Source-linked AI summary
Content-Based Collaborative Generation for Recommender Systems
Yidan Wang, Zhaochun Ren, Weiwei Sun, Jiyuan Yang, Zhixiang Liang, Xin Chen, Ruobing Xie, Su Yan, Xu Zhang, Pengjie Ren, Zhumin Chen, Xin Xin
TL;DR
Generative recommendation still faces the challenge of jointly modeling and aligning item content with collaborative signals. ColaRec uses sequence-to-sequence generation with collaborative GIDs, content aggregation, item indexing, and contrastive alignment. Experiments on four real-world datasets demonstrate its effectiveness and superior recommendation performance.
Problem
Existing generative recommendation methods do not effectively integrate and align item content with user-item collaborative signals in a unified framework.
Source: §2.2 · p. 3Method
ColaRec generates recommended-item GIDs from users' interacted-item content, constructs GIDs using a pretrained collaborative filtering model, and adds item indexing and contrastive alignment.
Source: §1 · p. 2, §4.5 · p. 5Results
Takeaways & Limitations
ColaRec provides an end-to-end generative recommender that integrates item content and user-item collaborative signals and aligns their representations.
Source: §6 · p. 9GID construction is expected to encode both collaborative signals and content, correlate related items, and uniquely identify each item.
Source: §4.2 · p. 4
Abstract
from arXiv · showhide
Generative models have emerged as a promising utility to enhance recommender systems. It is essential to model both item content and user-item collaborative interactions in a unified generative framework for better recommendation. Although some existing large language model (LLM)-based methods contribute to fusing content information and collaborative signals, they fundamentally rely on textual language generation, which is not fully aligned with the recommendation task. How to integrate content knowledge and collaborative interaction signals in a generative framework tailored for item recommendation is still an open research challenge. In this paper, we propose content-based collaborative generation for recommender systems, namely ColaRec. ColaRec is a sequence-to-sequence framework which is tailored for directly generating the recommended item identifier. Precisely, the input sequence comprises data pertaining to the user's interacted items, and the output sequence represents the generative identifier (GID) for the suggested item. To model collaborative signals, the GIDs are constructed from a pretrained collaborative filtering model, and the user is represented as the content aggregation of interacted items. To this end, ColaRec captures both collaborative signals and content information in a unified framework. Then an item indexing task is proposed to conduct the alignment between the content-based semantic space and the interaction-based collaborative space. Besides, a contrastive loss is further introduced to ensure that items with similar collaborative GIDs have similar content representations. To verify the effectiveness of ColaRec, we conduct experiments on four benchmark datasets. Empirical results demonstrate the superior performance of ColaRec.
1 INTRODUCTION
The introduction frames a gap in generative recommendation: existing approaches do not jointly model and align item content with user-item collaborative signals. ColaRec addresses this with direct GID generation, auxiliary item indexing, and contrastive alignment.
1 INTRODUCTION
- LLM-based recommendation can require grounding generated language to concrete items and struggles to rank target itemIDs from large candidate pools.These task mismatches limit the practical usage of LLM-based methods.Source: §1 · pp. 1–2
- Existing generative recommendation methods tend to omit either collaborative signals or item content, while simple concatenation has not effectively aligned them.The introduction contrasts content-derived GIDs that overlook collaboration with interaction-derived GIDs that omit content, and describes naive concatenation as suboptimal.Source: §1 · p. 2
- ColaRec combines item content and user-item collaborative signals in a sequence-to-sequence framework that directly generates the recommended item's GID.Its input aggregates content from a user's interacted items, while its GIDs come from a pretrained collaborative filtering model.Source: §1 · pp. 2–3
- Auxiliary item indexing maps item side information into GIDs, while contrastive loss encourages items with similar collaborative GIDs to have similar content representations.The indexing input includes both item text and users who interacted with the item.Source: §1 · p. 2
- On four public datasets, ColaRec outperforms related state-of-the-art baselines.The paper also identifies its encoder-decoder framework, indexing task, and contrastive loss as contributions.Source: §1 · pp. 2–3
2 RELATED WORK
The related-work section reviews collaborative filtering and generative approaches to recommendation, including language-based methods. It identifies joint modeling and alignment of content and collaborative signals as an open challenge.
2.1 Collaborative Filtering
- Collaborative filtering represents users and items through their interactions; approaches discussed range from matrix factorization to deep and graph-based models.The review also notes work that incorporates visual or textual content into collaborative filtering.Source: §2.1 · p. 3
2.2 Generative Models for Recommendation
- LLM-based approaches turn recommendation into prompted natural-language generation, but may require grounding text to items and face difficulty generating from large candidate pools.The review says this language-generation mismatch limits their practical usage.Source: §2.2 · p. 3
- Generative recommendation assigns items token-sequence GIDs and trains a sequence-to-sequence model to generate recommended-item identifiers directly.The review describes methods using content-derived, sequential, or other item identifiers.Source: §2.2 · p. 3
2 RELATED WORK
- Jointly modeling and aligning collaborative signals with item content in an end-to-end generative framework tailored for item recommendation remains an open challenge.The section's opening frames its scope as reviewing collaborative filtering and generative models for recommendation.Source: §2.2 · p. 3, §2 · p. 3
3 NOTATIONS AND TASK FORMULATION
The notation distinguishes users, items, interaction sets, content descriptions, and two item identifiers. The task is to generate recommendation GIDs autoregressively and return the highest-scoring items.
