Adaptive Factorization Network: Learning Adaptive-Order Feature Interactions

Authors: Weiyu Cheng, Yanyan Shen, Linpeng Huang3609-3616

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results on four real datasets demonstrate the superior predictive performance of AFN against the state-of-the-arts. In this section, we conduct experiments to answer the following research questions: RQ1: How do our proposed methods AFN and AFN+ perform against the state-of-the-art methods? RQ2: How does the performance of AFN vary with different settings of the hyper-parameters? RQ3: What are the learned feature orders in AFN, and can AFN find useful cross features from data?
Researcher Affiliation Academia Weiyu Cheng, Yanyan Shen, Linpeng Huang Shanghai Jiao Tong University {weiyu cheng, shenyy, lphuang}@sjtu.edu.cn
Pseudocode No The paper describes methods using equations and prose but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes We implement our methods using Tensorflow5. 5https://github.com/Weiyu Cheng/AFN-AAAI-20
Open Datasets Yes We conduct experiments with four publicly accessible datasets following previous works (Lian et al. 2018; He and Chua 2017): Criteo1, Avazu2, Movielens3 and Frappe4. 1http://labs.criteo.com/2014/02/kaggle-display-advertising-challenge-dataset/ 2https://www.kaggle.com/c/avazu-ctr-prediction 3https://grouplens.org/datasets/movielens/ 4http://baltrunas.info/research-menu/frappe
Dataset Splits Yes For each dataset, we randomly split the instances by 8:1:1 for training, validation and test, respectively.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running the experiments.
Software Dependencies No The paper mentions 'Tensorflow' and 'Adam optimizer' but does not specify their version numbers or other software dependencies with version information.
Experiment Setup Yes We apply Adam with a learning rate of 0.001 and a mini-batch size of 4096. The default number of logarithmic neurons is set to 1500, 1200, 800 and 600 for Criteo, Avazu, Movielens and Frappe datasets, respectively. We use 3 hidden layers and 400 neurons per layer by default in AFN. To avoid overfitting, we perform early-stopping according to the AUC on the validation set. We set the rank of feature embeddings to 10 in all the models.