Case Study: FPT-Play Video Recommender System

Industry: Media & Entertainment

The Client

FPT Play is an IPTV service operated by FTP Telecom. FPT Play is accessible through mobile, smart TV and web services. It is currently serving millions of users.

Business Needs

FPT Play wants to increase customer retention rate by recommending them the most relevant content.

Solutions

We collect user action log on FPT Play system via Flume.

We use Mapreduce to analyze user action history to get their personal preference and try to guess what users are interested in.

Recommended items for a given user are the combination of what are commonly hot and what he is personally interested in.

The Benefits

We get a relatively high average click rate on the recommended item.

User’s average time on site is significantly improved when users find their preferred items.

Technical Details

Personalized Video

  • Recommend videos watched by like-minded users. Need to discover users who have similar taste based on their watch history or social relationship. Implemented using Mahout’s matrix factorization collaborative filtering.

Co-watched Video

  • Recommended videos those usually watched next after current video. Implemented using Hadoop Map-reduce.

Trending Videos

  • Videos with highest “hot” score which is calculated from multiple factors: rising view count, high rating, rising number of comments, shares, likes… data collected from both internal (FPT play) and external sources (e.g. IMDB).
  • Implemented using Hadoop Map-reduce.

We are digitalizing the Media & Entertainment industry. Checking out our Service offering for Media & Entertainment here.