Language Representations Can be What Recommenders Need: Findings and Potentials
Authors: Leheng Sheng, An Zhang, Yi Zhang, Yuxin Chen, Xiang Wang, Tat-Seng Chua
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Surprisingly, our findings demonstrate that item representations, when linearly mapped from advanced LM representations, yield superior recommendation performance. This outcome suggests the possible homomorphism between the advanced language representation space and an effective item representation space for recommendation, implying that collaborative signals may be implicitly encoded within LMs. Empirical results show that such a simple model can outperform leading ID-based CF models on multiple datasets, which sheds light on using language representations for better recommendation. Moreover, we systematically analyze this simple model and find several key features for using advanced language representations: a good initialization for item representations, superior zero-shot recommendation abilities in new datasets, and being aware of user intention. |
| Researcher Affiliation | Academia | 1National University of Singapore 2University of Science and Technology of China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the model architecture and procedures using mathematical equations (1), (2), (3), (4) and descriptive text, but does not include a formally structured pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain any explicit statements about providing source code, a repository link, or code in supplementary materials for the methodology described. |
| Open Datasets | Yes | Table 1 reports the performance yielded by post-mapping representations on three Amazon datasets (Ni et al., 2019), comparing with classic ID-based CF baselines... Experimental settings. In zero-shot recommendation... we jointly train Alpha Rec on three source datasets (i.e., Books, Movies & TV, and Video Games), while testing it on three completely new target datasets (i.e., Movielens-1M (Harper & Konstan, 2016), Book Crossing (Lee et al., 2019), and Amazon Industrial & Scientific (Ni et al., 2019)). |
| Dataset Splits | Yes | We divide the history interaction of each user into training, validation, and testing sets with a ratio of 4:3:3, and remove users with less than 20 interactions following previous studies (Zhang et al., 2023b). We also remove items from the test and validation sets that do not appear in the training set, to address the cold start problem. |
| Hardware Specification | Yes | We conduct all the experiments in Py Torch with a single NVIDIA RTX A5000 (24G) GPU and a 64 AMD EPYC 7543 32-Core Processor CPU. |
| Software Dependencies | No | We conduct all the experiments in Py Torch with a single NVIDIA RTX A5000 (24G) GPU and a 64 AMD EPYC 7543 32-Core Processor CPU. We optimize all methods with the Adam optimizer. |
| Experiment Setup | Yes | We optimize all methods with the Adam optimizer. For all ID-based CF methods, we set the layer numbers of graph propagation by default at 2, with the embedding size as 64 and the size of sampled negative items |Su| as 256. We use the early stop strategy to avoid overfitting. We stop the training process if the Recall@20 metric on the validation set does not increase for 20 successive evaluations. In Alpha Rec, the dimensions of the input and output in the two-layer MLP are 3072 and 64 respectively, with the hidden layer dimension as 1536. We apply the all-ranking strategy (Krichene & Rendle, 2020) for all experiments... The hyperparameter search space is reported in Table 14. For Alpha Rec, the only hyperparameter is the temperature τ and we search it in [0.05, 2]. |