README.md

HINGAN

A GAN-based recommendation approach. This is the code accompanying the ICWS 2019 paper "Generative Adversarial Network Based Service Recommendation in Heterogeneous Information Networks" Paper link: [https://ieeexplore.ieee.org/abstract/document/8818434]

Experimental results

top-3

Methods	Recall	Precision	MRR	NDCG
BPR-SVD	0.2947	0.1321	0.2345	0.2637
BPRSLIM	0.0585	0.0328	0.0523	0.0617
SVD	0.2350	0.1069	0.1893	0.2134
POP	0.2223	0.1203	0.2467	0.2586
PaSRec	0.6894	0.3143	0.7717	0.7269
IRGAN	0.3171	0.1418	0.2137	0.2177
CFGAN	0.3216	0.1472	0.3331	0.3020
Ours	0.4917	0.2203	0.4926	0.4570

top5

Methods	Recall	Precision	MRR	NDCG
BPR-SVD	0.3629	0.1000	0.2585	0.2966
BPRSLIM	0.0849	0.0300	0.0640	0.0768
SVD	0.3434	0.0962	0.2253	0.2653
POP	0.2929	0.1000	0.2785	0.2947
PaSRec	0.7253	0.2046	0.7901	0.7360
IRGAN	0.3703	0.1038	0.2281	0.2400
CFGAN	0.3806	0.1087	0.3471	0.3260
Ours	0.5295	0.1466	0.5015	0.4705

top10

Methods	Recall	Precision	MRR	NDCG
BPR-SVD	0.4691	0.0665	0.2822	0.3343
BPRSLIM	0.1183	0.0223	0.0737	0.0892
SVD	0.4791	0.0686	0.2533	0.3127
POP	0.4051	0.0692	0.3040	0.3349
PaSRec	0.7727	0.1131	0.8048	0.7428
IRGAN	0.4474	0.0643	0.2379	0.2658
CFGAN	0.4621	0.0675	0.3602	0.3551
Ours	0.5635	0.0799	0.5060	0.4820

The importance of 8 metal paths

m-s-m-s is the most important

impact on hyperparameters

α 是G模型的正则化参数，同CFGAN

---

β 是G模型的正则化参数

---

Embedding size; 每个特征embedding的维度

λ 把训练集抽样为负样本的比例，CFGAN中是 按个数抽，这里是按照比例抽。