Contrastive Self-supervised Learning in Recommender Systems: A Survey

1 INTRODUCTION

• Sparse user–item interactions leave deep learning-based recommenders vulnerable to data sparsity, overfitting, and generalization error.These methods are usually trained with abundant labeled interaction data, despite the interaction space being sparse.Source: §1 · p. 2
• Self-supervised learning derives transferable knowledge from data through pretext tasks, without manually annotated labels, for use in downstream tasks.Source: §1 · p. 2
• Contrastive self-supervised learning is lightweight, encoder-independent, and typically needs no domain knowledge; it has become a dominant approach in self-supervised recommendation.The survey reports that related publications exceed 50% of SSL-based recommendation publications in the ACM Digital Library.Source: §1 · p. 2
• Earlier contrastive-learning reviews focus mainly on CV and NLP, while graph-learning reviews offer limited recommendation coverage and lack in-depth discussion.Recommendation differs through sparse inputs, user/item homophily, and task types such as bundle and multi-behavior recommendation.Source: §1 · p. 2
• Compared with, this survey offers a finer taxonomy, in-depth component analysis, and newer coverage of contrastive recommendation studies.It adds model-based augmentation and methods without augmentation to view-generation strategies.Source: §1 · pp. 2–3
• The survey collects studies through Google Scholar, searches in major conferences and journals, and reference checking to reduce omissions.Search terms combine “self-supervised” or “contrastive” with “recommend” or “collaborative filtering.”Source: §1 · p. 3

• The section introduces background knowledge and relevant concepts for contrastive-learning-based recommendation.Source: §2 · p. 3
• It briefly introduces contrastive learning and recommendation training strategies.Source: §2 · p. 3
• The survey also summarizes its notation in Table 1.Source: §2 · p. 3

2.1 Term Definitions

• Self-supervised learning needs no manual labels and generates supervision signals from data itself, while models are still trained in supervised settings.Source: §2.1.1 · p. 3

2.1.2 Pretext Tasks Versus Downstream Tasks.

• Pretext tasks are designed for models to solve and learn generalized representations; downstream tasks evaluate representation quality, including recommendation tasks such as sequential and social recommendation.Downstream recommendation tasks generally require manually annotated labels.Source: §2.1.2 · p. 3
• Table 1 collects the survey’s key notations.Source: Table 1 · p. 4

• Contrastive learning generates two views, encodes each with a usually shared encoder, then applies a contrastive loss to maximize agreement for positive pairs and minimize it for negative pairs.Positive pairs are the same instances across views; negative pairs are different instances across views.Source: §2.2 · pp. 4–5
• The pipeline’s encoder learns representations from the generated views, and its pretext decoder estimates agreement between instances.Source: §2.2 · p. 5

2.3 Training Strategy

• The two strategies used are pre-training and fine-tuning (P&F) and joint learning (JL).The survey presents their workflows in Fig. 2.Source: §2.3 · p. 5

2.3.1 Pre-training and Fine-tuning (P&F).

• P&F first pre-trains the encoder with contrastive tasks, then fine-tunes it with a recommendation decoder supervised by the downstream task.The pre-trained encoder parameters initialize the fine-tuning stage.Source: §2.3.1 · p. 5

2.3.2 Joint Learning (JL).

• JL jointly trains a usually shared encoder on contrastive and recommendation tasks, using the contrastive task as an auxiliary regularizer.Its objective combines contrastive and recommendation losses, with a trade-off hyperparameter controlling the contrastive loss contribution.Source: §2.3.2 · pp. 5–6

2.3.3 Discussion.

• P&F generalizes better across recommendation tasks and can use other-task data when target data are limited, but its two-stage process is more complex.Because pre-training uses unlabeled data, it may not explicitly learn task-specific features and can compromise performance on some tasks.Source: §2.3.3 · p. 6
• For a specific task, JL usually performs better and is easier to implement; for good performance across different tasks, the survey favors P&F.JL requires simultaneous training of multiple tasks and careful loss balancing.Source: §2.3.3 · p. 6

