Author: admin

 
Topic
Difficulty Level
(High / Low) Questions
Refs / Answers

1.
Text Mining
L
Explain :TFIDF,  Stanford NLP, Sentiment Analysis, Topic Modelling
 

2.
Text Mining
H
Explain Word2Vec. Explain how word vectors are created
https://www.tensorflow.org/tutorials/word2vec

3.
Text Mining
L
Explain Distance :  hamming,  cosine or eucleadean.
 

4.
Text Mining
H
How can I get single vector for sentence / paragraphs / document using word2vec ? 
https://radimrehurek.com/gensim/models/doc2vec.html

5.
Dimestion Reduction
L
Suppoese I have TFIDF matrix having dimensions 1000×25000. I want to reduce the the dimensions to 1000×500. What are the ways available ?
 PCA , SVD, (max df, min df, max features in TFIDF)

6.
Dimestion Reduction
H
Kernel PCA, tSNE
http://scikit-learn.org/stable/modules/decomposition.html#decompositions

7.
Supervised Learning
H
Uncorrelated vs highly corelated features :  How they will affect linear regression vs GBM vs Random Forest
GBM and RF are least affected

8.
Supervised Learning
L
If Metioned in Resume ask about : Logistic Regression, RF, Boosted Trees, SVM, NN
 

9.
Supervised Learning
L
Explain Bagging Vs Boosting
 

10. 
Supervised Learning
L
Explain how variable importance is computed in RF and GBM
 

11. 
Supervised Learning
H
What is Out Of bag in bagging
 

12. 
Supervised Learning
H
What is difference between adaboost and gradient boosted trees
 

13. 
Supervised Learning
H
What is learning rate ? What will happen if I increase my rate from 0.01 to 0.6 
The learning will be unncessarlity fast and the chances are that because of increased learning rate, global minima will be missed and weights will fluctuate. But if learning rate is 0.01, the learning will be slow and the chances are model will get stuck in local minima. Learning rate shoul dbe decided based on CV / parameter tuning

14.
Supervised Learning
L
How would you choose parameters of any model?
http://scikit-learn.org/stable/modules/grid_search.html

15.
Supervised Learning
L
Evaluation of Supervised Learning, Log Loss, Accuracy , sensitivity, specificity, AUC-ROC curve, Kappa
http://scikit-learn.org/stable/modules/model_evaluation.html

16.
Supervised Learning
L
My data has 1% Lable 1 and 99% lalel 0 , and my model has 99% accuracy? Should I be happy ? Explain Why
No. This might just mean that model has predicted all 0s with no intelligence. Look at Confusion Mat, Sensitivity Specificity, Kappa etc. Try oversampling, Outlier Detection , diferent algos like RusBoost etc

17.
Supervised Learning
H
How can I increase the percentage of Minority class representation in this case ?
SMOTE, Random Oversampling

18.
Unsupervised Learning
L
Explain Kmeans 
http://scikit-learn.org/stable/modules/clustering.html#clustering

19.
Unsupervised Learning
L
How to choose no of clusters in K means
https://www.quora.com/How-can-we-choose-a-good-K-for-K-means-clustering

20.
Unsupervised Learning
H
How to evaluate unsupervised learning algorithms
http://scikit-learn.org/stable/modules/clustering.html#clustering-performance-evaluation

21.
Unsupervised Learning
H
Which algorithm doesn’t require no of clusters as an input ? Birch , DBSCAN, etc
http://scikit-learn.org/stable/modules/clustering.html#overview-of-clustering-methods

22.
Unsupervised Learning
H
Explain AutoEncoder- Decoders
 

23.
Data Preprocessing
L
Normalising the data : How to normalise Train and Test data
http://scikit-learn.org/stable/modules/preprocessing.html#custom-transformers

24.
Data Preprocessing
L
Categorical variables : How to convert categorical variablesin to features 1- when no ordering, 2- when ordering
Dummy / one hot Encoding , Thermometer Encoding

25.
Unsupervised Learning
H
How kmeans will be affected in the presence of dummy variables
 

26.
Deep Learning
H
Deep learning : Explain activation function  : ReLu, Fermi / sigmoid , Tanh ,etc
www.deeplearningbook.org

27.
Supervised Learning
L
Explain Cross Validation : Simple,  , If it is time series data can normal cross validation work ?
http://scikit-learn.org/stable/modules/cross_validation.html

28.
Supervised Learning
L
Explain : Stratified and LOO CV
http://scikit-learn.org/stable/modules/cross_validation.html

29.
Supervised Learning
H
In Ensemble Learning, What is Soft Voting and Hard Voting
http://scikit-learn.org/stable/modules/ensemble.html#voting-classifier

30.
Supervised Learning
L
Ensemble Learning: If correlations of prediction between 3 classifiers is >0.95 should I ensemble the outputs? Why if Yes andNO?
 

