Federated recommendation is a typical application of federated learning, which can protect the privacy of users by exchanging models between users’ devices and central servers rather than users’ raw data. Recently, although some research in federated recommendation has made remarkable progress, there are still two major issues need to be addressed further due to the non-independent and identical distribution (Non-IID) data which is very common in federal recommendation systems. First, the communication load of the user device during training is heavy. Second, the trained local model lacks personalization. Aiming at the above problems, a federated recommendation algorithm based on user clustering and meta-learning, ClusterFedMet, is proposed to improve communication efficiency and recommendation personalization simultaneously. In ClusterFedMet, users are clustered into different clusters according to their data distribution, and user sampling are performed based on the clustering result, thus reduce harmful interference among users with different data distribution. The model is trained with meta-learning, which can generate more personalized local models. During meta-learning, a controller which can dynamically tune the hyperparameters for users is designed to achieve better performance. According to weights, gradients, and losses of each step, the controller can find a learning rate suitable for each user’s local data and model. We perform evaluations for the proposed algorithm on two public datasets, and the results demonstrate that our algorithm outperforms other advanced methods in terms of recommendation accuracy and communication efficiency.

中文翻译：

---

一种基于用户聚类和元学习的联合推荐算法

联邦推荐是联邦学习的典型应用，它通过在用户设备和中央服务器之间交换模型而不是用户的原始数据来保护用户的隐私。近年来，尽管联邦推荐方面的一些研究取得了显着进展，但由于联邦推荐系统中非常常见的非独立同分布（Non-IID）数据，仍然存在两个主要问题需要进一步解决。首先，训练时用户设备的通信负载较大。其次，经过训练的本地模型缺乏个性化。针对上述问题，提出一种基于用户聚类和元学习的联邦推荐算法ClusterFedMet，以同时提高通信效率和推荐个性化。 ClusterFedMet中，根据用户的数据分布将用户聚类到不同的簇中，并根据聚类结果进行用户采样，从而减少不同数据分布的用户之间的有害干扰。该模型通过元学习进行训练，可以生成更加个性化的本地模型。在元学习过程中，设计了一个可以为用户动态调整超参数的控制器，以实现更好的性能。根据每一步的权重、梯度和损失，控制器可以找到适合每个用户本地数据和模型的学习率。我们在两个公共数据集上对所提出的算法进行了评估，结果表明我们的算法在推荐准确性和通信效率方面优于其他先进方法。