How is Predictive Analytics Driving Business Performance?
Recently, I got to attend the Predictive Analytics Summit in San Diego. It felt great to be in a room full of data scientists from around the world—all my hidden statistics, operations research, and even modeling background came back to me instantly. I was most interested to learn what this vanguard was doing as well as any lessons learned that could be shared with the broader analytics audience. Presenters ranged from Internet leaders to more traditional companies like Scotts Miracle Gro. Brendan Hodge of Scotts Miracle Gro in fact said, as 125 year old company, he feels like “a dinosaur at a mammal convention”. So in the space that follows, I will share my key take-aways from some of the presenters.
Fei Long from 58.com
58.com is the Craigslist, Yelp, and Monster of China. Fei shared that 58.com is using predictive analytics to recommend resumes to employers and to drive more intelligent real time bidding for its products. Fei said that 58.com has 300 million users—about the number of people in the United States. Most interesting, Fei said that predictive analytics has driven a 10-20% increase in 58.com’s click through rate.
Ian Zhao from eBay
Ian said that eBay is starting to increase the footprint of its data science projects. He said that historical the focus for eBay’s data science was marketing, but today eBay is applying data science to sales and HR. Provost and Fawcett agree in “Data Science for Business” by saying that “the widest applications of data mining techniques are in marketing for tasks such as target marketing, online advertising, and recommendations for cross-selling”.
Ian said that in the non-marketing areas, they are finding a lot less data. The data is scattered across data sources, and requires a lot more cleansing. Ian is using things like time series and ARIMA to look at employee attrition. One thing that Ian found that was particularly interesting is that there is strong correlation between attrition and bonus payouts. Ian said it is critical to leave ample time for data prep. He said that it is important to start the data prep process by doing data exploration and discovery. This includes confirming that data is available for hypothesis testing. Sometimes, Ian said that this the data prep process can include inputting data that is not available in the data set and validating data summary statistics. With this, Ian said that data scientists need to dedicate time and resources for determining what things are drivers. He said with the business, data scientist should talk about likelihood because business people in general do not understand statistics. It is important as well that data scientist ask business people the so what questions. Data scientist should narrow things down to a dollar impact.
Barkha Saxena from Poshmark
Barkha is trying to model the value of user growth. Barkha said that this matters because Poshmark wants to be the #1 community driven marketplace. They want to use data to create a “personal boutique experience”. With 700,000 transactions a day, they are trying to measure customer lifetime value by implementing a cohort analysis. What was the most interesting in Barkha’s data is she discovered repeatable performance across cohorts. In their analysis, different models work better based upon the data—so a lot of time goes into procedurally determining the best model fit.
Meagan Huth from Google
Meagan said that Google is creating something that they call People Analytics. They are trying to make all people decisions by science and data. They want to make it cheaper and easier to work at Google. They have found through their research that good managers lower turnover, increase performance, and increase workplace happiness. The most interesting thing that she says they have found is the best predictor of being a good manager is being a good coach. They have developed predictive models around text threads including those that occur in employee surveys to ensure they have the data to needed to improve.
Hobson Lane from Sharp Labs
Hobson reminded everyone of the importance Nyquist (you need to sample data twice as fast as the fastest data event). This is especially important for organizations moving to the so called Internet of Things. Many of these devices have extremely large data event rates. Hobson, also, discussed the importance of looking at variance against the line that gets drawn in a regression analysis. Sometimes, multiple lines can be drawn. He, also, discussed the problem of not having enough data to support the complexity of the decision that needs to be made.
Ravi Iyer from Ranker
Ravi started by saying Ranker is a Yelp for everyone else. He then discussed the importance of have systematic data. A nice quote from him is as follows: “better data=better predictions”. Ravi discussed as well the topic of response bias. He said that asking about Coke can lead to different answer when you ask about Coke or Coke at a movie. He discussed interesting how their research shows that millennials are really all about “the best”. I see this happening every time that I take my children out to dinner—there is no longer a cheap dinner out.
Ranjan Sinha at eBay
Ranjan discussed the importance of customer centric commerce and creating predictive models around it. At eBay, they want to optimize the customer experience and improve their ability to make recommendations. eBay is finding customer expectations are changing. For this reason, they want customer context to be modeled by looking at transactions, engagement, intent, account, and inferred social behaviors. With modeling completed, they are using complex event processing to drive a more automated response to data. An amazing example given was for Valentine Day’s where they use a man’s partner’s data to predict the items that the man should get for his significant other.
Andrew Ahn from LinkedIn
Andrew is using analytics to create what he calls an economic graph and to make professionals more productive. One area that he personally is applying predictive analytics to is with LinkedIn’s sales solutions. In LinkedIn Sales Navigator, they display potential customers based upon the sales person’s demographic data—effectively the system makes lead recommendations. However, they want to de-risk this potential interaction for sale professionals and potential customers. Andrews says at the same time that they have found through data analysis that small changes in a LinkedIn profile can lead to big changes. To put this together, they have created something that they call the social selling index. It looks at predictors that they have determined are statistically relevant including member demographic, site engagement, and social network. The SSI score is viewed as a predictive index. Andrew says that they are trying to go from serendipity to data science.
Robert Wilde from Slacker Radio
Robert discussed the importance of simplicity and elegance in model building. He then went through a set of modeling issues to avoid. He said that modelers need to own the discussion of causality and cause and effect and how this can bias data interpretation. In addition, looking at data variance was stressed because what does one do when a line doesn’t have a single point fall on it. Additionally, Robert discussed what do you do when correlation is strong, weak, or mistaken. Is it X or Y that has the relationship. Or worse yet what do you do when there is coincidental correlation. This led to a discussion of forward and reverse causal inference. For this reason, Robert argued strongly for principal component analysis. This eliminates regression causational bias. At the same time, he suggested that models should be valued by complexity versus error rates.
Parsa Bakhtary from Facebook
Parsa has been looking at what games generate revenue and what games do not generate revenue for Facebook—Facebook amazingly has over 1,000 revenue bearing game. For this reason, Facebook wants to look at the Lifetime Value of Customers for Facebook Games—ithe dollar value of a relationship. Parsa said, however, there is a problem, only 20% pay for their games. Parsa argued that customer life time value (which was developed in the 1950s) doesn’t really work for apps where everyones lifetime is not the same. Additionally, social and mobile gamers are not particularly loyalty. He says that he, therefore, has to model individual games for their first 90 days across all periods of joining and then look at the cumulative revenue curves.
So we have seen here a wide variety of predictive analytics techniques being used by today’s data scientists. To me this says that predictive analytical approaches are alive and kicking. This is good news and shows that data scientists are trying to enable businesses to make better use of their data. Clearly, a key step that holds data scientist back today is data prep. While it is critical to leave ample time for data prep, it is also essential to get quality data to ensure models are working appropriately. At the same time, data prep needs to support inputting data that is not available within the original data set.
Solution Brief: Data Prep
Author Twitter: @MylesSuer