31.
Optimisation
H
What is regularisation, is linear regression regularised , if no then how it can be regularised
L1, l2 : See Ridge and lasso 

32.
Supervised Learning
L
Which algorithms will afected by Random Seed : Logistic regression, SVM, RandomForest,  Neural nets
RF and NN

33.
Supervised Learning
H
What is Look Ahead Bias ? How it can be identified ?
 

34.
Supervised Learning
H
Situation : I have 1000 Samples and 500 Features. I want to select 50 features. I Check the correlation of each of the 500 variable with Y using 100 samples and then use top 50. After doing this step I run  cross validation on 1000 sample. What is the problem here ?
This has Look Ahead Bias

35.
Optimisation
H
Explain Gradient Descent. Which one is better Gradient Descent or SGD or ADAM ?
http://ruder.io/optimizing-gradient-descent/

36.
Supervised Learning
L
Which algorithm is faster : GBM Trees or xgBoost ? Why
Xgboost :
https://arxiv.org/abs/1603.02754

37.
Deep Learning
H
Explain back progapagation 
www.deeplearningbook.org

38.
Deep Learning
H
Explain Softmax
www.deeplearningbook.org

39.
Deep Learning
H
DL : For Time series which archeture is used : MLP / LSTM / CNN ? Why ?
www.deeplearningbook.org

40.
Deep Learning
H
Is it required ot normalise the data in neural nets ? Why ?
www.deeplearningbook.org

41.
Optimisation
L
My Model has Very High Variance but Low Bias. Is this overfitting or underfitting ?
If ans is Overfitting ( Which is correct) how can I make sure I don’t overfit.  

42.
Deep Learning
H
Explain Early Stopping
http://www.deeplearningbook.org/contents/regularization.html#pf20

43.
Deep Learning
H
Explain Dropout. Is bagging and dropout similar concepts ? If No , what  is the difference ?
http://www.deeplearningbook.org/contents/regularization.html#pf20 https://goo.gl/gWrdWD #DataScience #Cloud

Social media has exploded. And it is rapidly becoming the choice of marketers who want to gain a large audience, spread their brands and develop relationships that will result in trust and ultimately sales.

But competition is fierce, and marketers who must craft content are finding it increasingly difficult to cover all of the social media platforms they want to. One of the solutions is to use AI-powered tools to save manpower and to improve efficiency.

Use of AI Tools on Social Media

The AI tools available now allow marketers to generate more content faster; bots can analyze the performance of content; AI tools find and sort data on customer profiles by mining information of users; AI tools are also used to monitor all comments of users/consumers; and marketers even use AI to collect data on influencers and match the right influencers to their brands.

One Key AI Use – Translation

There are more than two billion users of Facebook, and the majority are not English-speaking. And as businesses and their marketers are increasing their attempts to reach a global audience, the need for translation of social media posts is real.

In the past, Facebook has used a machine translation program, which was less than ideal, for certain.

Now, Facebook has developed an improved translation service, powered by neural networks, a type of AI. It is called “convolution neural network (CNN).”

While in the past, the Facebook translation program performed translation, word by word, this new AI-based translation focuses on phrases and, over time, phrases and phrases within context of the content.

According to Facebook, its new translations are more accurate than previous machine translations. As well, the AI machines continue to “teach” themselves to interpret large language data sets, in order to understand natural language, including idioms and slang. So far, accuracy has increased by about 20% over earlier machine translations.

The Beauty of Facebook’s New AI Translation Feature

Here’s the thing. There are two key features that users will love:

* They will have the option to automatically translate anything into an array of languages. If something is posted in English, and someone in France has their preferred language listed as French, any post will automatically be translated into French. Think of the benefit to marketers when they can simply post something in English and not worry about a translation into the languages of consumers they are targeting globally.
* Users in foreign countries will have the option to “fix” translation errors. This is huge. Recipients of translated posts have the option to go in and correct translation errors; they are thus allowing the machines to learn more and use those corrections in any future translations they do. Over time, user involvement in this process will mean that marketers get their messages across without laborious manual translations.

Where Does This Leave the Competition

Facebook has obviously set the bar AI translation. While its function is not by any means perfect, it is a massive improvement over previous translation activity. And, as its machines continue to get better at natural language processing, translations will become more accurate.

Until then, of course, marketers will continue to use the most reputable and largest translation companies to localize their messages for foreign audiences. But, as the Bob Dylan song says, “The Times, They are a’Changin.’”

There is certainly the incentive for other social media platforms to develop their own AI translation functions too. Google is also moving forward on AI translation. Social media platforms like Instagram, Twitter, and others will have to get on board as well, if they want to remain competitive and continue to capture a large share of the reading and viewing market.

Marketers will gravitate toward those channels that offer accurate AI translations, as they strive to reach global markets. The savings in time and money are just too attractive. And the opportunities to connect with a global audience for marketing purposes is just too critical.

The Future is Now

AI is disrupting almost every sector of the economy. Social media should be no different. AI is already providing marketers with a wealth of data they can use to reach their audiences on social media platforms. Translation through AI technology is a logical outcome, as marketers are continually finding ways to streamline and spread their brand messages to a wider audience. There will soon come a day when accurate and contextually correct translations are available on all social media platforms – automatically, with no need for additional work on the part of brand marketers. https://goo.gl/vFTmZV #DataScience #Cloud