Every business faces an open-ended challenge: to maximize
the revenue opportunity out of each interaction with every
customer. Businesses offering a wide range of products or
services face the additional challenge of matching the right
product or service based on immediate browsing and search
activity along with the historical data for the customer.
Recommender systems are designed to tackle this business
need. Their impact can be huge: Amazon, whose catalog
size is unmatched, reports that 35% of its sales comes from
cross-sell and up-sell recommendations.
Recommendation is a never-ending battle, for accuracy and
In today's hypercompetitive world with the sub-second
response requirements, just having any old recommender
isn't enough. Traditional recommenders perform global
statistical computations offline, using snapshots of data that
can be days old. They don't have the real-time modeling and
nuanced profiling that is needed today.
Real-time speed is essential. Website visitors have short
attention spans: catch them now, or they're gone. For
retailers, their product catalogs can change by the minute.
Recommenders need to quickly understand the profile of their
client, align that with the rapidly changing profiles of the larger
customer base and product catalog, and produce engaging,
personalized recommendations. Traditional recommendation
engines built on the relational databases aren’t able to keep
up with these requirements.
Why TigerGraph, a Native Parallel Graph Database for Real-time Deep Link Recommendation Engine?
Deep Link Recommendation Engine
Using a graph for recommendation analytics is the first step
towards faster and more personalized recommendations.
Consider that the standard collaborative filtering algorithm is
a 3-hop graph query:
(People → who bought the Product ← which You just bought)
→ also bought these other Products
But all graph databases are not created equal. TigerGraph
easily handles 3 to 10+ hops, while many others struggle with
more than 2 hops. TigerGraph's Deep Link Analytics let you
customize and extend your analytics -- adding in hops to
consider product features, customer demographics, and the
context of the current situation -- which result in more
accurate and more personalized recommendations.
For example, It takes 2 hops to find similar shoppers:
Shopper → (Demographics) → Similar Shoppers
So demographic-aware collaborative filtering can be
implemented in 5 hops: Shopper → Products purchased → Other Shoppers who
purchased the product → Demographics → Other Shoppers
that belong to the Demographics → Other Products
This is illustrated with the example, where Customer 1 has
purchased a toy batmobile and light up shoes in prior visits.
Customer 2 and Customer 3 have both purchased the
batmobile toy, and belong to the demographics of “Suburban
affluent parent”. Customer 4 belongs to that demographic as
well and has purchased the video game “Super Mario Party
for Nintendo Switch”. Based on the 5-hop analysis, the video
game, “Super Mario Party for Nintendo Switch” is the
recommended product for Customer 1.
TigerGraph's native parallel graph delivers not only more
personalized results, but it does it in real-time. The result is
the capture of key “Business Moments”, transient
opportunities where people, businesses, data and "things"
work together dynamically to create value. TigerGraph lets
businesses capture these
moments to personalize the customer experience which leads
to more transactions.
One of the leaders in mobile e-commerce uses TigerGraph to
help drive online, personalized recommendations for its 300
million users. They use TigerGraph to model its catalog of
millions of products from thousands of vendors, and its
consumer data, including each shopper's real-time behavior
on their website. TigerGraph increased the company's query
speed by 100x compared to their previous solution and has
helped it scale from a startup to Billions in annual revenue in
just six years competing and winning against Amazon. The
result is a better shopping experience with offers that are
more likely to result in an initial sale, an up-sell, and cross-
Building the Next Generation Recommendation engine with AI and Machine Learning
Graphs with real-time deep link analytics are also powering
the next generation of AI-based recommenders using
machine learning to understand user behavior and preferred
actions better than before. Have you ever made a major
online purchase, such as a TV or washing machine, and then
the recommender asks if you would like to buy a similar
product? No! An intelligent recommender would recommend
Similarly, an intelligent recommender
can use the shopper's browsing behavior to guess at what
stage a shopper is at: looking around, evaluating multiple
options, price shopping, vacillating, etc. Then the
recommender can make an appropriate suggestion to nudge
the shopper along.