3.1 Notations
- For user u and item i, I+_u denotes u's interacted items, U+_i denotes users who interacted with i, and c_i denotes item i's content description.Source: §3.1 · p. 3
- ColaRec's overview assigns each item a GID from a GNN-based collaborative-filtering model and includes recommendation and item-indexing tasks.The overview also identifies ranking and contrastive losses.Source: Figure 3 · p. 4
- A randomly assigned single token identifies an item as its atomic identifier, distinct from its generative identifier.Source: §3.1 · p. 4
- The item's GID is a sequence of tokens, with l denoting its length.Source: §3.1 · p. 4
3.2 Generative Recommendation
- Generative recommendation takes information describing a user's interacted items and generates a list of item GIDs autoregressively.Source: §3.2 · p. 4
- The recommender ranks items by p(u,i) and returns the top-n highest-scoring items for user u.Source: §3.2 · p. 4
4 METHODOLOGY
ColaRec shares an encoder-decoder framework across recommendation and item indexing, combining content inputs with collaborative GIDs. Hierarchical clustering, indexing, and auxiliary ranking and contrastive objectives support this integration.
4.1 Overview of ColaRec
- ColaRec's shared encoder-decoder model maps a user's historical item content to a recommended item's GID, while item indexing maps item side information to its GID.The recommendation task combines content and collaborative signals; indexing aligns them.Source: §4.1 · p. 4
4.2 Generative Identifier Construction
- Hierarchical clustering of pretrained LightGCN item representations forms a K-ary tree whose root-to-leaf paths serve as collaborative GIDs.Codebook embeddings at each GID position incorporate content information during item indexing.Source: §4.2 · p. 4
4.3 User-Item Recommendation
- For recommendation, ColaRec aggregates content tuples of a user's interacted items and uses an encoder-decoder Transformer to generate the target item's GID.T5 is fine-tuned as the Transformer, and constrained beam search limits generation to valid GIDs.Source: §4.3 · p. 5, §4.5 · p. 5
4.4 Item-Item Indexing
- The item-indexing task maps from item content's semantic space into the interaction-based collaborative space using the shared model and codebook embeddings.Its input includes item textual information and information about users who interacted with the item.Source: §4.4 · p. 5
4.5 Multi-Task Training
- ColaRec trains recommendation and item-indexing tasks jointly, adding ranking and contrastive losses to align content information with collaborative signals.The ranking loss separates interacted from sampled non-interacted items; the contrastive loss pairs items with overlapping GID prefixes and uses a negative with no overlapping prefix.Source: §4.5 · p. 5
5 EXPERIMENT
The experiments evaluate ColaRec on four public datasets. They cover three Amazon product categories and Recipe from Food.com.
5 EXPERIMENT
- The experiments address comparisons with existing recommendation methods, the effects of joint multi-task training, and how GID design affects recommendation performance.The experiment section poses these as three research questions.Source: §5 · p. 6
5.1 Datasets
- The evaluation uses Beauty, Sports and Outdoors, Cell Phones and Accessories, and Recipe datasets, filtering out users and items with fewer than five interactions.Item content comes from title, brand, and categories for Amazon, and name, description, and tag for Recipe.Source: §5.1 · p. 6
- The dataset setup spans three Amazon product categories and one Food.com recipe dataset, with users and items below five interactions removed.The resulting dataset statistics are presented in Table 1.Source: §5.1 · p. 6
5.2 Evaluation Protocols
The evaluation uses repeated cross-validation with recall and NDCG, ranking against the full item set. ColaRec is compared with conventional collaborative-filtering and generative recommendation baselines.
5.2 Evaluation Protocols
- Recall@n and NDCG@n evaluate recommendations against the whole item set, using per-user 8:1:1 train/validation/test splits and three-run averages.Recall measures ground-truth item inclusion in the recommended list; NDCG also accounts for ranking.Source: §5.2 · p. 6
5.3 Baselines
- Baselines include NeuMF, LightGCN, SimpleX, NCL, MultiVAE, DiffRec, DSI, TIGER, and LC-Rec.The set spans conventional collaborative-filtering methods and generative recommendation models.Source: §5.3 · p. 6
- P5-based baselines are excluded because their general-recommendation prompts require candidate items, preventing ranking over the whole item set within limited input length.The DSI comparisons include random-string GIDs and content-embedding-based hierarchical K-means GIDs.Source: §5.3 · p. 6
5.4 Implementation Details
Implementation settings specify GID and clustering sizes, sampling procedures, embedding dimensions, and optimization choices.
