Opera Solutions Analytics

Netflix and the Netflix Prize: With more than 10 million subscribers and ’08 revenues of $1.4BN, Netflix (NASDAQ: NFLX) is the largest online movie rental subscription service in the U.S. Subscribers select titles at the company’s website, aided by a proprietary recommendation service (“Cinematch”); and then receive a stream of DVDs through the mail. To improve Cinematch’s performance, Netflix launched the Netflix Prize in late 2006, offering a $1MM award to the team that could improve predictive performance by 10% or more. The contest originally attracted over 35,000 teams from around the world, and ended with over 4,200 active teams from the worlds of academia, industry, and business.

A Highly Complex Problem: Predicting future movie preference based on past ratings may sound simple, but it requires breakthroughs in modeling and analytics as well as computing power that only recently has become widely available. Developing collaborative filtering models on a dataset of this size – 100MM ratings of 18,000 movie titles submitted from 500,000 randomly-chosen, anonymous customers – requires a sophisticated and artful application of multiple machine learning algorithms. Only 1% of the customer/movie matrix is populated with ratings, so models must be able to work effectively with significant data gaps. User mood and preferences change over time, and one Netflix account may have multiple users and raters, so the data contain significant noise and variation. And the predictive variables (movie ratings) are also the explanatory variables, creating a very thorny joint prediction problem. Solving the Netflix problem requires a new class of models and analytic techniques that are non-linear, adaptive and “learning,” non-parametric, capable of handling high dimensional input and output, and able to capture both local and global effects.

Opera Solutions’ Performance and Breakthrough Approaches: Along with our partner, Vandelay Industries !, Opera finished the competition in 4th place, maintining its position as the leading consulting firm in the contest for over a year. Our San Diego-based team of senior scientists applied advanced non-linear modeling and enhancement techniques, including Neural Networks, Matrix Factorization/Singular Value Decomposition, Restricted Boltzmann Machines, Stochastic Gradient Descent, and a new approach to K-Nearest Neighbor, “Learning KNN.” The team also has broken new ground in combining the results of these various modeling approaches via ensemble techniques; the current consolidated output is a combination of 700+ individual models and has upwards of 5 billion parameters.

Real World Applications: Opera Solutions has had significant success in applying its analytic advances to real-world business problems that also require cutting-edge predictive capabilities. Examples:

•Credit Management: a Top 5 card issuer had tightened credit among a large group of consumers, generating intense consumer/press backlash. Using Neural Networks and multi-dimensional Binning, Opera Solutions eliminated 20% of the group from adverse actions without increasing exposure, improving loyalty, reducing customer complaints, and retaining $20MM in profits that would otherwise have been lost.

•Fraud Reduction: a credit card company wished to reduce “bust-outs” – the situation in which a customer suddenly maxes out credit limits and then rapidly defaults. Opera’s Neural Network models not only found significantly more bust-out candidates, but also found them earlier, sharply reducing the client’s financial exposure.

•Cross-Selling: a client in the grocery business faced stagnant same-customer sales. Using a unique combination of non-linear models to predict which products their customers were most likely to buy, Opera integrated these recommendations into a revised sales strategy to increase revenues by 15-20%.

•Marketing: a client sought to improve investment across a variety of customer acquisition channels, and linear methods were not able to create predictive relationships between specific marketing channels and new customer applications. Using an Ensemble approach of Neural Network techniques, Opera improved performance by 6.6X; we then applied an Optimization Solver to maximize marketing ROI across channels.

Opera Solutions’ predictive modeling achievements, honed in the Netflix Prize competition, represent a transformative opportunity to deepen companies’ understanding of customers and prospects and predict their behavior. For marketers, these capabilities provide sharply enhanced abilities to offer customers goods and services they are most likely to value. For credit-granting institutions, these new techniques represent more effective ways to strip out risk and volatility from revenue streams. In short, these techniques move companies to the next level in bringing new life and expression to individuals through data.

More About Opera's Participation in the Netflix Challenge