3.1 Unified Framework

• The framework generates multiple views of the data, then uses a contrastive pretext task to maximize agreement between positive pairs across views.The views can contain representations at different scales, such as an item or a sequence of items.Source: §3.1 · p. 6

3.2 Proposed Taxonomy

• The taxonomy distinguishes methods by view-generation strategy, pretext task, and contrastive objective, while leaving the task-determined encoder outside its scope.The survey presents representative methods in Table 2.Source: §3.2 · pp. 7–8
• View generation is classified by whether augmentation is used, pretext tasks by same- versus cross-scale contrasting, and objectives by whether they provide a mutual-information lower-bound estimate.These dimensions describe how views are generated, how supervision signals are obtained, and how mutual information is measured.Source: §3.2 · p. 9
• Component choices should fit the input data and downstream task; the survey recommends designing the pretext task first, then selecting compatible views and objectives.For example, feature-based augmentation may not suit a pretext task modeling sequential item relationships.Source: §3.2 · p. 9

4 VIEW GENERATION

• View generation uses augmentation when multiple views are unavailable and can use naturally existing views directly; augmentation may be data-based or model-based.Data-based methods transform input data, while model-based methods perturb the encoder.Source: §4 · p. 9, §4.1 · p. 10

4.1.1 Data-based Augmentation.

• Data-based augmentation covers graph structure, interaction sequences, and feature vectors, with methods including graph perturbation, subgraph sampling, sequence completion, and feature transformation.Graph-based methods transform adjacency structure, while subgraph sampling selects nodes and corresponding edges; sequence-based methods transform interaction sequences.Source: §4.1.1 · pp. 10–12

4.1.2 Model-based Augmentation.

• Model-based augmentation creates multiple representations of the original data by perturbing the encoder, including message dropout, embedding or parameter noise, and architecture changes.Different dropout masks yield views from the same input, while architecture perturbations can contrast shallow and deep features.Source: §4.1.2 · pp. 12–13

4.2 Without Augmentation

• Without augmentation, methods contrast naturally available views such as domains, knowledge graphs, auxiliary behaviors, or user-item and user-bundle interactions.Examples include treating behavior types as views and contrasting across recommendation domains.Source: §4.2 · pp. 13–14

4.3 Discussion

• Data-based augmentations are common and easy to implement, but manual trial-and-error selection can limit generalizability, and some transformations may damage semantics and recommendation performance.The survey notes that these effects apply to some data-based augmentations, not necessarily all of them.Source: §4.3 · p. 14
• Without-augmentation strategies preserve semantics but can require costly, domain-specific knowledge and lack view randomness; model-based methods generalize better but still need manual tuning.The survey also identifies architecture-based perturbation design as challenging.Source: §4.3 · p. 14
• Hybrid strategies combine advantages, but adding data-based augmentation to naturally available views introduces randomness while retaining manual trial-and-error selection.For cold-start, the survey advises avoiding strategies that drop data and instead considers adding or substituting interactions or using model-based strategies.Source: §4.3 · p. 14

5 PRETEXT TASK

• Contrastive learning maximizes agreement between positive pairs with the same semantic information and minimizes agreement between unrelated negative pairs.Source: §5 · p. 15
• The survey divides contrastive pretext tasks into same-scale contrasting and cross-scale contrasting according to instance scale.Source: §5 · p. 15
• Local, contextual, and global scales correspond respectively to nodes, subgraphs, and whole graphs in graph data, or items, subsequences, and whole sequences in sequence data.Source: §5 · p. 15

5.1 Same-Scale Contrasting

• Same-scale contrasting divides into local-local, contextual-contextual, and global-global subtypes, with methods organized by recommendation task.Source: §5.1 · p. 15

