Case study: IMVU Recommendation Engine
Cloudwick Machine Learning helps IMVU stay ahead of the game
IMVU, Inc. (www.imvu.com) is an online social entertainment destination where members use 3D avatars to meet new people, chat, create and play games with friends. IMVU has over 50 million registered users, 10 million unique visitors per month and three million monthly active users. IMVU has the world’s largest virtual goods catalog of more than 20 million items, most of which are created by members.
IMVU is free and users can simply register and select a personal avatar. Avatars can be customized by purchasing items from its virtual goods catalog. The company provides promotional credits to get started after which users can purchase additional digital products to create personalized experiences.
Founded in 2004 and located in Redwood City, CA IMVU is backed by venture investors Menlo Ventures, Allegis Capital, Bridgescale Partners and Best Buy Capital.
IMVU is unique in that it enables people to socialize in their alter ego—the virtual platform protects users and allows to drop inhibitions, encouraging them to fantasize and create imaginary personalities. Users have created thousands of chatrooms based on different interests where discussions are rich and real—hooking users, enticing them to come back to join the conversations again and again.
The IMVU Virtual Shop is the largest of its kind, with approximately 7,000 new items added each day. The marketplace allows users to personalize, customize, stylize, and animate 3D avatars and environments to the hearts’ content.
This is where the company makes money. But the platform has to ensure members are spending time chatting, interacting, creating personas and adding dimensions with jewellery, clothing, cars, new hairstyles, decorating the rooms—mimicking the real world while creating a fantasy.
The business imperative for the platform is to extend the duration of user engagement as opportunities are directly proportionate to the time members engage with the platform. This is driven by two conditions: larger the community, more the scope of engagement; and ability to provide compelling features, environments to attract sticky users.
While IMVU has succeeded in the first objective of increasing registrations, it has struggled to convert them into active users. Therefore one of the immediate goals of IMVU was to enable users to create personalized chatrooms by providing appropriate recommendations. Currently chatroom recommendations are based on popularity which does not necessarily reflect individual likings.
IMVU turned to Cloudwick to design intelligent algorithms to get deep insights into user behaviour and create personalized recommendation engine.
The Cloudwick Solution
Cloudwick designed Machine Learning algorithms on Amazon EMR clusters using Spark to analyse logs from S3 which captured details of different aspects of user behaviour—which screens user clicked, which chatroom user went, how much time user spent in different chatrooms, etc.
Algorithms based on Collaborative Filtering help make automatic predictions about the interest of user by collecting preferences from many users. Cloudwick prepared an initial matrix by putting raw data in data pre-processing modules. The matrix was created by mapping user id and rooms user visited with the number of times user visited own room viz-a-viz other rooms. The goal was to assess users’ room preference based on the strength of observation which includes frequency of visits and the duration of time spent in different rooms.
During data preparation stage, the Cloudwick team iterated several times to normalize as data sets were very large: 24 million active users and more than 500,000 chat rooms.
Once data was ready, Cloudwick applied machine learning algorithms and again iterated to find the best model to deploy and extract the explicit preference of users. While Spark collaborative modelling enables to predict missing entries by learning the latent factors, the iterations must be carefully executed, taking into consideration business goals and quality of data output that would best meet goal requirements.
By closely studying usage patterns and iterating the algorithm, Cloudwick created the application aligned to business requirement. The deployment used r4.8x large EMR clusters of 8 nodes, and Apache Zeppelin to write the Spark code.
The Next Steps
IMVU has started using the model to get insights into user behaviour and preference. Cloudwick will design another model in the second phase based on Content Filtering which takes into account the description of the item and profile of user preference. The hybrid model will enable the team to build a complete recommendation engine.