由买买提看人间百态

topics

全部话题 - 话题: accuracy
首页 上页 1 2 3 4 5 6 7 8 9 10 下页 末页 (共10页)
s****l
发帖数: 10462
1
来自主题: Biology版 - Oxford nanopore 要出USB sequencer了
看来年底之前,是骡子是马就差不多知晓了。
ILlumina是不是有15%的股份,还有 right to negotiate until 2016?
如果accuracy不错的话,Ion Torrent比较麻烦 --- 快速,价格都比不上,accuracy又
不高很多。
nanopore sample preparation 似乎很简单,才6个eppendorf管子就解决了。是不是我
看错了?
s****l
发帖数: 10462
2
来自主题: Biology版 - Oxford nanopore 要出USB sequencer了
看来年底之前,是骡子是马就差不多知晓了。
ILlumina是不是有15%的股份,还有 right to negotiate until 2016?
如果accuracy不错的话,Ion Torrent比较麻烦 --- 快速,价格都比不上,accuracy又
不高很多。
nanopore sample preparation 似乎很简单,才6个eppendorf管子就解决了。是不是我
看错了?
G***G
发帖数: 16778
3
来自主题: Biology版 - 3 classes
for 2 classes classification, we have specificity and sensitivity to
evalute the prediction performance.
for a 3 classes classification, what methods can measure the prediciton
performance in addition to the accuracy?
for example
class1 class2 class3
class1 10 2 2
class2 0 5 1
class3 0 2 6
The accuracy is (10+5+6)/(10+2+2+5+1+2+6).
any other measurement we can use?
s****l
发帖数: 10462
4
MinIon的数据已经出来了一些了。
大概看了下,总体上讲,还有很远的路要走。能不能真正进入市场还很难说。
single read accuracy只有60~70%,相当坑爹,但是可以长达好几千甚至好几万个碱
基。大部分有效的在几千的水平吧。
有的人说error is biased,也就是说很多reads都会在同一区域出错,这样一来,high
depth coverage也不太能帮上忙。但是也有的说,error is not biased,那么多测若
干遍也就能提高consensus accuracy了。
最大的好处就是便携和便宜了
i*******i
发帖数: 145
5
这个不光是技术上的问题。
“Single Molecule, Real-Time (SMRT®) DNA sequencing achieves highly
accurate sequencing results, exceeding 99.999% (Q50) accuracy, regardless of
the DNA’s sequence context or GC content. - http://blog.pacificbiosciences.com/2013/01/a-closer-look-at-accuracy-in-pacbio.html
首先这个说明pacbio的准确率足够绝大多数的分析。
“The stats are quite fascinating. They have 66 SMRT cells producing 32,559,
803,198 bases of post-filtered nucleotides. - http://www.homolog.us/blogs/blog/2013/10/22/pacbio-releases-... 阅读全帖
s****9
发帖数: 932
6
来自主题: Biology版 - 牛皮终于吹破了 (转载)
即使有breakthrough,第一步也只能在Cost,accuracy/reproducibility,sample
volume中选择2个。当她们选择了low volume和低cost时,必定以accuracy作为代价。
Medical diagnostics需要的technology advancement绝不是一个辍学的本科生拍拍脑
袋就可以完成的。这个和facebook需要的技术含量完全不在一个数量级上。
s****9
发帖数: 932
7
来自主题: Biology版 - 牛皮终于吹破了 (转载)
即使有breakthrough,第一步也只能在Cost,accuracy/reproducibility,sample
volume中选择2个。当她们选择了low volume和低cost时,必定以accuracy作为代价。
Medical diagnostics需要的technology advancement绝不是一个辍学的本科生拍拍脑
袋就可以完成的。这个和facebook需要的技术含量完全不在一个数量级上。
x****6
发帖数: 4339
8
By Christopher Weaver,
John Carreyrou and
Michael Siconolfi
April 18, 2016 6:37 p.m. ET
296 COMMENTS
Federal prosecutors have launched a criminal investigation into whether
Theranos Inc. misled investors about the state of its technology and
operations, according to people familiar with the matter.
