Scipy_v1_reprint
0 likes275 views
This document discusses using SciPy and real-time big data for site optimization. It describes a system using Kafka, Storm, SciPy and a Bayesian bandit algorithm to test different variations of content like headlines on a website to maximize click-through rates. The system analyzes impressions and click data in real-time to decide which variation to show users and improves its confidence in the estimated click-through rates of each variation as it receives more data.
Scipy_v1_reprint
- 1. SciPy and Real-time Big Data for Site Optimization Pyleus Message Processor Bolt Pyleus Event Worker Bolt Pyleus SciPy Optimizer Bolt Pyleus Update Messenger Bolt SciPy Bayesian Bandit Spout Spout Application State Visitors to Bankrate.com Impressions and Clicks Improve User Experience For more info, contact: Winnie.Cheng@bankrate.com Which Variation to show Bankrate.com Data Science and Engineering Team Example: Pick better story headlines Objective for Site Optimization: Enable fast and cost-efficient ways of testing new designs to improve user experience Algorithmically decide which of two headlines to show user to maximize click-thru-rate (CTR) Computation Framework with Kafka-Storm Simulation Results With more data, algorithm becomes more confident of estimated CTR for each variation Bayesian Multi-Armed Bandit algorithm on Storm Topology decides how often to show each variation by analyzing impressions and clicks Iteration: 100 W1: 56.37% W2: 42.63% Iteration: 1000 W1: 9.82% W2: 90.18% Iteration: 2000 behavior reversal W1: 64.06% W2: 39.94% Iteration: 3000 W1: 94.84% W2: 5.16% Iteration: 4000 W1: 97.12% W2: 2.88%