operate effectively and timely on large datasets despite complex multi-domain model structures.The challenge arises from both the complexity of cross-domain models and the need to integrate information from multiple domains.
  • Explainability: Explainable cross-domain recommendation must accommodate domain differences in metadata and account for both domain-specific and domain-invariant representations.The survey connects these challenges to the readability, transparency, persuasiveness, and trustworthiness of recommendations.

8 CONCLUSION

The survey systematically organizes cross-domain recommendation research through taxonomies of scenarios and tasks, method summaries, dataset coverage, and proposed future directions.

  • Conclusion: The survey presents a two-level taxonomy, organizes methods by recommendation scenario, introduces commonly used datasets, and discusses future research directions.Its stated aim is to clarify the field for newcomers and experts.
Loading 2108.03357v2…