Learning Feature Interactions with Lorentzian Factorization Machine
Authors: Canran Xu, Ming Wu6470-6477
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on several benchmark datasets show that, even without using any deep learning layers on top of the embedding layer, the proposed approach will achieve the state-of-the-art performance with reducing up to 82.9% in training parameters and 69.7% in training time compared to the strong deep learning baselines such as Deep FM, DCN and x Deep FM. We conduct extensive experiments on a wide range of real-world datasets. Our results demonstrate that Lorentz FM achieves state-of-the-art performance, compared to existing deep learning approaches. |
| Researcher Affiliation | Collaboration | Canran Xu,1 Ming Wu2 1e Bay Inc. 2Department of Electrical Engineering, Tongji University, China |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions implementing Lorentz FM and baselines in PyTorch, but does not provide concrete access to its own source code for the methodology described. |
| Open Datasets | Yes | We use Steam, Movie Lens and KKBox datasets to evaluate the recommendation task and Avazu dataset for the CTR task. ... Steam1 is a dataset crawled from Steam database, which includes rich information such as users playing hours, games price, category and publisher etc. ... Movie Lens2 is a set of benchmark datasets for evaluating recommendation algorithms. ... KKBox3 dataset is adopted from the WSDM cup 2018 Challenge provided by the music streaming service KKBox. ... Avazu4 is an ads click-through dataset with more than 40 millions instances. ... 1http://cseweb.ucsd.edu/ jmcauley/datasets.html#steam data 2https://grouplens.org/datasets/movielens/1m/ 3https://www.kaggle.com/c/kkbox-music-recommendation-challenge 4https://www.kaggle.com/c/avazu-ctr-prediction |
| Dataset Splits | Yes | For the three recommendation datasets we process the dataset as follows: ... randomly splitting {10K, 10K} samples for {validation, testing}, and the rest as training set. For the Avazu dataset, we remove the timestamp field and then replace features less than 5 times by a universal unknown tag. We randomly select 80% samples for training and, 10% for validation and 10% for testing. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions implementing Lorentz FM and all baselines in Py Torch5 (Paszke et al. 2017) and using Adam optimizer (Kingma and Ba 2015), but it does not specify concrete version numbers for PyTorch or any other software libraries required for replication. |
| Experiment Setup | Yes | The embedding size is set to 10 for all models for fair comparisons. For the recommendation task, we tune the learning rate [0.05, 0.1, 0.2, 0.3] with RSGD and batch size [64, 128, 256, 512]. For the CTR prediction task, the learning rate is tuned [0.1, 0.2, 0.3] and batch size is set to 4096. Baselines Details The loss function is BPR loss for the recommendation task and BCE loss for the CTR task. We tune the dropout rate p {0, 0.1, . . . , 0.5}, the MLP layers [[512, 256], [400, 400], [100, 100]], and the batch size [64, 128, 256, 512]. Learning rate is set to 0.001 with Adam optimizer (Kingma and Ba 2015). We use an L2 regularization with λ = 10 5. |