从零开始的论文之旅-摘要&介绍部分

最近在写一篇关于多视角回归的英文论文，针对写作过程中的不同问题与心得在此记录一下。

（1）摘要

摘要是一篇论文的核心，也是审稿重点照顾的部分。一个不会犯错在平均线以上水平的论文摘要可以按照十句法书写。

第一句：简要的概括你的研究方向本质（将你的论文工作、研究方向（领域）直接呈现给读者）

正确例句：Multi-view learning focuses on how to improve the learning performance through multiple learners with different views.（多视角学习的目的）

错误：Semi-supervised learning （SSL） lies at the core of machine learning （ML）.（无用的废话，将SSL换成Co-training，deep-learning都行）

第二句：研究方向已有的工作（在论文研究方向的范围内，所完成的工作）

例句1：For semi-supervised regression (SSR), a multi-view learner explores more hidden or less easily observable information presented by unlabeled data.（多视角在半监督回归的工作）

例句2：Existing rules are usually evaluated through two measures, namely support and confidence. （已有的指标）

第三句：现有工作的不足（呈上启下，阐述为什么要做这个方法）

例句1：Howerver, a mount of studies reveals that a few low-quality instances and pseudo labels may decrease the quality of the final regressors.（将未标记数据加入训练集可能引起性能降低）

例句2：However, these two measures may not be enough to describe the strength of a rule. （指标存在缺陷，不能很好的反映特征）

第四句：概述本文的工作

套路句：在本文中，针对什么样的问题，我们提出了什么样的方法。

例句1：In this paper, we propose the multi-view regression through serial updating and safe labeling (MESS) algorithm to face the such fundamental problem.

例句2：In this paper, we introduce granular association rules with four measures to reveal connections between granules in two universes, and propose three algorithms for rule mining.

第五到八句：向读者描述你算法的优势，具体做法，算法框架等

例1：

First, the serial updating technique is employed the training set of multiple regressors. 
This strategy is fully comply with safe regression framework and leads better performance than traditional parallel methods.
Second, a new safe regression form is designed to improve the label quality. 
This new method regards the trusty confident pseudo labels as the base line and can improve the quality of pseudo labels without side effects.（描述优势）

例2：

With this approach, all possible itemsets are divided into three regions, namely the positive, the boundary and the negative region.     
Positivein the positive region are already frequent.             Boundaryin the boundary region are infrequent, however may be frequent after data increment in the near future.       
Negative regionin the negative region will not be frequent even after data increment.       
Therefore to keep the frequent itemsets up-to-date, one only need to check those in the boundary region, and the runtime is saved.(算法细节)

第九到十句：实验结果，简要的概述你的实验最终结果（自己算法在多少个数据上比别人好，超过了什么算法）

例1：

Experiments were undertaken on ten regression data sets. 

 The results show that MESS is superior to state-of-the-art semi-supervised regressors.

例2：

Experiments on two real world datasets show that the approaches dedicated to three subtypes are 2-3 orders of magnitudes faster than the one for the general case.

总体摘要如下：

Multi-view learning focuses on how to improve the learning performance through multiple learners with different views.
      For semi-supervised regression (SSR), a multi-view learner explores more hidden or less easily observable information presented by unlabeled data.
      Howerver, a mount of studies reveals that a few low-quality instances and pseudo labels may decrease the quality of the final regressors.
      In this paper, we propose the multi-view regression through serial updating and safe labeling (MESS) algorithm to face the such fundamental problem.
      First, the serial updating technique is employed the training set of multiple regressors. 
      This strategy is fully comply with safe regression framework and leads better performance than traditional parallel methods.
      Second, a new safe regression form is designed to improve the label quality. 
      This new method regards the trusty confident pseudo labels as the base line and can improve the quality of pseudo labels without side effects.
      Experiments were undertaken on ten regression data sets. 
      The results show that MESS is superior to state-of-the-art semi-supervised regressors.

（2）介绍

介绍是对于摘要的扩充，一个合格的介绍每一段应该是对摘要部分的扩充。

part1：研究方向本质，以及所选研究领域的优势（比摘要部分要详细）

Semi-supervised learning (SSL) aims to improve the learning performance with the help of additional unlabeled data.
It learns more useful information through considering the data distribution presented by labeled and unlabeled data.
Therefore, the learning performance is fully related to the quality of additional unlabeled data. （SSL的本质与摘要类似）      
In the field of regression, multi-view learning strategy can cope efficiently with the aforementioned phenomena.
It explores unlabeled data with the redundant information learned from multiple different views.（multi-view的本质和做法）
The pseudo labels of different learners can be integrated to provide richer information.
In this way, the multi-view learners can discriminate more hidden insights upon unlabeled instances than single view.
Such a learning method has received much attention owing to the immense data demands in real-world applications.（这三句阐述multi-view的好处，为part2做铺垫）

part2：part1方法虽好，但是存在问题（不从问题角度来说，可以从自己的受到什么启发角度出发）

 Despite the advantages of multi-view learning and SSL, the unlabeled data remain in training set still might decrease learning performance.（阐述可能存在的错误）
In the training process, the selection of some unlabeled instances or their pseudo lalbels may be inappropriate.
These inappropriate instances and labels will decrease the learning performance and not be corrected in the future training process. （对错误进一步延展）
It is highly desirable to reduce the accumulation of such instances and labels.（定性为接下来自己的说法铺垫）

part3：摘要第四到八句，对算法具体描述

In this paper, we propose the multi-view regression through serial updating and safe labeling (MESS) algorithm to face the above problem.（摘要第4句，介绍中的更加更加详细）
First, a serial updating manner is employed to update the training set of different regressors.
Figure \ref{fig:SPOR} illustrates our serial updating manner.
In each iteration, all regressor will evaluate sampled unlabeled instances $\mathbf{U}$ and select the most confident instance among in most views.
Then, the common confident instance will be included in the training set of the current view.
This updating manner may guarantee the reliability of selected confident instances than parallel way.（对于第五六句的详细补充，详细到具体操作）

     
Second, a safe multi-view regression framework is designed to improve the reliability through pseudo-label ensemble.
Safe learning is a specially SSL form that can often improve the learning performance without side effects.
In MESS, we regard the semi-supervised prediction of selected confident instance in current view better than its supervised prediction.
In this way, the pseudo labels of selected instances will be considered as the base line.
Each instance will learn a safer pseudo label from other views before being included in the training set.（对应摘要第7至8句，详细到具体操作）
Through above improvements, the regressors will obtain a more stable performance using the safe training set.（阐述优势）

part4：实验部分（这部分对应摘要8，9句，可以详细描述实验数据的性质，指标，对比算法之类）

Experiments were undertaken on 10 data sets with 3100--9600 instances and 8--21 attributes.
These data sets include health (parkinsons), physical measurements (abalone), engineering (folds5x2\_pp), etc.（数据性质）
We compared the MESS algorithm with state-of-the-art semi-supervised regression (SSR) algorithms.
The pairwise $t$-test with 0.05 significance level were used to verify the significance of the performance differences between MESS and the other algorithms.（指标）
The results showed that MESS outperforms all of these competing algorithms in terms of mean squared error.
The implementation of MESS is available at in which all source code is accessible.（结果说明）

part5：固定部分（介绍论文结构相当于一个目录作用）

The rest of the paper is organized as follows:
Section \ref{Section: Related Work} introduces the related work in SSL fields.
Section \ref{Section：Model} presents our algorithm framework and key issues in learning strategy.
Section \ref{section: Experiments Results} describes the experimental procedure and lists some results.
Finally, Section \ref{section:Conclusion and Future Work} presents conclusions and outlines further research trends.