SCI一区：IEEE TSMC老师2021年2月修改稿反馈

引入

  写好论文的初稿给老师看，当时甩在大群里的，包括老师在内的其他老师都说可以，当时就飘了。
  第二天老师开始细看，反馈回来的东西就和自己的很不一样，/xk
  这里记录下来，加以反思，希望对看到这里的伙伴有所帮助。
  注：批语是老师当时写下的；说明是自己回忆老师说的或者自己想的。

1 摘要部分

  1）简要地对多示例进行介绍：

原句：The processing objects of multi-instance learning (MIL) are labeled bags. One bag corresponds to one sample, and the label is given at the bag level.
批语：能写成这样已经比较清楚了，但没力量，没能把问题的重要性强调出来。
修改：Multi-instance learning (MIL) is more general and challenging than traditional supervised learning in that labels are given at the bag level.
思考：增加了与传统有监督学习的对比，这里也表明了自己的场景，同时对多示例的特征进行了简要说明。

  2）已有的多示例处理方式：

原句：By performing feature mapping, MIL can be converted to a single-instance leaning problem and be solved with traditional supervised learning.
批语：避免词语的重复
修改：The popular feature mapping approach converts each bag into an instance in a new feature space.
思考：词语的重复是一个经常遇见的问题，当然这一句要和上一句结合起来看。

  3）承上启下：

原句：However, most existing mapping methods cannot maintain the distinguishability of bags in the new feature space and consider the self-reinforcement ability of model.
批语：这里self-reinforcement ability of model很突兀，读者一点点准备都没有
修改：However, most existing mapping methods cannot maintain the distinguishability and self-reinforcement ability of bags.
思考：在这篇文章中，针对可区分性和自我学习的机制是两个亮点。所对比的六个方法中，严格意义上只有一个对可区分性进行了说明。所有的算法均是模型训练–>测试的过程。
最开始时，我觉得保留自我学习能力不太合适，不过仔细想想，这两个技术是本质上都是包的选择技术，作用于包上。
再次修改：maintain the distinguishability of bags, and the MIL model does not support self-reinforcement.
思考：这种是逻辑错误，如果不是很了解本方向的话确实不容易看出来。

  4）本文干嘛：

原句：we propose the multi-instance ensemble leaning
批语：低级拼写错误，leaning -> learning，不止一次！
思考：这种真的就是要靠细心，特别是latex中，有的单词虽然写的和自己想要的不一样，但是又没有拼写错误。

  5）辨别性分析技术是什么：

原句：First, discriminative analysis optimizes the bag selection process by considering both the spatial and label distribution of data.
批语：discriminative analysis 作为主语不合适
修改：First, considering the space and label distribution of the data, the bag selection process is optimized through discriminant analysis.
思考：考虑什么？什么怎么样？句式记下。

  6）自增强技术怎么做？

原句：Second, self-reinforcement mechanism obtains a more distinguishability dBagSet by designing the state and action transfer strategy.
批语：要么用the … mechanism这种有定冠词的固定搭配，要么都不用。
修改：Second, with the state and action transfer strategy, a more distinguishability dBagSet is obtained through self-reinforcement.
思考：啊啊啊~~~读起来的感觉就不一样了。

  7）模型集成：

原句：The classifier ensemble technique constructs a series of weighted models with these dBagSets, and then obtains the final prediction by weighting sum.
批语：weight出现了两次，而且这里既涉及到classifier building，又涉及到ensemble，所以说成classifier ensemble technique不合适。
其实这里还应该把如何加权写清楚，否则没特色。
修改：The ensemble technique constructs a series of classifiers with these dBagSets, and then obtains the final prediction by weighted sum.

  8）一句话带过实验：

原句：Experiments show that our algorithm outperforms the state-of-the-art MIL mapping solutions in general.
修改：Experimental results show that ELDB is superior to the
state-of-the-art MIL mapping solutions.
思考：其实诸如这样的总结性句子有很多风格，并没有说一定要用哪种，这里是为了把in general去掉，哈哈。

2 引入部分

  1）与传统SIL的比较，引出MIL：

原句：Compared with traditional single-instance learning (SIL), the processing data of multi-instance learning (MIL)　learner is a set of bags with bag-level label instead of a set of instances with instance-level label.
批语：learning learner，这种重复不靠谱
label应该用复数
其实这句话的词语重复率还是高！
修改：Compared with traditional single-instance learning (SIL), the processing data of multi-instance learning (MIL) is a set of bags with bag-level labels instead of a set of instances with instance-level labels.
思考：降重复还是不简单哈，这里的instance就有四个/xk

  2）标准多示例假设：

原句：Moreover, a bag is positive if it contains at least one positive instance, otherwise it is negative.
批语：这里不是Moreover的逻辑，而是指出了应用中的初衷
修改：In applications, a bag is labeled as positive if it contains at least one positive instance, otherwise it is negative.

  3）MIL是干嘛：

原句：The task of MIL is to train a classifier by processing the training set, and then label the testing bag.
批语：写得没营养，谁不是训练/测试呢？
修改：Accordingly, the task of MIL is to classify bags instead of individual instances.
思考：MIL与传统有监督学习的区别在于所处理的数据结构，一个是包，一个则是实例。
对于传统的来说，分类、回归、聚类等都是常规操作，MIL也是能够搞的。
哈哈

  4）例如…等等：

原句：such as … and so on.
思考：例如就是在举例了哈，例如我和你喜欢吃苹果，其实没有必要再来什么等等哈。

  5）就问你饶不饶：

原句：find or use a method to directly measure the distance or similarity between bags
说明：特别是在引入里，不要把重点引入别人做法的细节上，你这里是为了给别人说清楚，距离就距离，相似性就相似性，别人什么几个OR
修改：measure the distance between bags

  6）说了大类型，想引入一下小类型：

原句：This work follows the last way with the following subcategories.
说明：由远及近是论文写作中的一种常用方式，上一段已经介绍了大方向，这里其实只需要很自然的过渡。
修改：The last category can be further divided into the following
subcategories.

