Personalized Question Routing via Heterogeneous Network Embedding
Authors: Zeyu Li, Jyun-Yu Jiang, Yizhou Sun, Wei Wang192-199
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that Ne Rank significantly outperforms competitive baseline question routing models that ignore the raiser information in three ranking metrics. In addition, Ne Rank is convergeable in several thousand iterations and insensitive to parameter changes, which prove its effectiveness, scalability, and robustness. We conduct extensive experiments on two real CQA datasets and evaluate the routing performance of Ne Rank via three ranking metrics. |
| Researcher Affiliation | Academia | Zeyu Li, Jyun-Yu Jiang, Yizhou Sun, Wei Wang University of California, Los Angeles Los Angeles, CA, 90095, USA {zyli, jyunyu, yzsun, weiwang}@cs.ucla.edu |
| Pseudocode | No | The paper describes algorithms and processes in text and figures, but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Available at: https://github.com/zyli93/Ne Rank |
| Open Datasets | Yes | Two datasets of two real-world CQA websites with specific topics are employed to evaluate Ne Rank: Biologyand English. Each dataset2 contains all questions raised before December, 2017 and all users historical asking and answering records. 2Available at: https://archive.org/details/stackexchange |
| Dataset Splits | No | CQA networks are built from 90% of questions and the corresponding users to generate training walks. The rest 10% of questions and the corresponding raisers and answerers for testing. |
| Hardware Specification | Yes | All experiments are conducted on a single 16GB-memory Tesla V100 GPU in an 512GB memory Nvidia DGX-1. |
| Software Dependencies | Yes | Ne Rank is prototyped by Python 3.6.4 and Py Torch 0.4.04. |
| Experiment Setup | Yes | The walks are generated from metapath AQRQA with the default length of 13 (three cycles) and the default node coverage of 20 (each node is covered at least 20 times in walk generation). The window size of Skip-gram model is set as 4; we use 3 negative samples per positive sample; and the dimension of learned embeddings is set as 256. We use a 64-channel CNN for ranking and the 300-dimensional Google News pretrained word2vec model3 to build the embedding matrix x for questions. |