Walgreens Boots Alliance Inc. and the New York State Department of Health
have received subpoenas in recent weeks seeking documents and testimony
about representations made to them by the Palo Alto, C... 阅读全帖

发帖数: 1
9
导语:回观国内市场,这些不靠谱、不合医学逻辑和没有临床价值的事情反而愈演愈烈
,耗费了大量的资本和医疗资源,可谓几大乱象
本文作者代号HHH,目前为一家人工智能医疗创业公司的CEO,有10多年硅谷的高科技公
司的工作经历(Google, Cisco等),斯坦福商学院Sloan Fellow。
业内人的“针砭时弊”:AI医疗界的3大乱象以及如何评价医学人工智能的成果?
With machine learning situated at the peak of inflated expectations, we
can soften a subsequent crash into a “trough of disillusionment” by
fostering a stronger appreciation of the technology’s capabilities and
limitations.
今年6月底的《新英格兰》刊出了一篇观点文章《Machine Learning and Prediction
in Medicine — Beyond the Peak of ... 阅读全帖
k********s
发帖数: 320
10
来自主题: Chemistry版 - 问一个qtof的问题
Q-tof accuracy is very temperature dependent. For good accuracy you need to
use lockspray or at least adjust Veff using Leucine Enkephalin on daily
basis.
k********s
发帖数: 320
11
来自主题: Chemistry版 - 问一个qtof的问题
1. If you use external lockmass (lockspray)- which gives you better accuracy
than internal lockmass, then you don't need to recalibrate.
2. If you don't use lockmass but rather use Leu or other reference standards
to adjust Veff, and if the change of Veff is very small, then you don't
need to recalibrate - according to Water's engineer. However, based on my
experience, fresh calibration does give you improved accuracy - not dramatic though.
3. What is your RMS from calibration?
4. What's the acc
p**********m
发帖数: 472
12
来自主题: Chemistry版 - 问一个qtof的问题

accuracy
standards
dramatic though.
i am looking for smaller than 5ppm.
after temperature change, i hardly can get 5ppm accuracy even with lockmass.
will it be harm to qtof if temperature change dramatically on daily basis.
withine 5oC range.
b*******g
发帖数: 1309
13
来自主题: Chemistry版 - Quadrupole TOF MS有啥特别的吗?
I bet what you mentioned is ESI-Q-TOF,
ESI-Q-TOF is similar as ESI-QqQ
TOF used here instead of Q is to improve mass accuracy.
QToF is much cheaper than FT MS, but the mass accuracy is also good (1-5ppm)
, and the scan speed is very fast, so it become more and more popular.
With the MALDI source, there is instrument called MALDI-ion trap-TOF,
produced by Shimadzu. Ion trap is used to select and fragment ions, then you
can do MS/MS or more.
There are also some home-made MALDI-quadrupole-TOF instr... 阅读全帖
k********s
发帖数: 320
14
来自主题: Chemistry版 - Possible identity of a compound?
Synapt accuracy is enough. Orbi won't help you much with accuracy
improvement over Synapt. Paste UV spectrum, MS/MS and also accurate mass
spectrum here so people can help you better.
What's the collision energy you were using? - Please also provide system
information, waters? Thermo? what model?
Do you have a GC-MS with Direct Probe (DI) capability?

not
k********s
发帖数: 320
15
来自主题: Chemistry版 - Possible identity of a compound?
Synapt accuracy is enough. Orbi won't help you much with accuracy
improvement over Synapt. Paste UV spectrum, MS/MS and also accurate mass
spectrum here so people can help you better.
What's the collision energy you were using? - Please also provide system
information, waters? Thermo? what model?
Do you have a GC-MS with Direct Probe (DI) capability?

