[论文解读] Diffusion Cross-domain Recommendation

Yuner Xuan|arXiv (Cornell University)|Feb 3, 2024

Recommender Systems and Techniques被引用 6

一句话总结

DiffCDR 引入一个基于扩散概率模型的跨域推荐，将用户嵌入从丰富的辅助域转移到稀疏的目标域，使用扩散模块和对齐模块，在冷启动和暖启动下实现强性能。

ABSTRACT

It is always a challenge for recommender systems to give high-quality outcomes to cold-start users. One potential solution to alleviate the data sparsity problem for cold-start users in the target domain is to add data from the auxiliary domain. Finding a proper way to extract knowledge from an auxiliary domain and transfer it into a target domain is one of the main objectives for cross-domain recommendation (CDR) research. Among the existing methods, mapping approach is a popular one to implement cross-domain recommendation models (CDRs). For models of this type, a mapping module plays the role of transforming data from one domain to another. It primarily determines the performance of mapping approach CDRs. Recently, diffusion probability models (DPMs) have achieved impressive success for image synthesis related tasks. They involve recovering images from noise-added samples, which can be viewed as a data transformation process with outstanding performance. To further enhance the performance of CDRs, we first reveal the potential connection between DPMs and mapping modules of CDRs, and then propose a novel CDR model named Diffusion Cross-domain Recommendation (DiffCDR). More specifically, we first adopt the theory of DPM and design a Diffusion Module (DIM), which generates user's embedding in target domain. To reduce the negative impact of randomness introduced in DIM and improve the stability, we employ an Alignment Module to produce the aligned user embeddings. In addition, we consider the label data of the target domain and form the task-oriented loss function, which enables our DiffCDR to adapt to specific tasks. By conducting extensive experiments on datasets collected from reality, we demonstrate the effectiveness and adaptability of DiffCDR to outperform baseline models on various CDR tasks in both cold-start and warm-start scenarios.

研究动机与目标

• 通过利用辅助域数据解决跨域推荐系统中的冷启动问题。
• 提出基于扩散的映射模块以跨域转移用户嵌入。
• 通过对齐模块稳定跨域转移。
• 纳入目标域标签数据以使扩散输出与任务目标对齐。
• 在真实世界的亚马逊域CDR任务中，在冷启动和暖启动设置下展示有效性。

提出的方法

• 引入扩散模块（DIM），通过在源域嵌入条件下反向扩散来生成目标域用户嵌入。
• 增加对齐模块（ALM），以减少随机性并将转移后的嵌入与真实目标嵌入对齐。
• 使用快速扩散求解器实现 DIM 推断。
• 使用 DIM 损失使预测噪声与真实噪声匹配，以及将 ALM+任务损失将转移嵌入与目标域评分联系起来。
• 通过将 ALM 输出与目标域评分相结合来采用面向任务的损失，以适应特定任务。
• 进行消融研究以分离 DIM、ALM 和面向任务的学习的贡献。

Figure 1. Illustration of the DiffCDR concept. (a) Learning the Diffusion Module (DIM) $p_{\theta}$ on the embeddings of overlapping users (b) Generating transferred embeddings via reverse diffusion process for cold-start users conditioned on corresponding latent vectors from auxiliary domain.

实验结果

研究问题

• RQ1 DiffCDR 在冷启动和暖启动设置下相对于最先进的跨域推荐基线的表现如何？
• RQ2 DIM 与 ALM 组件对性能的贡献，以及包括目标标签任务损失的影响？
• RQ3 为什么基于扩散的转移在跨域知识迁移中会带来改进？
• RQ4 与传统方法相比，DiffCDR 的推理吞吐量是多少？

