Big Data

Building Effective go to Market Organizations

Derivatives

Entrepreneurship Through Acquisition

International Commercial Policy

BEG2MO Class #5 GTM INFLUENCES TYPE OF SALES PERSON 23 Relationship Account Executive Account Manager Sales Development Rep “Senior Sales Manager” – Boeing Retail Sales Associate Specialist and “Genius” – Apple “Educator” - lululemon Internet Telesales Rep Customer Service Rep “Customer Loyalty Agent” – Zappos Channel Relationship Manager Partner

BEG2MO_-_Class_5_PDF.pdf-BEG2MO Class #5 GTM INFLUENCES TYPE

BEG2MO Class #4 ALIGNING SALES WITH ORGANIZATION STRATEGY 11 • The link between company strategic goals and sales objectives must be crystal clear. • Research indicates that only 10% of companies strategic plans are effectively executed and on average firms deliver between 50% and 60% of the financial performance their strategies promise. (HBR Putting

BEG2MO_-_Class_4_-_PDF.pdf-BEG2MO Class #4 ALIGNING SALES WITH

[7] Big Data: Clustering Matt Taddy, University of Chicago Booth School of Business faculty.chicagobooth.edu/matt.taddy/teaching Clustering via Mixture Models You’ve seen lots of models for [ y | x ] (and [ y | d , x ] , etc) . Today is about models for x . We’ll work with mixtures: assume that each x i is drawn from one of K different mixture components , p k ( x ) . Even if the individual components are very simple, their mixture can yield all sorts of complicated distributions p ( x )= π 1 p 1 ( x )+ ...π K p K ( x ) Two algorithms: I K -means with Normal p k ( x ) I topic models with Multinomial p k ( x ) 2 Supervision The goal in everything we do has been Dimension Reduction . DR : move from high dimensional x to low-D summaries. Dimension reduction can be supervised or unsupervised. Supervised: Regression and classiﬁcation HD x is projected through β into 1D ˆ y Outside info (y) supervises how you simplify x . Unsupervised: Mixture and Factor Models x is modeled as built from a small number of components. You’re ﬁnding the simplest representation of x alone. We always want the same things: low deviance, without overﬁt. 3 Clustering: unsupervised dimension reduction Group observations into similar ‘clusters’, and understand the rules behind this clustering. I Demographic Clusters: Soccer moms, NASCAR dads. I Consumption Clusters: Jazz listeners, classic rock fans. I Industry Clusters: Competitor groups, supply chains. The fundamental model of clustering is a mixture : observations are random draws from K distributions, each with different average characteristics. Top down Collaborative Filtering: group individuals into clusters, and model average behavior for each. 4 The K -means Mixture Model Suppose you have K possible means for each observed x i : E [ x i | k i ]= μ k i , where k i ∈{ 1 ... K } e.g., if k i = 1, x i is from cluster 1: E [ x i 1 ]= μ 11 ... E [ x ip ]= μ 1 p . Then you have a mixture model. For new x with unknown k , E [ x ]= p ( k = 1 ) μ 1 + ... + p ( k = K ) μ K DR: Given μ k ’s, you discuss data in terms of K different types, rather than trying to imagine all possible values for each x . 5 Mixture Model: without knowing membership ‘ k ’ 5 10 15 20 25 30 35 probability density function x The marginal density has multiple modes; one for each μ k . 6 Mixture Model: breaking into K components 5 10 15 20 25 30 35 probability density function x Here, we have K = 4 different cluster centers. Should it be 5? 7 K -means K -means algorithm clusters data by ﬁtting a mixture model. Suppose data x 1 ... x n comes from K clusters. If you know membership k i for each x i , then estimate ˆ μ k = 1 n k X i : k i = k x i where { i : k i = k } are the n k observations in group k . K -means ﬁnds k = k 1 ... k n to minimize the sum-of-squares X k = 1 X i : k i = k ( x i - ˆ μ k ) 2 Mixture deviance: total sums of squares within each cluster. 8

BUSN | 41201 | Big Data | lecture notes | [7] Big Data: Clustering Matt 

What does it mean to be ‘big’? Big in both the number of observations ( size ‘n’ ) and in the number of variables ( dimension ‘p’ ). In these settings, you cannot : Look at each individual variable and make a decision ( t-tests ). Choose amongst a small set of candidate models ( F-test ). Plot every variable to look for interactions or transformations. Some BD tools are straight out of previous statistics classes ( linear regression ) and some are totally new ( trees, PCA ). All require a diﬀerent approach when n and p get really big. 8 Course Schedule subject to change... [1] Data: Computing, plotting, and principles. [False?] discovery. [2] Regression: A grand overview, linear and logistic. [3] Model Selection: penalties, information criteria, cross-validation [4] Treatment Eﬀects: HD controls, propensity scores, bootstrap [5] Classiﬁcation: Multinomials, KNN, sensitivity/speciﬁcity, DMR Midterm! [6] Networks: co-occurance, directed graphs, Page Rank [7] Clustering: Mixture models, k-means, and association rules. [8] Factors: Latent variables, PCA, PCR, and PLS. [9] Trees: CART and random forests, ensembles [10] Text Mining: topic models, sentiment prediction, deep learning 9 We’ll be working with real data analysis examples I Mining client information: Who buys your stuﬀ, what do they pay, what do they think of your new product? I Online behavior tracking: Who is on what websites, what do they buy, how do/can we aﬀect behavior? I Collaborative ﬁltering: predict preferences from people who do what you do; space-time recommender engines. I Text mining: Connect blogs/emails/news to sentiment, beliefs, or intent. Parsing unstructured data, e.g. EMR. I Big covariates: mining data to predict asset prices; using unstructured data as controls in observational studies . Many are applicable to marketing, but we’re far more general. 10 All of our analysis will be conducted in R This is the real deal: industrial strength software for data analysis. It’s free, cross platform, and hugely capable. Academics (stats, marketing/ﬁnance, genetics, engineering) , companies (EBay, Google, Microsoft, Boeing, Citadel, IBM) , and governments (Rand, DOE National Labs, Navy) use R . Since R is free, you’ll always be able to use it. 11 A huge strength of R is that it is open-source. This is also why it is sometimes a bit unpolished. R has a core , to which you can add contributed packages . These add-ons are as varied as the people who write them. Some are speciﬁc, others general, some are great, some suck. R is not without ﬂaws, but neither are the other options. e.g., I like python, but the community of stats developers is smaller and you need to be a more careful programmer. Some students prefer to wrap R in an IDE; e.g. R-studio. 12

BUSN | 41201 | Big Data | lecture notes | What does it mean to be ‘big’?

Contents CHAPTER i: Where The $100 Million Carne From • Start Here • The “Hot Coals” Sketch • The Painful Planning Mistake Boards and Sales VPs ..Make Everv Year • Have You Ever Felt Like an Utter Failure? • The $100 Million Sales Process • Make Your Lack of Money an Advantage CHAPTER 2: Coid Calling 2.0, Ramp Sales Fast Without Coid Calis • The First Breakthrough

All courses in BUSN

Popular Documents