not
j***h
发帖数: 4412
16
Earthquake detection systems can sound the alarm in the moments before a
big tremor strikes—time enough to save lives
Japan’s system, which went live in 2007, makes heavy use of personal
technology. Alerts go out not only on television and radio but through
special receivers in homes, offices and schools. Pop-up windows on
computers show a real-time map with the epicenter’s location and the
radiating seismic waves. A timer counts down to the shaking at your
location and highlights predicted inte... 阅读全帖
h***o
发帖数: 539
17
来自主题: Computation版 - 问个专业问题,请帮助看看
f(x+dx) - f(x) = SIGMA{f_n(x)dx^n/n!}, n from 1 to infinity
if f_n(x)dx^n/n! > 0 for any n, higher order calculation has better
accuracy.
otherwise, it's hard to say.
However, if you increase the the order significantly, it's relatively safe
to say that you also get better accuracy.
b*****y
发帖数: 163
18
来自主题: Computation版 - let me ask a question again

They talked about solving a system of ODEs which consists of state variables
with different magnitudes. If you want to solve the system with one tolerance,
you have to make sure every equation achieve that accuracy. 10e-06 means
the tolerance (eps here). The scaling part comes in because some state
variables may not need such a high accuracy as 1e-06, so you use scaling
vector to make the magnitudes of state variables close to each other.
To clarify the terminology, in double precision under IE
i******m
发帖数: 10
19
来自主题: Computation版 - KDV type PDE
The FD scheme in the (N. J. Zabusky and M. D. Kruskal, Phys. Rev. Lett. 15 (
1965), 240) is a classic one. Fast and accurate. You can try this. Highly
recommend.
But if you need more accuracy, you can try the spectral method. It has "
infinite" accuracy, and no phase error. However, you may only impose the
periodic BC for this method.
w*****f
发帖数: 61
20
【 以下文字转载自 EE 讨论区 】
发信人: websurf (天天冲浪), 信区: EE
标 题: 求助: about leave-one-out cross validation
发信站: BBS 未名空间站 (Thu Feb 21 10:34:05 2008)
I am using LOOCV to estimate the classification accuracy because of small
sample size.
There are 3 classes. The sample size of these classes are [ 10 12 15]. I
used LOOCV on 10+12+15=37 samples.
My question is will the unbalanced sample size affect the estimation of
classifcation accuracy. Do I need to change the prior possiblity in the
classifier?
谢谢大牛们指点
m****i
发帖数: 577
21
来自主题: Computation版 - 包子请帮忙,关于python
文科PHD.选了一门PYTHON课。。。从来没有code过。
第一次作业,真的一头雾水。
从"C“步就不会了。。。
有人帮忙指点怎么写吗?
有视频讲类似的问题吗?
包子感谢。
Coding problem: Write a self-contained & fully functional Python script
that:
a. Loads the necessary libraries to run (e.g., scikit-learn)
b. Loads the dataset “digits” (see below for description) from the
scikit-learn example datasets contained in sklearn.datasets
c. Instantiates a DecisionTreeClassifier
d. Splits the dataset in training and testing sets using KFold Cross
Validation (use K=10)... 阅读全帖
s**g
发帖数: 66
22
Wien-Bridge has decent distortion performance. But the amplitude control may
not be as good as you like to see.
DDS uses high speed DAC with built-in look-up table to generate waveform (
sine/triangular/square). The frequency accuracy is realized by internal
oscillator or PLL referenced to an external crystal; amplitude accuracy is
obtained by the DAC.
Add LPF or BPF to suppress distortion and noise (to meet 1mVrms spec). Add
buffer if you like to scale up swing or to drive considerable load or
g****t
发帖数: 31659
23
请问为啥Wien-Bridge类型的振幅控制会不好?
我觉得精度不够就是二极管工作区的曲线太陡或者太不陡,
把Wien-Bridge里的二极管接个可调电阻,
然后调一下说不定精度就上去了吧?

Wien-Bridge has decent distortion performance. But the amplitude control may
not be as good as you like to see.