5.1.1 Local-Local Contrasting.

• SGL contrasts representations of the same user or item across augmented interaction-graph views, using node dropout, edge dropout, or random walks and joint recommendation-loss training.It uses LightGCN [39] and treats different nodes in different views as negatives.Source: §5.1.1 · pp. 15–16
• Follow-on interaction-graph methods vary view construction through ego-net perturbation, graph diffusion, SVD, embedding noise, cross-layer comparison, or learnable graph-generation strategies.Examples include DCL [60], GDCL, LightGCL [7], SimGCL, XSimGCL, GCARec, LDA_GCL [43], AdaGCL [45], and VGCL.Source: §5.1.1 · pp. 16–17
• Local-local methods also contrast representations across distinct graphs, including interaction-derived user-item, item-item, user-user, and hypergraph views, or augmented subgraphs and paths.Examples include MCLSR, HCCF, SGCCL, and MPT [35].Source: §5.1.1 · pp. 17–18
• Knowledge-graph methods use semantic graph views to contrast item or user representations, and may guide view generation through edge importance or rationale scores.MCCLK and KACL contrast across graph views; KGCL uses knowledge-guided augmentation with intra- and inter-view contrasting, whereas KGRec uses inter-view contrasting only.Source: §5.1.1 · p. 18
• Local-local contrasts extend to multi-behavior recommendation through positives sampled by target-behavior similarity, social recommendation through interaction and social-network graph views with tri-training-predicted positives, and cross-domain recommendation by maximizing MI between same-node source- and target-domain representations.Source: §5.1.1 · pp. 18–19
• Other local-local applications include group, sequential, session, and bundle recommendation, with contrasts between graph views, layers, or predicted session items.S2-HHGR uses coarse- and fine-grained user dropout; DCRec adjusts contrastive regularization using user conformity and interest; COTREC predicts session-item positives and negatives; CrossCBR [63] contrasts corresponding user and bundle views.Source: §5.1.1 · pp. 20–21

5.1.2 Contextual-Contextual Contrasting.

• Contextual-contextual methods discriminate contextual representations: CHEST contrasts augmented HIN subgraphs, HMG-CR contrasts hyper meta-graphs for behavior patterns, and GCL4SR contrasts sampled sequence subgraphs.CHEST also combines generative pretext tasks with progressive curriculum pretraining; HMG-CR treats adjacent hyper meta-graphs as negatives; GCL4SR uses subgraphs from the same sequence as positives.Source: §5.1.2 · p. 21
• MISS [30] contrasts feature-vector representations, making different interests within categorical and sequential features dissimilar while aligning representations of the same interest.It uses two CNN-based models to extract multiple interests from each feature vector.Source: §5.1.2 · p. 21

5.1.3 Global-Global Contrasting.

• CL4SRec and related sequential methods contrast global sequence representations generated from augmented views, with variants changing the augmentations or positive pairs.CL4SRec uses cropping, shuffling, and masking; CoSeRec [59] adds item substitution and insertion, ContraRec also treats sequences sharing a target item as positive, and TiCoseRec [21] uses time-interval-based augmentations. IOCRec contrasts intention representations of augmented sequences with the same intent.Source: §5.1.3 · p. 22
• Other data-augmented sequential methods adapt sampling or augmentation to news frequency, curriculum learning, user/item neighbors, or item importance.MCCM varies news masking/substitution probability by frequency; CCL uses mask-and-fill recovery and an easy-to-difficult curriculum; MIC uses nearest neighbors as positives and different k-means++ clusters as negatives. EC4SRec uses explanation-based item scores to guide positive and negative sampling.Source: §5.1.3 · pp. 22–23
• Model-based sequential methods generate views through dropout, masking, or learned augmenters, and vary whether they use negative samples or contrast across networks.DuoRec [73] also samples sequences with the same target item as positives; CBiT [23] treats all K masked sequences as positives; CLUE omits negatives; FDSA_CL contrasts feature- and interaction-sequence representations. EMKD [24] contrasts within and across encoder networks, while MCLRec [71] learns two augmenters and contrasts four representations.Source: §5.1.3 · pp. 23–24
• Session-based methods contrast session representations across views or models, or between sessions linked by similarity.DHCN uses intra- and inter-session hypergraphs and perturbed representations; OD-Rec contrasts teacher and student representations of the same session; CGL aligns similar connected sessions and separates unconnected ones.Source: §5.1.3 · p. 24
• Feature-based methods contrast augmented feature vectors or align feature and collaborative representations, with CL4CTR avoiding negative samples.CFM uses feature dropout and masks related features to make its pretext task harder; CLCRec treats the same item's feature and collaborative representations as positives and different items' representations as negatives.Source: §5.1.3 · p. 24

