With the rapid development of storage and computing power on mobile devices,
it becomes critical and popular to deploy models on devices to save onerous
communication latencies and to capture real-time features. While quite a lot of
works have explored to facilitate on-device learning and inference, most of
them focus on dealing with response delay or privacy protection. Little has
been done to model the collaboration between the device and the cloud modeling
and benefit both sides jointly. To bridge this gap, we are among the first
attempts to study the Device-Cloud Collaborative Learning (DCCL) framework.
Specifically, we propose a novel MetaPatch learning approach on the device side
to efficiently achieve "thousands of people with thousands of models" given a
centralized cloud model. Then, with billions of updated personalized device
models, we propose a "model-over-models" distillation algorithm, namely
MoMoDistill, to update the centralized cloud model. Our extensive experiments
over a range of datasets with different settings demonstrate the effectiveness
of such collaboration on both cloud and devices, especially its superiority to
model long-tailed users.

本文提出了一种名为 “Device-Cloud Collaborative Learning” 的框架，运用 “MetaPatch 学习” 及 “MoMoDistill” 算法实现了设备端和云端之间的合作学习，取得了较好的效果。

设备 - 云协作学习推荐

Device-Cloud Collaborative Learning for Recommendation

In this paper, we try to predict the winning team of a match in the
multiplayer eSports game Dota 2. To address the weaknesses of previous work, we
consider more aspects of prior (pre-match) features from individual players'
match history, as well as real-time (during-match) features at each minute as
the match progresses. We use logistic regression, the proposed Attribute
Sequence Model, and their combinations as the prediction models. In a dataset
of 78362 matches where 20631 matches contain replay data, our experiments show
that adding more aspects of prior features improves accuracy from 58.69% to
71.49%, and introducing real-time features achieves up to 93.73% accuracy when
predicting at the 40th minute.

本文提出了多种模型，使用赛前个人数据和实时数据对多人电子竞技游戏 Dota 2 中比赛获胜队伍进行预测，其中包括逻辑回归，属性序列模型，以及它们的组合，实验表明，添加赛前和实时的特征信息可以将精度从 58.69% 提高至 71.49% 和 93.73%。