DDS uses high speed DAC with built-in look-up table to generate waveform (
sine/triangular/square). The frequency accuracy is realized by internal
oscillator or PLL referenced to an external crystal; amplitude accuracy is
obtained by the DAC.
Add LPF or BPF to suppress distortion and noise
S*******p
发帖数: 261
24
Please contact Shaz Samual directly!
FYI:
One of our clients is the global leading technology provider to the
international energy industry with the world's most advanced intelligent
metering, data collection and utility software solutions. It is their vision
to provide responsible management and conservation of the world's most
valuable resources requiring optimization of the delivery and use of energy
and water. Their Systems Engineering group is rapidly expanding and we need
high impact indi... 阅读全帖
K**********n
发帖数: 1197
25
来自主题: EE版 - 版上有人懂这个技术没?
学术界珍稀有限的原版外文资料,
Remote Mind Control Technology

Reprinted from SECRET AND SUPPRESSED: BANNED IDEAS AND HIDDEN
HISTORY, edited by Jim Keith, $12.95, available from
1-800-680-INET.
There had been an ongoing controversy over health effects of electromagnetic
fields (EMF) for years (e.g., extremely low frequency radiation and the
Navy's Project Seafarer; emissions of high power lines and video display
terminals; radar and other military and industrial sources of radio
frequencies and micr... 阅读全帖
a******e
发帖数: 80
26
想购买一款compact的flip chip bonder, 但因为以前没用过flip chip bonding
technology,不知道该如何挑选。不知道业界内有什么好的品牌,求建议
1. 大型设备就算了,希望体积不要太大,compact一些
2. 希望最好能同时支持wafer-to-wafer packaging和die-level packaging
3. 我们主要的使用的wafer的尺寸是6 inch和8 inch
4. 关于chip pad上的bump,希望能支持solder bump和gold bump
5. 最好bonder能同时完成underfilling,而不用买额外的设备了
6. 至于placement accuracy,我心里还没有具体的要求。silicon die上的IO pad一般
是100um X 100um的,两个pad之间的距离(边缘到边缘)大概也有100um。这种情况是
不是+/- 5um的accuracy就足够了?
非常感谢。
c**********e
发帖数: 694
27
来自主题: Pharmaceutical版 - 工作机会(LCMS, Toxicology)
Position Summary:
The Certifying Scientist (CS) is a scientist who operates in a
multidisciplinary environment in which skills and knowledge in a variety of
subject areas relating to drug testing are required. The primary
responsibilities of the CS are to ensure the accuracy and integrity of the
drug testing process and to subsequently evaluate these test results before
the release of the test reports to physicians or other health care
professionals.
Responsibilities:
•Operation of LCMS fo... 阅读全帖
J****S
发帖数: 15
28
Pls contact: d******[email protected] for more information
Requirements:
1、BS in Clinical Medicine/Basic Medicine from top universities
2、Overseas Ph.D in lifescience(no clincal experiences need) / or have
clinical trial management experiences(no further degree needed)
Responsibilities:
1、Work as a CTL on Medium to High complexity studies including multi-
country / multi-centre trials and lead the Clinical Trial Team (CTT) in
development of the protocol and related documents.
*CTL will involve the... 阅读全帖
R*k
发帖数: 264
29
来自主题: Physics版 - 文章出来了:arXiv:1109.4897
http://arxiv.org/abs/1109.4897
这个网页现在很难打开,估计很多人在看,呵呵
The OPERA neutrino experiment at the underground Gran Sasso Laboratory has m
easured the
velocity of neutrinos from the CERN CNGS beam over a baseline of about 730 k
m with much higher
accuracy than previous studies conducted with accelerator neutrinos. The mea
surement is based on highstatistics
data taken by OPERA in the years 2009, 2010 and 2011. Dedicated upgrades of
the CNGS
timing system and of the OPERA detector, as well as a high preci... 阅读全帖
C*******M
发帖数: 153
30
来自主题: Statistics版 - How to use Theil's U2 inequality
The following table shows the number of actual and expected deaths in the
last ten years:
Actual: 12 15 18 21 24 27 28 25 23 18
Expected: 20 22 24 26 27 26 24 22 19 16
I am wondering if I can use Theil's U2 inequality for accuracy the above
expected value. The formula is attached and I am confused about the formula.