  7）想说说别人的不足，从而引出自己的工作：

原句：However, most existing mapping approaches cannot maintain the distinguishability of bags in the new feature space.
批语：However作为转折词，前面的句子一般是赞扬，后面则表示批评。你这里，前面没有赞扬。
修改：Unfortunately, most existing mapping approaches cannot maintain the distinguishability of bags in the new feature space.

  8）提出疑问：

原句：First, how to maintain or even improve the distinguishability of bags in the new feature space by using the spatial distribution and label information of data. Second, how the model acquires the self-reinforcement ability.
说明：这里本身就是疑问的形式，直接？打底的话，可以更有气势。进一步，疑问的话：1）怎样写好论文？2）怎样通过A、B和C写好论文？后者加以限定，前者则把重心放在怎么写好上。
修改：a) How to improve the distinguishability of bags in the new feature space? and b) How does the model acquire the self-reinforcement ability?

  9）开局一张图：

原句：Figure 1 compares our algorithm with the traditional bag-based mapping (TBBM) method using an example.
说明：图的本身就是一种说明；少用我们的算法什么的，既然都有了算法名，就不要让它在家里孵蛋蛋；当然要秀肌肉的时候另当别论。
修改：Figure 1 compares ELDB with the traditional bag-based mapping (TBBM) algorithm.

  10）开局一张图的描述：

原句：First…Second
说明：以上是一种描述形式，a)… and n)是另一种。想想同一个意思的句子如何写成两种句子甚至更多，这些都是需要注意的。
再次修改：ELDB introduces a self-reinforcement mechanism into TBBM. -> b) ELDB introduces a self-reinforcement mechanism to learn and update the existing dBagSet;
说明：修改前：A引入B到C，相当于给C打广告；修改后：A引入什么干嘛，这些都是你自己的贡献。

  11）因此，我们有以下优势：

原句：Consequently, ELDB has the following advantages.
说明：Consequently的结论性很强，能上段推出目前吗？
修改：Our proposed ELDB has the following advantages.、

  12）与A相比我们怎样？：

原句：ELDB may be easier to distinguish than the TBBM
说明：可区分的能力，到底是模型的呢？还是它有这样的能力？
修改：ELDB has better distinguishing ability than TBBM

  13）多个模型的好处：

原句：Third, the label of a bag will no longer be determined by a single model, but will be predicted by the weighted sum of multiple models.
说明：我不是不喜欢你，只是我们不合适，这里不是男女朋友谈恋爱哈，直接说不香吗？
修改：Third, the label of the bag is predicted by the weighted sum of multiple models to obtain better stability.、

  14）和非辨别性算法比：

原句：Comparison with non-discriminative methods: In order to ensure the distinguishability of bags in the new feature space, the spatial distribution and label distribution are analyzed theoretically.
说明：
修改：Compared with the state-of-the-art map-based methods, the mapping vectors of bags generated by ELDB have better distinguishability in terms of theoretical analysis;

3 相关工作

  1）总领句子：

原句：Our work is directly related to three directions of research.
说明：多收集相关句子。
修改：Three lines of researches are directly related to our work.

  2）核心是什么：

原句：The core is how to choose some instances, and then use or design a mapping function to convert bags into a new feature space.
说明：现在回过头去看以前的句子，确认有够绕的。
修改：The core part is to design a mapping function based on selected instances and transform bags into the new feature space.

  3）依托别人的做法，萌生了现有的点子：

原句：We referred to the core ideal of this method…
说明：说是参考别人，便弱化了自己的功劳。
修改：We borrow the core idea of this method…

  4）基于什么做什么：

原句：Once we have a series of dBagSet, we can construct mapping function as follows
说明：笑死，哈哈。确实是很中式的翻译了：
修改：With a dBagSet, the mapping function is constructed as follows

4 算法部分

  1）参数介绍：

原句：${\lambda }_{ij}$ is a scale parameter, and $\forall i,j:{\lambda }_{ij}>0$
说明：能一句话说完就别两句话
修改：${\lambda }_{ij}>0$ is a scale parameter,

  2）分析和讨论怎么写？

说明：算法提出，有哪些关键的技术，这些都是讲你怎么做？而分析和讨论，则是对你为什么这么做进行说明。

5 实验部分

  1）提出的一些问题：

原句：How sensitive is the parameters of ELDB?
说明：当时写完都没有很在意这个句子的意思是否正确，哈哈
修改：How sensitive is the ELDB algorithm to parameter settings?

  2）这类方法的做法是：

原句：This types of algorithms focus on designing or employing a kernel, such as Gaussian kernel, Fisher kernel or Isolation kernel.
说明：我真是爱把别人到底怎么做的说明一下；一般而言，A专注于设计B干嘛，才算是完整的说明了。
修改：Gaussian, Fisher and Isolation kernels are often employed for this type of algorithms.

  3）它用什么做什么：

原句：Its procedure is to use the Gaussian mixture model (GMM) to extract…
说明：再次强调一哈，它的步骤是用什么做什么->就是用什么做什么，完毕。
修改：It uses Gaussian mixture model (GMM) to…

  4）参数的设置：

原句：the number of component for GMM was set to [1, 2, 3]
说明：这个参数想说明的意思是从[1, 2, 3]中枚举，那这样写不就行了？
修改：the number of components for GMM is enumerated in…

  5）实验说明：

原句：Tables 3 and 4 report the classification experimental results of ELDM and four types of algorithms.
修改：Tables 3-4 compare the performance ELDB with these four types of algorithms.