Deep Interest Evolution Network for Click-Through Rate Prediction
Authors: Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, Kun Gai5941-5948
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments on both public and industrial datasets, DIEN significantly outperforms the state-of-the-art solutions. |
| Researcher Affiliation | Industry | Guorui Zhou,* Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, Kun Gai Alibaba Inc, Beijing, China {guorui.xgr, mouna.mn, fanying.fy, piqi.pq, weijie.bwj, ericzhou.zc, xiaoqiang.zxq, jingshi.gk}@alibaba-inc.com |
| Pseudocode | No | The paper describes mathematical formulations and network structures but does not include pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/mouna99/dien. |
| Open Datasets | Yes | public Dataset Amazon Dataset (Mc Auley et al. 2015) is composed of product reviews and metadata from Amazon. |
| Dataset Splits | No | For training set, we take the ads that clicked at last 49 days as the target item. Each target item and its corresponding click behaviors construct one instance. Using one target item a as example, we set the day that a is clicked as the last day, the behaviors that this user takes in previous 14 days as history behaviors. Similarly, the target item in test set is choose from the following one day, and the behaviors are built as same as training data. |
| Hardware Specification | No | ii) Batching: adjacent requests from different users are merged into one batch to take advantage of GPU. |
| Software Dependencies | No | The paper mentions various model architectures and techniques (e.g., GRU, LSTM, RNN) but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | There are 6 FCN layers used in industrial dataset, the dimensions area 600, 400, 300, 200, 80, 2, respectively, the max length of history behaviors is set as 50. Figure 2: Learning curves on public datasets. α is set as 1. |