There are several accuracy measurements: Mean Absolute Deviation (MAD), Root
Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Theil's
Measure of Inequality (U2), Co
n*****s
发帖数: 10232
31
来自主题: Statistics版 - how do you deal with sparse data?
这个post作者介绍用bootstrap方法,缺点之一是这样就没有一个单一的model。看到下
面一个人的回复,我感觉更感兴趣
“I’ve built models with 90/10, 95/5 or worse without resampling with good
success whether using logistic regression, neural networks, or some kinds of
trees. The key is thresholding the posterior probability estimate from the
model at the level of the a priori probability (if you want to compute
classification accuracy or use confusion matrices). ”
我好像没用过这个prior/post probability的方法(是bayesian?)。这里说的
classification accuracy和confusion
c*****l
发帖数: 297
32
来自主题: Statistics版 - 哪位用R做过CART MODEL
it will have over-fitting problem, you will to use 9 -fold cross
validation test accuracy or 5 fold cross validation test or leave one
out validation test accuracy
s*****9
发帖数: 285
33
来自主题: Statistics版 - DC Entry-level SAS PROGRAMMER PHARM TRAINER
ABBOTT钱不是很多,CONTRACT2HIRE,2轮PHONE SCREEN,第一轮我面过,感觉还行。
Clinical Statistician I – B ABLJP00027218
Abbott Labs
Abbott Park, IL 60064
8 – 12/2011
Experience in database management (preferably claims and other healthcare
databases), SAS programming and analysis skills are also a plus. Understands
and appropriately uses advanced epidemiological research techniques: •
;Understands primary and secondary research techniques as well as the
elements of more complex quantitative research methods, s... 阅读全帖
s*****9
发帖数: 285
34
来自主题: Statistics版 - DC Entry-level SAS PROGRAMMER PHARM TRAINER
ABBOTT钱不是很多,CONTRACT2HIRE,2轮PHONE SCREEN,第一轮我面过,感觉还行。
Clinical Statistician I – B ABLJP00027218
Abbott Labs
Abbott Park, IL 60064
8 – 12/2011
Experience in database management (preferably claims and other healthcare
databases), SAS programming and analysis skills are also a plus. Understands
and appropriately uses advanced epidemiological research techniques: •
;Understands primary and secondary research techniques as well as the
elements of more complex quantitative research methods, s... 阅读全帖
A*******s
发帖数: 3942
35
来自主题: Statistics版 - proc nlmixed很豪放
问问大牛们, 你们搞nlmixed的时候容易收敛么?
不用说别的复杂问题,就sas自己的例子,很豪放,不容易收敛
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/d
这个初始值要是我自己猜的话怎么都不行... 搞grid search来选初始值也不行。
比如说我把sample code改了parms statement
parms beta1=-3.01 to 3 by 0.5 beta2=-3.02 to 3 by 0.5 beta3=-3.03 to 3 by 0.
5
s2b1 =0.1 to 2 by 0.1 cb12 =-1 to 1 by 0.1 s2b2 =0.1 to 2 by
0.1 s2=0.1 to 2 by 0.1/best=50;
结果出来
ERROR: Quadrature accuracy of 0.000100 could not be achieved with 31 points.
The achieved
accuracy ... 阅读全帖
c***i
发帖数: 15
36
来自主题: Statistics版 - eBay上海在招人
pls send email to xwang2 AT ebay DOT com if you are interested.
Search Metrics, Staff Data Mining Analyst
Job Responsibilities
eBay Search Metrics is building world-class engineering teams in Shanghai,
focusing on search related systems and services such as user behavior/intent
analytics and crowd sourcing judgments. The team’s work will directly
impact
eBay revenue by improving both search engine performance and user
experience on eBay.
