Dot Product is All You Need: Bridging the Gap Between Item Recommendation and Link Prediction (2409.07433v1)

Abstract: Item recommendation (the task of predicting if a user may interact with new items from the catalogue in a recommendation system) and link prediction (the task of identifying missing links in a knowledge graph) have long been regarded as distinct problems. In this work, we show that the item recommendation problem can be seen as an instance of the link prediction problem, where entities in the graph represent users and items, and the task consists of predicting missing instances of the relation type <<interactsWith>>. In a preliminary attempt to demonstrate the assumption, we decide to test three popular factorisation-based link prediction models on the item recommendation task, showing that their predictive accuracy is competitive with ten state-of-the-art recommendation models. The purpose is to show how the former may be seamlessly and effectively applied to the recommendation task without any specific modification to their architectures. Finally, while beginning to unveil the key reasons behind the recommendation performance of the selected link prediction models, we explore different settings for their hyper-parameter values, paving the way for future directions.

• The paper demonstrates that recasting item recommendation as link prediction using dot product embeddings yields competitive performance.
• It evaluates popular link prediction models, including DistMult, CP, and ComplEx, against state-of-the-art recommendation systems on benchmark datasets.
• Results reveal that richer embedding representations improve metrics like Recall@20 and nDCG@20, bridging theoretical insights with practical applications.

Dot Product is All You Need: Bridging the Gap Between Item Recommendation and Link Prediction

This paper titled "Dot Product is All You Need: Bridging the Gap Between Item Recommendation and Link Prediction" investigates whether link prediction models can be effectively utilized for item recommendation tasks without requiring any specific modifications to their architectures. The central premise is that item recommendation can be viewed as a link prediction problem, with users and items represented as entities in a knowledge graph connected via a single type of relation: "interactsWith."

Abstract

Item recommendation focuses on predicting user interactions with new items, while link prediction identifies missing links in a knowledge graph. This paper proposes that the two problems share intrinsic similarities, particularly when approached via factorisation-based techniques, where the interaction score is computed using the dot product of latent embeddings. It reviews and evaluates three popular link prediction models—DistMult, CP, and ComplEx—against ten state-of-the-art recommendation systems. The empirical results suggest that link prediction models are competitive without requiring architectural modifications.

Key Insights

Recasting Item Recommendation as Link Prediction:

• The authors propose treating item recommendation as link prediction by framing users and items within a knowledge graph, connected through an "interactsWith" relation.
• Formally, this approach redefines the interaction matrix of a recommendation system as a knowledge graph, where item recommendation translates into predicting missing interactions.

Selected Models and Baselines:

• Link Prediction Models:
  • DistMult: Uses a factorisation model with tri-linear dot product for scoring.
  • CP (Canonical Tensor Decomposition): Similar to DistMult but with double representations for entities.
  • ComplEx: Encompasses complex-valued embeddings and computes the real part of the tri-linear dot product.
• Item Recommendation Models:
  • Ranging from simple heuristics like MostPop and Random to modern GNN-based approaches like NGCF, DGCF, and LightGCN.
  • Evaluation included user-kNN and item-kNN as traditional baselines, as well as advanced GNN approaches like UltraGCN and GFCF.

Experimental Setup

The experimental setup included the use of three standard benchmarks: Gowalla, Yelp 2018, and Amazon Book. These datasets were recast from their native formats to a knowledge graph consisting of user-item interactions, allowing seamless training and evaluation of link prediction models within the item recommendation context.

Evaluation Metrics:

• Recall@20 and nDCG@20 were utilized as primary evaluation metrics.
• Modifications to the LibKGE framework were necessary to compute these metrics directly for item recommendation tasks.

Results

The performance of link prediction models was comparable to state-of-the-art item recommendation systems. DistMult, CP, and ComplEx were able to consistently place in the top-5 models in terms of both Recall@20 and nDCG@20 on the Yelp 2018 and Amazon Book datasets. Additionally, results for higher embedding sizes indicated better performance in general, validating the common wisdom in both link prediction and recommendation system literature that richer representations capture more relevant interactions.

Implications and Future Work

Theoretical Implications:

• The work substantiates theoretical speculations about the shared optimization properties and performance potentials between item recommendation and link prediction tasks.
• It encourages a rethinking of model architectures and training paradigms, suggesting that innovations from the link prediction domain can seamlessly improve item recommendation systems.

Practical Implications:

• Practically, existing link prediction models can be deployed for item recommendation tasks without significant modifications, leveraging their robust training mechanisms and efficiency.
• This also implies that businesses and platforms using item recommendation systems could potentially integrate or switch to tested link prediction models, reducing time-to-market for new system updates.

Future Directions:

The paper prompts further investigation into other facets of link prediction models, particularly:

• Experimenting with bi-directional predicates or employing advanced regularization techniques.
• Exploring different loss functions and optimizing the interplay of hyper-parameters.
• Building more complex models by integrating multi-relational knowledge graphs with richer interaction semantics.

Conclusion

This paper offers a compelling case for bridging the gap between item recommendation and link prediction. Their results convincingly demonstrate that off-the-shelf link prediction models are competitive for item recommendation tasks. The implications of this research are extensive, suggesting a unified approach could be beneficial for both academic research and practical applications in AI-driven recommendation systems.