Everyone’s Preference Changes Differently: A Weighted Multi-Interest Model For Retrieval
Authors: Hui Shi, Yupeng Gu, Yitong Zhou, Bo Zhao, Sicun Gao, Jishen Zhao
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments have been done on various industrial-scale datasets to demonstrate the effectiveness of our approach. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science and Engineering, University of California San Diego, La Jolla, United States, {hshi, jzhao, sicung}@ucsd.edu 2Pinterest, San Francisco, United States, {yupeng, yzhou, bozhao}@pinterest.com. |
| Pseudocode | No | The paper describes its methods through text and mathematical formulas but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | 1The code is available at: https://github.com/ shihui2010/MIP |
| Open Datasets | Yes | Three public datasets are used: Amazon-book 2 (hereinafter, Amazon), Taobao3, and Movie Lens4. ... 2https://jmcauley.ucsd.edu/data/amazon/ 3https://tianchi.aliyun.com/dataset/data Detail?data Id=649 4https://www.kaggle.com/grouplens/movielens-20m-dataset |
| Dataset Splits | Yes | We then split each user s engagement history to non-overlapping sequences of length 100, and use the first 50 items to learn the user embeddings and the last 50 items as labels (as used in Cen et al. (2020)). Any sequence shorter than this threshold are discarded. For each sequence, another 50 negative samples are uniformly sampled at random from the items that the user does not interact with. Our goal is to rank the positive items (that users have actually interacted with) higher than the negative items (random). AND Table 3. Dataset statistics. ... # Training Seq 57,165 127,212 343,171 # Test Seq 5,000 5,000 10,000 # Validation Seq 5,000 5,000 10,000 |
| Hardware Specification | Yes | All the models are trained for 100 epochs on a NVIDIA Tesla T4 GPU with an early stop strategy that stops the training when validation AUC does not improve for 20 epochs. AND In this section, we profiled the model latency on a desktop computer with a 12-core Intel i7-8700k CPU, and a single Nvidia Ge Force RTX 2080 Ti GPU. |
| Software Dependencies | Yes | The neural network training and inference are on the GPU with vanilla Py Torch framework (version 1.12) without any further optimization on the computation. |
| Experiment Setup | Yes | For fair comparison, we set up the configurations as follow: (1) item and user embedding vectors have the same size (d = 32); (2) the number of attention heads is the same (H = 8) if the model includes a multi-head attention module; (3) the baseline models should have similar or more parameters than MIP. We let the hidden size in GRU4Rec (Hidasi et al., 2015) be 128, the key and query projected dimension (dmodel in Equation 3) is labeled in place with the results, and if the model contains a position-wise FFN (Equation 6), it will be a two-layered fully-connected structure with a hidden size of 32 each. AND All the models are trained for 100 epochs on a NVIDIA Tesla T4 GPU with an early stop strategy that stops the training when validation AUC does not improve for 20 epochs. |