We are seeking a data mining analyst to provide applied re... 阅读全帖
p********a
发帖数: 5352
37
☆─────────────────────────────────────☆
TNEGIETNI (lovewisdom) 于 (Tue Oct 11 20:01:18 2011, 美东) 提到:
朋友告诉我说他要来,还有Harvard的Carl Morris等,$200的Registration fee对我来
说不是一笔小数目。我其实并不欣赏他的bootstrap法,曾公开批评过这个方法的逻辑
错误。如果去,我跟他讲什么呢?版上的恶朋好友可否给点建议?谢谢。
☆─────────────────────────────────────☆
angelsun (安吉笋) 于 (Tue Oct 11 20:38:46 2011, 美东) 提到:
在哪里?什么时候?讲座是什么topic的?

☆─────────────────────────────────────☆
statcompute (statcompute) 于 (Tue Oct 11 23:51:23 2011, 美东) 提到:
$200 for a lecture by Efron? it i... 阅读全帖
k*******a
发帖数: 772
38
来自主题: Statistics版 - a R question
这个好像没啥好办法, 见R-FAQ
7.31 Why doesn't R think these numbers are equal?
The only numbers that can be represented exactly in R's numeric type are
integers and fractions whose denominator is a power of 2. Other numbers have
to be rounded to (typically) 53 binary digits accuracy. As a result, two
floating point numbers will not reliably be equal unless they have been
computed by the same algorithm, and not always even then. For example
R> a <- sqrt(2)
R> a * a == 2
[1] FALSE
R> a * ... 阅读全帖
c********h
发帖数: 330
39
来自主题: Statistics版 - What if binary variable is very unbalanced
如果是做response的话,看accuracy的时候要注意,因为认为所有都是男的,accuracy
还有0.9,一般用precision recall。另外太unbalanced的话,可能fitting不converge
c***z
发帖数: 6348
40
来自主题: Statistics版 - 请教logistic regression问题
consider F1 score for accuracy instead of the vanilla accuracy
h******s
发帖数: 3420
41
Contractor 看人了。
进去是第一步,后面是剪刀,好的越来越好,差的越混越差。
我觉得主要的区别就是 accuracy 吧,对 trial 对 definition 写报告的 accuracy.
这个是programmer 而言。
分析不是最重要的,但是你要么不做,要么完全做的分毫不差。
我也看到很多尤其女生,逻辑不好,所以 code 很容易错,小错不断,这种日子就非常
难过,想拿高薪是不可能的,能保工作就不错。要么做master level biostatistician
, coding 要求不高,但是没前途。
h******s
发帖数: 3420
42
Contractor 看人了。
进去是第一步,后面是剪刀,好的越来越好,差的越混越差。
我觉得主要的区别就是 accuracy 吧,对 trial 对 definition 写报告的 accuracy.
这个是programmer 而言。
分析不是最重要的,但是你要么不做,要么完全做的分毫不差。
我也看到很多尤其女生,逻辑不好,所以 code 很容易错,小错不断,这种日子就非常
难过,想拿高薪是不可能的,能保工作就不错。要么做master level biostatistician
, coding 要求不高,但是没前途。
s**********8
发帖数: 25265
43
来自主题: MedicalDevice版 - key device GMP component - design control
Design Controls
INTRODUCTION
Coverage
QUALITY SYSTEM
Personnel Training
DESIGN AND DEVELOPMENT PLANNING
Interface
Structure of Plans
DESIGN INPUT
Input Checklists
DESIGN REVIEW
Combination Devices
Preparation For Reviews
Why Design Reviews
Types Of Design Review Meetings
Design Review Requirements
End Of Initial Design
DESIGN OUTPUT
Documenting Design Output
Acceptance Criteria
Design Output Approval
DESIGN VERIFICATION AND VA... 阅读全帖
w***n
发帖数: 9040
44
来自主题: Tri版 - 910XT下水初体验
Troubleshooting Pool Swim Data
I wanted to briefly talk to this, since I’ve seen a number of folks ask
about accuracy in the pool. Some have had the distance issues where the
unit reports longer than normal. And a few people have issues where it
reports shorter. With that, I wanted to provide some tips based on my using
it for the past 5+ months. During which time, I’ve had it measure
distance in a pool wrong only once – due to having to stop mid-way down the
lane and converse with my lovely... 阅读全帖
o******s
发帖数: 1364
45
来自主题: Archery版 - dual cam需要多久做一次synchronize
I second this comment. When I started I believed accuracy is the key.
