Machine learning models depend on often large amounts of training data for supervised learning tasks. This data may be expensive to collect, especially if it requires human labeling (as in for document or image classification) and raises some particular quality issues. For example, how do we ensure that human agreement is high and what do we do in the event that it is not? Also, when your data is expensive to tag, how do you ensure that you have the smallest set possible that is representative of all your features? This talk will address these and other issues associated with gathering crowd-sourced, hand-coded data sets for supervised machine learning models.

The problem of predicting, for each type of crime, the crime frequency in a specific area on a specific day can be framed as a regression problem on crime frequencies and Twitter data: given (1) the last 31 days of Twitter activity geo-tagged in the immediate area, (2) the last 31 days of Twitter activity in the general area, and (3) the historical crime frequencies of the general area for the past year, predict the crime frequency for the next day in that location. Inherent in this problem description is the following prediction: the taxonomy used in tweets from around the area (such as a large fraction of restaurants with low ratings, or lots of tweets about how unsafe people feel in that area) contains information that could be used to build predictors of future crime frequencies for different types of crime. Assuming that the time series can be modeled as deviation from a periodic function, and incorporating this assumption into the model, may potentially produce better crime frequency estimates than directly predicting crime frequencies. The proposed research has implications for decision makers concerned with geographic spaces occupied by Twitter users. This session will cover these analytical results, which were produced by an extended group of graduate students and researchers at New York University.

The use of machine learning is a common theme in organizations today, yet most people still struggle with its definition given its many different levels. In this session, we attempt to eliminate this confusion by exploring a number of machine learning algorithms ranging from the simple to the more complex. We observe the use of these algorithms across a variety of industries as well as different behaviours such as customer response and customer risk. Alongside the comparison of machine learning algorithms, we also look at the impact of the data and how feature engineering impacts a given solution.

The model variable selection process is a key component of predictive analytics. Whereas logistic regression depends on feature selection/discovery being done beforehand, Random Forests and GBM tree based machine learning approaches intrinsically enact feature selections. However, default options for Random Forest and GBM are biased towards continuous variables, less favorable for categorical and binary variables. Unbiased solutions are very computational intensive. We propose a new approach that combines the strengths of intrinsic feature selection from Random Forests and GBM to detect non-linear relationships, enact scaling, and find interactions, and integrate these non-linear transformations into a traditional approach. A 7-10% lift is observed.

Accurate forecasting of customer demand for products is critical to success in the semiconductor industry. A diverse product portfolio, detailed customer mappings, dynamic market conditions, and extended production cycle times all sharpen the need for a reliable and responsive automated forecasting solution. I will describe a custom forecasting model recently developed by and now in use at Micron Technology Inc. that combines machine learning algorithms, established time series modeling techniques, and human expertise in leveraging existing forecasting infrastructure to achieve significant improvements in forecast accuracy across multiple levels of the product and customer hierarchy.

Three Steps for Improving Data Quality for Predictive Analytics

Bad data is the enemy of predictive analytics. From confusion about units or measure, to missing definitions, to data values that are just plain wrong, bad data gets in the way or, worse, leads predictions awry. Unfortunately, there is no "silver bullet" solution for all the problems that can arise.

Fortunately, practitioners can take steps to alleviate the bad data problems:

1. Understand quality levels and evaluate whether the data are fit-for-use in their analyses.
2. Use "rinse, wash, and scrub" routines to clean some bad data, and
3. Address the root causes of poor quality data longer-term.

This presentation briefly summarizes the issues, puts the steps above in context, and shows they contribute to better predictive analytics.

Preeminent consultant, author and instructor Dean Abbott, along with Rexer Analytics president Karl Rexer, field questions from an audience of predictive analytics practitioners about their work, best practices, and other tips and pointers.

For some of us, predictive accuracy is paramount when we assess our models: PCC, ROC AUC, Type I and Type II errors, etc. However, in other applications, the interpretation of predictive models is paramount so we understand why the model behaves the way it behaves. For this reason, many practitioners end up building models that are easier to interpret rather than models that are more accurate: regression or decision trees in particular. Neural Networks and Model Ensembles fall out of favor in these applications because they are perceived to be "black boxes".

This talk will describe an approach to determine the relative influence of each input variable in any predictive model using input shuffling, no matter how simple or complex the model. Interpretation of linear regression, logistic regression, neural networks, and Random Forest models will be compared and contrasted.

Software and analytics have been eating the world for a long time, and law and government are next. Businesses are increasingly transcending physical boundaries into new, unregulated virtual domains, forcing companies, developers and regulators to take a hard look at how data is collected, stored, and used. And new laws around the world are beginning to force the technology industry to rethink how it approaches the law. This talk will explain how and why the worlds of law and technology are colliding, and what this means for data-driven companies, the technology industry, and governments and citizens around the world.