5.4 Implementation Details
- GID length is 3; hierarchical K-means uses 32 clusters, except Recipe uses 48.Source: §5.4 · p. 6
- Negative instances for the BPR and contrastive losses are uniformly sampled, and investigating negative-sampling strategies is left for future work.Users and item-indexing inputs also use randomly sampled interacted items and users, respectively.Source: §5.4 · p. 6
- ColaRec's user, item, and codebook embeddings are 512-dimensional; AdamW uses a 5e-4 learning rate and batch size 128.Baseline user and item embedding sizes are also set to 512 for comparison.Source: §5.4 · p. 6
5.5 Performance Comparison (RQ1)
ColaRec performs strongly against collaborative-filtering and generative baselines on overall users, with an even more pronounced advantage for long-tail users. Its reported comparisons cover four datasets and multiple recommendation metrics.
5.5.1 Comparison on whole users.
- ColaRec achieves the best results on all four datasets except for NDCG@5 in Beauty, where it is comparable to NCL, and improves Recall@20 over prior CF-based and generative models.The relative Recall@20 improvements are 3.87% on Beauty, 3.54% on Sports, 6.72% on Phone, and 3.88% on Recipe.Source: §5.5.1 · p. 7, Table 2 · p. 7
- ColaRec improves Recall@5 over DSI-R and DSI-S by 47.89%, 38.13%, 80.82%, and 26.11% across the four datasets, respectively.The authors also report that ColaRec consistently outperforms TIGER and LC-Rec.Source: §5.5.1 · p. 7
- Among generative methods, DiffRec and TIGER can show promise in specific scenarios but underperform strong CF methods on general recommendation, while ColaRec is competitive with CF methods across datasets.The authors describe the improvement as more significant for long-tail users.Source: §5.5.1 · p. 7, §5.5.2 · p. 7
5.5.2 Comparison on long-tail users.
- With the head-to-long-tail user ratio set at 20%:80%, ColaRec significantly outperforms all baselines on Recall for long-tail users across four datasets, using user-item interactions and item content information.Source: §5.5.2 · p. 7, Table 3 · p. 8
5.6 Ablation Study (RQ2)
Ablations indicate that each tested ColaRec component contributes to recommendation performance. Removing textual content, ranking loss, indexing, or contrastive loss leads to a decline.
5.6 Ablation Study (RQ2)
- Removing textual content produces Recall@5 drops of 20.99%, 17.65%, 14.63%, and 8.59% across the four datasets.The authors interpret this result as evidence for the importance of incorporating textual item content.Source: §5.6 · p. 8
- Removing the pairwise ranking loss Lbpr reduces performance, supporting its role in prioritizing positive items within generative recommendation.The ablation compares ColaRec with a variant trained without Lbpr.Source: §5.6 · p. 8
- Removing either the item-item indexing task or contrastive loss Lc consistently reduces performance across the four datasets.The authors say these techniques align item content information with user-item collaborative signals.Source: §5.6 · p. 8
- The ablation study concludes that every tested ColaRec component is essential for improving recommendation performance.Source: §5.6 · p. 8
5.7 GID Investigation (RQ3)
The GID investigation compares identifier types and examines how GID length and cluster count affect generation. Collaborative GIDs perform best in the reported comparison, while length and cluster count involve trade-offs in decoding search space and difficulty.
5.7 GID Investigation (RQ3)
- The GID analysis covers identifier type as well as the length and K-means cluster count used in GID construction.Source: §5.7 · pp. 8–9
5.7.2 Impact of GID Hyper-parameters.
- With GID length l varied from 1 to 4, l=3 gives the best performance in most cases by balancing decoding steps against each step’s search space.The authors report that l=4 adds autoregressive decoding steps, increasing generation difficulty and inference latency; l=3 is their default.Source: Figure 4 · p. 9, §5.7.2 · p. 9
- With GID length fixed at 3, increasing cluster count K to 64, 96, or 128 typically slightly decreases performance, with a more notable decrease on Beauty.A larger K increases decoding search space; K must be large enough to encode all items but controlled thereafter.Source: Figure 5 · p. 9, §5.7.2 · p. 9
5.7.1 Effect of different GID types.
- ColaRec’s collaborative GIDs attain the best performance across four datasets among the tested identifier types, including item-ID, random, and content-based GIDs.The authors attribute the advantage over content-based GIDs to collaborative signals and the advantage over item-ID GIDs to explicitly representing item correlations.Source: §5.7.1 · p. 9
6 CONCLUSIONS AND FUTURE WORK
ColaRec combines item content and user-item collaborative signals in an end-to-end generative recommender, with item indexing and contrastive loss supporting alignment. Future work includes GID construction, alignment, negative sampling, and model efficiency.
6 CONCLUSIONS AND FUTURE WORK
- ColaRec integrates item content and user-item collaborative signals in an end-to-end generative framework, using an auxiliary item indexing task and contrastive loss to align their representations.Source: §6 · p. 9
- Future work includes investigating additional GID construction methods and more effective approaches for aligning content information with collaborative signals.Source: §6 · p. 9
- The authors identify negative sampling and improving generative recommendation efficiency as further research directions.Proposed negative-sampling directions include using GIDs to sample informative negatives or generative models to produce synthetic negative instances.Source: §6 · p. 9