Now I understand if I cannot get to the deer, accuracy means nothing.

leave
dual
out
c***z
发帖数: 6348
46
来自主题: DataSciences版 - [Data Science Project] Location data quality
Had some more progress. Using some better data, and after correcting for
flipped clusters (i.e. usually the bad points are in cluster 1, but
occasionally they like cluster 2 better), I had 95% accuracy in clustering
the points.
Now the bootstrap step, I labeled test points with its nearest neighbor, and
had 80% accuracy using a majority vote by the metrics. I am modifying
the algorithm so that I can allow more false positives and less false
negatives, as required by the business.
The real headac... 阅读全帖
c***z
发帖数: 6348
47
来自主题: DataSciences版 - [Data Science Project] Location data quality
Had some more progress. Using some better data, and after correcting for
flipped clusters (i.e. usually the bad points are in cluster 1, but
occasionally they like cluster 2 better), I had 95% accuracy in clustering
the points.
Now the bootstrap step, I labeled test points with its nearest neighbor, and
had 80% accuracy using a majority vote by the metrics. I am modifying
the algorithm so that I can allow more false positives and less false
negatives, as required by the business.
The real headac... 阅读全帖
c********1
发帖数: 60
48
来自主题: DataSciences版 - 刚入行新人的两个问题
刚刚转入data science不到两个月,所知甚少,如果有大牛觉得我的问题太过白痴,请
见谅!
这几天在做一个project,有了两个发现,也算是问题,在此分享,希望版上各位同仁
大牛能够解答。
(1)用Radial SVM做classification,一开始忘了normalization(数据的scale差异极
大,有的在0-1波动,有的超过1000,的确有这个必要),赶紧perform normalization
后发现accuracy,sensitivity, kappa这些metric基本没有significant improvement (
平均就增加了0.05左右)。不知道这是为何?
(2)Andrew Ng在他的cousera lecture vedio里说如果number of training samples
is intermediate compared to the number of features (他说的范围是指前者10-10
,000,后者1-1000),radial SVM要比linear SVM好。我的data set有99个obs, 20个
... 阅读全帖
w**2
发帖数: 147
49
来自主题: DataSciences版 - 刚入行新人的两个问题
我试着来回答一下:
(1)用Radial SVM做classification,一开始忘了normalization(数据的scale差异极
大,有的在0-1波动,有的超过1000,的确有这个必要),赶紧perform normalization
后发现accuracy,sensitivity, kappa这些metric基本没有significant improvement (
平均就增加了0.05左右)。不知道这是为何?
The kernel's in SVM makes computation in high dimension more convenient.
Depends on how it maps to the higher dimension, if it is Radial, it follows
the Gaussian distribution so there is the flavor of standardization in there
. Besides, your regularization C also controls how much wei... 阅读全帖
w**2
发帖数: 147
50
来自主题: DataSciences版 - 一个面试题(predictive model) (转载)
这题考的是evaluation metrics,像precision, recall, f1之类的。给的数据是
target variable是predicted target variable。
从confusion matrix入手。分析一下根据business model,是optimize precison还是
recall。然后具体怎么做。
为什么不是accuracy,你可以说accuracy可以很高,error很低的情况可能classifier
总是predict majority。
如果又多个model,你可以比较他们的auc score, indicating how well a classifier
separate the two classes
不要担心,多面几次就好了。
首页 上页 1 2 3 4 5 6 7 8 9 10 下页 末页 (共10页)