Something about Interview of Data Scientist

trend of general data scientist interviews


earlier,

more like software engineer interviews
heavy on coding
light problem solving and probablity

currently 

more balance btw coding & ML statistics
coding still dispensable
0-1 coding problems in phone interviews
1-2 coding problems in onsite interviews

new requirement
deep understanding into algorithms and metrics
context/domain knowledge for problems


example:


为什么decision tree会overfitting
minimum split减少nonlinear bagging 
greedy method，criterion，一个数据点，generalize所有数据
一个数据点，variance大，参数越多数据越少
层数，最小node size random forest 
performace metrics 
inbalanced data 
down-sampling 1／10 down settling 十倍数目
AOC 不变

数据不平衡 up sampling down sampling
smote sampling package

overfitting

为什么方法1比2好，参数这么调
limited area collection prepare
bayes rule—
debug bootstrap 
A,b test p-value —phone interview
industrial blogs and papers 
work on public competition data
with guidance from profs 
focus on 2-3 areas


suggestions

Regression likelihood
P value常见
Statistic tests f test 选择变量
面经，有限范围

Leetcode medium 高频题

How about AI

only if you really understand and have hand on experience with dl and ai 
do not claim it on your resuyme
if really interested,build up DL and AI experience and knowledge

explain RNN why it can work for certain problems
CNN
what is backward prop why its nessesary 
why gradient vanishing why bad how solve
what is ReLu advantage
how to tune deep neural network



language understanding translation recommendation
cnn image 构造 每一步 】
参数怎么调
initialization



summary 

plan 
know yourself and set goals
lay out:
1. coding sql
2. probability
3, ml
4, project experience

go deep to ml and projects
optimize ROI of interview preparation 


算法 掌握程度 background 投入时间 coding probability ml algorithm case study
时间 return

算法挖得很深 算法比较
curiosity critical thinking
ROI return of investment 
Boosting 为什么比较好为什么比RF好










评论

发表评论

此博客中的热门博文

8 Link Analysis

图片
Pagerank—link analysis  Early web search: extract keywords-->queries match -->page ranked by  occurrences of keywords Term Spam: (TS) Disguise a page as sth it is not about  Pagerank: vs TS Rank pages by linkage too How many pages point to a page= # incoming links => importance Link spam (fade incoming pointers) Manually identify some important pages and assign them page ranks , then let ranks propagate to other nodes. Random surfer model  Random surfer start from any page follow its outgoing links randomly   来，去，都是随机的 Pagerank = Prob ( a random surfer lands on the page ) More robust than manual approach , a collective voting scheme Limiting distribution  Graph is strongly connected , a node can reach any other node in the graph But cannot have dead end & spider traps A--->B, C, D;    B--->A, D,(A,D->C);     C--->A(->B,C,D);   ...
继续阅读

1 Map reduce problems

图片
PROBLEMS Join   Matrix multiplication 2 / 1 phase  Matrix vector multiplication  POINTS Map reduce input output Memory requirement  Reducer size  Communication cost P1 Join:  R ( A , B )  S ( A , C )   1. Map r ( a , b ) => ( a , ( 'R' , b ));  s ( a , c ) => ( a , ( 'S' , c ))   2. Reduce ( a , [( 'R' , b ), ( 'S' , c1 ), ( 'S' , c2 )]) => ( a , ( b , c1 )), ( a , ( b , c2 ))   Dangling tuples: Key with values from only 1 relation  ( a , [( 'R' , b )]) => left dangling ; ( a , [( 'S' , c )]) => right dangling  P2 Matrix-multiplication  C mXn = A mXp B pXn  2 phase: a. phase1  product  Map : Extract key from A ’ s c , B ’ s r A [ i , k ] => ( k , ( 'A' , i , A [ i , k ])) ; B [ k , j ] => ( k , ( 'B' , j , B [ k , j ]))   Communication cost : mp + np  Reduce...
继续阅读

NoSql and AWS DynamoDB practices

图片
no sql and dynamoDB https://mp.weixin.qq.com/sbiz=MzAxMzA2MDYxMw==&mid=2651560478&idx=1&sn=e8ce8d0ff0067059cf9946eee50bcbf6&chksm=80576574b720ec626102d9eca08da2424fedb3d0d58770cc26ed5ae8402a37ee97ae2c4441b1&mpshare=1&scene=24&srcid=0416hM7NzIQUooRU92fn4VOv&key=017ace5cd749d20d13db07527244da37a6c4e17bce708cfa37916f41f4f2cf0450b750489db9c37ee67227bbbd43207b03aaed62b9cea6e3d931a61a83e29b5ce94e014a496ea83f5effe88ca0c7d4f8&ascene=0&uin=ODc5NTM5MzAx&devicetype=iMac+MacBookPro13%2C2+OSX+OSX+10.12.3+build(16D32)&version=12020110&nettype=WIFI&fontScale=100&pass_ticket=TZiwhj6StjhSR05qVUyKEcClLlArTC5CqpFW3qEaROK%2FfI7ojwh9a1esoikANmQh Relational database: Mature and stable Feature rich versatile query language Acid  No sql  Flexible  Scaled out easily Data replicated over multi-servers Weak consistency high availability Scale up out Scale up vertical , CPU Scale out adding more computers ...
继续阅读