主要发现

贝塔	CDR 任务	指标	目标域	CMF	EMCDR	SSCDR	LACDR	PTUPCDR	DiffCDR	改进幅度
20%	Task1 Video → Music	MAE	4.4546	1.4642	1.3596	1.1757	1.1295	1.1099	1.0435*	6.0%
20%	Task1 Video → Music	RMSE	5.1338	1.9571	1.6615	1.4911	1.4358	1.4543	1.3840*	3.6%
20%	Task1 Video → Music	N@20	0.00253	0.00508	0.00977	0.00932	0.00984	0.00978	0.01026*	4.3 %
20%	Task1 Video → Music	H@20	0.00033	0.00084	0.00229	0.00212	0.00228	0.00236	0.00238*	1.1 %
50%	Task1 Video → Music	MAE	4.4884	1.6710	1.6891	1.4320	1.3502	1.2842	1.2367*	3.7%
50%	Task1 Video → Music	RMSE	5.1790	2.2076	2.0368	1.8248	1.7510	1.7340	1.6859*	2.8%
50%	Task1 Video → Music	N@20	0.00251	0.00403	0.00898	0.00793	0.00893	0.00828	0.00915*	1.9%
50%	Task1 Video → Music	H@20	0.00033	0.00068	0.00193	0.00164	0.00199	0.00179	0.00202*	1.9%
80%	Task1 Video → Music	MAE	4.4959	2.2327	2.1980	1.8162	1.6886	1.6174	1.5606*	3.5%
80%	Task1 Video → Music	RMSE	5.1830	2.8868	2.5713	2.3090	2.2238	2.2429	2.1754*	2.2%
80%	Task1 Video → Music	N@20	0.00248	0.00348	0.00622	0.00578	0.00606	0.00545	0.00665*	6.9%
80%	Task1 Video → Music	H@20	0.00033	0.00051	0.00124	0.00111	0.00124	0.00107	0.00136*	9.7%
20%	Task2 Book → Video	MAE	4.1807	1.4742	1.1305	0.9774	0.9681	1.0728	0.9476*	2.1%
20%	Task2 Book → Video	RMSE	4.7496	1.9180	1.4215	1.2356	1.2311	1.3745	1.2338*	-0.2%
20%	Task2 Book → Video	N@20	0.00245	0.00578	0.01898	0.02066	0.01850	0.01821	0.02073	0.3%
20%	Task2 Book → Video	H@20	0.00043	0.00124	0.0064	0.00676	0.0056	0.00594	0.00697*	3.1%
50%	Task2 Book → Video	MAE	4.1951	1.5651	1.1863	1.0193	1.0077	1.1116	0.9953	1.2%
50%	Task2 Book → Video	RMSE	4.7693	2.0341	1.4993	1.3089	1.3051	1.4425	1.3155	-0.8%
50%	Task2 Book → Video	N@20	0.00274	0.00536	0.01924	0.02041	0.01875	0.01785	0.02047	0.3%
50%	Task2 Book → Video	H@20	0.00044	0.00107	0.00642	0.00675	0.00535	0.00575	0.0068	0.7%
80%	Task2 Book → Video	MAE	4.2384	2.2379	1.3445	1.1469	1.1151	1.2072	1.0846*	2.7%
80%	Task2 Book → Video	RMSE	4.8198	3.1740	1.6946	1.4871	1.4660	1.5968	1.4695	-0.2%
80%	Task2 Book → Video	N@20	0.00258	0.00412	0.01906	0.01949	0.01710	0.01520	0.01960	0.6%
80%	Task2 Book → Video	H@20	0.00040	0.00073	0.00628	0.00636	0.00512	0.00484	0.00634	-0.3%
20%	Task3 Book → Music	MAE	4.5190	1.7976	1.6425	1.3073	1.1945	1.2556	1.1220*	6.1%
20%	Task3 Book → Music	RMSE	5.1838	2.3545	1.9873	1.6599	1.5771	1.6730	1.5390*	2.4%
20%	Task3 Book → Music	N@20	0.00196	0.00383	0.01193	0.01179	0.01367	0.01006	0.01374	0.5%
20%	Task3 Book → Music	H@20	0.00035	0.00071	0.00323	0.00313	0.0037	0.00275	0.00382*	3.2%
50%	Task3 Book → Music	MAE	4.4953	2.0002	1.9364	1.5183	1.3925	1.4304	1.3077*	6.1%
50%	Task3 Book → Music	RMSE	5.1685	2.6001	2.2966	1.9467	1.8644	1.9475	1.8255*	2.1%
50%	Task3 Book → Music	N@20	0.00200	0.00341	0.00994	0.00964	0.01058	0.00804	0.01082*	2.3%
50%	Task3 Book → Music	H@20	0.00028	0.00059	0.00253	0.00247	0.00277	0.00206	0.00281*	1.7%
80%	Task3 Book → Music	MAE	4.5133	2.5014	2.3448	1.8849	1.7107	1.7016	1.5871*	6.7%
80%	Task3 Book → Music	RMSE	5.1960	3.1740	2.7035	2.3517	2.2468	2.3248	2.2110*	1.6%
80%	Task3 Book → Music	N@20	0.00170	0.00275	0.00705	0.00652	0.00658	0.00682	0.00722*	2.3%
80%	Task3 Book → Music	H@20	0.00027	0.00046	0.00176	0.00158	0.00165	0.00107	0.00179*	1.6%

• DiffCDR 在亚马逊数据集上的冷启动和暖启动跨域推荐任务中优于若干基线（CMF、EMCDR、SSCDR、LACDR、PTUPCDR）。
• 消融研究表明 DIM、ALM 与面向任务的学习各自对性能提升有贡献。
• 采用完整 DAT 配置的 DiffCDR 在多任务和多冷启动水平上取得最佳结果，在 MAE、RMSE、N@20、H@20 相较最强基线有显著提升。
• 采用快速 DIM 求解器显著提升推理速度，同时Accuracy损失不大。
• 可视化显示 DiffCDR 将用户因子更连贯地转移到目标域，优于其他方法。

Figure 2. Architecture of DiffCDR. It consists of a Diffusion Module (DIM) and an Alignment Module (ALM). We also employ a diffusion solver to accelerate the inference of the DIM. The recommendation results are produced by element-wisely multiplication with the output of ALM and the item embeddings