5.2 Cross-Scale Contrasting

• Cross-scale tasks pair representations from different scales, with local-contextual, local-global, and contextual-global as the three subtypes.The survey frames the categories by the scales of the instances being contrasted.Source: §5.2 · p. 24
• NCL contrasts users or items with cluster prototypes, while ICL and SITN use sequence-cluster or cross-domain user-cluster pairs.NCL also contrasts representations across GNN layers; ICL additionally contrasts sequences, and SITN also contrasts users across domains.Source: §5.2 · p. 25
• MHCN contrasts user, user-centered sub-hypergraph, and hypergraph representations, while CubeRec pairs group-intersection representations with member users.SMIN constructs context using a substructure-aware adjacency matrix; CubeRec uses group intersections to model common interests.Source: §5.2 · p. 25

5.2.2 Local-Global Contrasting.

• Local-global methods pair edge and graph representations in EGLN, and node and graph representations in HGCL, which also adds cross-type contrasting.GroupIM contrasts member-user and group representations, using preference-biased sampling of non-member users as negatives.Source: §5.2.2 · p. 26

5.2.3 Contextual-Global Contrasting.

• Contextual-global methods include BiGI’s edge-context and graph pairing, MMSSL’s modality-specific and overall user pairing, and C2DSR’s single- and cross-domain sequence pairing.SSI contrasts a subsequence with its full sequence; SESRec uses positive and negative subsequences, while S3-Rec and TCPSRec define multiple sequence-, item-, attribute-, and periodicity-level tasks.Source: §5.2.3 · pp. 26–27

5.3 Discussion

• Cross-scale methods may need extra modules to generate large-scale representations and context-extraction strategies, increasing complexity.With a shared encoder, local-global graph methods first generate node representations and then aggregate them with a readout function.Source: §5.3 · p. 27
• Cross-scale contrasting models belonging relationships between small and large scales, while same-scale contrasting usually distinguishes different instances.Cross-scale methods usually draw negative samples from small-scale representations.Source: §5.3 · p. 27
• Cross-scale contrasting can add information to small-scale representations, but may also introduce irrelevant information as noise.Source: §5.3 · p. 28

6 CONTRASTIVE OBJECTIVE

• The objective is to maximize agreement between positive pairs and minimize it between negative pairs, with mutual information expressed through KL divergence.The survey classifies objectives by whether they provide a lower-bound estimate of mutual information.Source: §6 · p. 28

6.1 Bound Objective

• Bound objectives estimate mutual information through a lower bound; the survey identifies Jensen-Shannon and InfoNCE estimators as used in recommendation.The Jensen-Shannon estimator replaces KL divergence with Jensen-Shannon divergence and is described as more efficient than DV estimation.Source: §6.1 · p. 28, §6 · p. 28, §6.1.1 · p. 28

6.1.2 InfoNCE Estimator.

• InfoNCE is a popular MI lower-bound objective, and its loss commonly uses cosine similarity with a temperature parameter.In a mini-batch of size N + 1, the other N instances serve as negatives for each instance.Source: §6.1.2 · p. 29

