Every November 11, Chinese consumers embark on an online shopping spree that stops only when they “chop off” their hands, as the saying goes. This year, 11.11 sales totaled US$25.3 billion across e-commerce giant Alibaba’s eight shopping platforms, mostly Taobao and Tmall. The figure is double that of Black Friday and Cyber Monday totals combined.
US$1.51 billion worth of sales were recorded in the first three minutes of 11.11, which is also known as “Singles Day” or “Double 11.” Last year it took six minutes to reach the same amount. With so many sales in so little time, it’s no surprise that artificial intelligence is playing a growing role in the largest online shopping event in human history.
（Double 11 Super Project: Man-Machine Collaboration of Unprecedented Scale）
Rong Jin, Dean of Alibaba Institute of Data Science and Technologies, says the application of deep learning at Alibaba can be categorized into four main areas: computer vision for visual search, image classification and cross-media retrieval; combinatorial optimization for 3D bin packing and automatic banner design; speech and NLP technology for acoustic models, dependency parsing and mimicked QA; and model simplification for model compression.
11.11 is a day of non-stop action for the several thousand front-line Alibaba engineers in Hangzhou, whom colleagues have nicknamed the “siege lions.” In the quiet hours before Singles Day’s midnight start, some opt for a stroll by the west lakes, while others try to relax by their computer screens with a cup of Dragon Well tea.
In late October, 200 engineers performed the first 11.11 ‘stress testing’ to determine whether Alibaba’s hardware infrastructure could handle Singles Day’s peak levels. Alibaba also partnered with 500 banks, enterprises, logistics companies and government agencies to conduct “full-link” stress testing.
Full-link stress testing is an 11.11 rehearsal with no users involved. The process simulates sales using a similar online environment, the number of users, and transaction scenarios and scale. It then initiates a tuning process to optimize system functionality for peak performance. The goal is to monitor generic metrics such as CPU, memory, hard drives, reaction time; and also to ensure the system is protected.
Alibaba’s Vanguard Program is responsible for the stress-testing and capacity tuning processes. This year’s team required just three stress-testing sessions, down from 10 last year. The machines themselves can automatically make more than 50% of system repairs.
To ensure the billions of transactions run as planned, Alibaba uses AI external and internal feedback systems. DingTalk, an office communication tool similar to Slack, is used in conjunction with online chatbots to automatically log staff problems into the system. The system then applies machine learning algorithms to conduct a cluster analysis.
This automates the internal feedback loop and shortens time spent on communication. On Singles Day this basic application helps log and track all reported problems on shopping portals. Questions are processed by machines before they triage to DingTalk.
Online shoppers know the power of customer comments and ratings. While internal feedback systems monitor technical issues, external monitoring focuses on user comments.
Alibaba’s machine learning algorithm uses unsupervised clustering analysis for automatic comment monitoring, so problems are identified in real-time. The system de-noises incoming data, pre-processes it using business rules, then proceeds to use optimized K-medoids, CNN, and TextRank algorithms to do a clustering analysis. A final run through with AdaBoost meta-algorithm ensures generalization capabilities.
The E-commerce boom has transformed China’s logistics industry, with more than 60% of delivery parcels now originating from online orders. The workload skyrockets on Singles Day, and Alibaba Group affiliate Cainiao logistics does the heavy lifting.
A logistics niche applying deep learning technology is 3D bin packing, where one of the first questions is how to effectively pack a parcel using the smallest box possible. Using Deep-Q networks, Alibaba can determine the best fit, which has saved 5% of overall packaging costs.
While a packaging savings of 5% may sound trivial, it adds up fast over the over 812 million orders made on Singles Day. Optimization efforts have also cut delivery time from an average of nine days in 2013 to three days last year.
Amid the frenzy, Yang Lu, a Cainiao Logistics engineer in his early 20s, somehow found the time to complete an AI music composition program side project called MusicGo, which creates rap music in celebration of Single’s Day.
Lu collected ten thousand rap lyrics online, preprocessed them to remove punctuation and special characters, then trained a seq2seq model to generate its own rap lyrics. The final composed song is called The Rap of T-mall 11.11, a title inspired by the Chinese reality TV show The Rap of China.
Here’s a Google translation of the lyrics using the latest state-of-the-art NLP technology with some editorial touch-ups:
Come again a year Double 11,
Let me buy buy spend all the strength;
Come again this year Double 11,
Dark technology to change life, break the blockade that surrounds our region;
Smart technology connects you and me, light Double 11 night with ten thousand lights!
There are robots in the rookie, warehouse staged,
Cumulative enough intelligence they start to take the initiative,
Picking and packing smooth customs clearance,
From Hangzhou to Dongguan, from Shanghai to the Central Plains!”