6.2 Non-Bound Objective

• Triplet and BYOL losses lack proof that maximizing them maximizes mutual information, so they are classified as non-bound objectives.Source: §6.2 · p. 29

6.2.1 Triplet Loss.

• Triplet loss makes positive-pair agreement exceed negative-pair agreement rather than minimizing negative-pair agreement, and it is sensitive to its margin.Its performance depends on selecting informative, difficult-to-discriminate positive and negative samples.Source: §6.2.1 · p. 29, §6.3 · p. 30

6.2.2 BYOL Loss.

• BYOL maximizes positive-pair agreement without negative samples, but typically needs other designs, such as momentum encoders or stop-gradient, to prevent collapse.Source: §6.2.2 · p. 29, §6.3 · p. 30

6.3 Discussion

• InfoNCE is widely used and has lower estimated-MI variance than JS, but its need for many negatives raises computational and time costs.JS can achieve better performance with small batches.Source: §6.3 · p. 29
• InfoNCE is generally a good choice, JS may suit limited batch sizes, and triplet loss fits cases where pairs need not be absolutely discriminated.BYOL is efficient but needs collapse-prevention design; the survey also describes reweighting reliable data for noise.Source: §6.3 · p. 30

7.1 View Generation

• Current views mostly drop interactions or shuffle sequences; future strategies should adapt across tasks, remain efficient, and may update dynamically.Existing strategies often rely on intuitive designs and may not suit downstream tasks.Source: §7.1 · pp. 30–31

7.2 Pretext Task

• Pretext tasks should extract task-useful knowledge: CGI targets minimum sufficient information but is graph-specific, while combining and adaptively weighting tasks merits study.Different tasks can capture different information, so combining tasks may further improve recommendation performance.Source: §7.2 · p. 31
• Uniform negative sampling risks false negatives, while easy negatives provide little information; recommendation-specific and efficient sampling strategies remain open problems.Methods explored in computer vision [19] [47] are specifically designed for image data and are difficult to apply to recommendation.Source: §7.2 · p. 31

7.3 Contrastive Objective

• InfoNCE's KL-based mutual-information estimate can be asymmetric and unstable; Wasserstein discrepancy [25] is sequential-only, and V-information merits exploration as an alternative agreement measure.The applicability of Wasserstein discrepancy to other recommendation tasks remains to be explored.Source: §7.3 · p. 31

7.4 Miscellaneous

• Offline CL models may lag dynamic preferences; online learning and conversational recommendation using real-time feedback are proposed directions.Conversational recommendation models current preferences through multi-turn interaction with users.Source: §7.4.1 · pp. 31–32

7.4.1 Meeting Real-World Recommendation.

• CL denoising is mainly implicit, and explicit denoised views remain rarely explored; debiasing directions include popularity and selection bias.Denoising may impair diversity; DCRec adjusts contrastive regularization by disentangling user conformity and interest.Source: §7.4.1 · p. 32
• Multimodal contrastive learning may exploit video, text, and images, particularly with sparse interactions, but only a few studies explore it.The survey identifies deriving useful knowledge from multimodal data as a promising direction.Source: §7.4.1 · p. 32

7.4.2 Learning with Advanced Techniques.

• KD–contrastive combinations remain early-stage, alongside opportunities to combine CL with semi-supervised, meta-, and curriculum learning or use AutoML to select configurations.AutoML is proposed to reduce human effort in choosing among view-generation strategies, contrastive tasks, and encoders.Source: §7.4.2 · pp. 32–33

8 CONCLUSION

• The survey proposes a unified framework and organizes CL-based recommendation methods by view generation, pretext task, and contrastive objective.The review is presented as comprehensive and systematic.Source: §8 · p. 33
• For each framework component, the survey provides descriptions and discussion intended to guide method choice.Source: §8 · p. 33
• The survey discusses open issues and promising future research directions for contrastive self-supervised recommendation.Source: §8 · p. 33