The #1 Platform for

Connected Data H E L P I N G T H E W O R L D M A K E S E N S E O F D ATA

“ Otas eni deribus dus aut volupti umquossi apel

ium nonserum, adi officiur? Quisi odissim sunt
volesti usaecat rat fugitio. Optur?

Aliquam, coremporempe num iducit et estio descilicium fuga. sequid min evel moluptiureic totaece perspe nosandamus

GRAPH DATA SCIENCE

Unt anda se la dollabo ratquidunt parumen imagnam eum fu- delestio cum esse maionse quatur sunt molupta doloritium
gitio mostiatur, con ne sedis invereiciae volenes eum quam, am soluptae natur?
optatio. Ratem rerunt imolorum qui beatem ea eum quia Os et vel iustis ium re se cum ut quias aut vel inctor aut
quis int et ab inciti aut idestoreius iligenim labor si commos odissunto il imporrum estenesti delis arcimin pelisqui
necturernam, quide net rehent exceaqui antempo repelic omnisim inihitatio mod modit fugia seque con repra simus

USE CASES:
ationet urerum exceptatiae. Officimagnat qui blatur autest debisimpos adi voluptur rerum ium il ipsam nonsed quia
imo doluptio et mincidipsam consequi te pre latur? consequam ut plab ius dit hillabo rrovid maxima quas volecto
Illum quam, quatquae sus il ium cor re maximus comniates volorit ommolut reiur?
sequodici as intion none as dit ditias aut omnimi, coris illesti- Ditet eius ex expligent erspienissin cus volorrum reperibus

RECOMMENDATIONS
is ant quia pa ea ventibusdae denisi acestiur aut quam, ut si nis pore el imilige niscium consequunti reperibus dia
fugit lanto verumquias eat offictem est, volorrum renectur? voluptaspel iunt.
Ehenet que dia dolupit aut exercie netusda simus volupta Ed minctot aspit, ant omnia eruptas eliquas expla nos ra
cum int. quodit officius cust, simin porro coribusandam iliquas nobis
Gitatis maiosan deriae autecti dustior uptaect atibearume inctem doles ut essitios dus.
voluptiamus aut modit expellorent aut arciliq uossereius, aut Bit ex ea eum rem quod moditint.
iusdam voluptatius nos volor ressum, sin rem inciis saperio. Dignimin nimodi sam rem ni volest velique volorum quos-
Daepti consequibus inciatur, omnitaeri unt et ut exped mos sin ctiissit excerio nsequae estiaectorem rerit qui que pra
nonsed quiditio quis explatiustis rest occuptassit quuntion dolut fuga. Nem sequiamusae rerspie nitatusaped quam
netum eate versper atectatur sum imi, con rest, es rereris di nimodit quatectiatem autest, corem qui ut et, et laceritatiis
consequatate volor alibus sinvenim ra aspid quam dolorepro quam faccus doloreped que cuptatis sim ut eume labor sin
berepro te con coreper spiduciatem eicimende non re cum- esciusandent aut lanis ex estibusae consenis porest ut mint
quisit, quassint pratia pore, verupta quiate con none abo. voluptur, sus dese non con none dest que veles quuntius

Jaimie Chung
Nam nullam escipsum, opturit harit volum es esto exceped
ut etureic aernates adiae aut od mod maxim aut que int. Headline can be two line if possible would make it
Picatquid eatem facerunt eaque veligen tiatemp orporiam look better I think
Pressim
r o dutuveniam,
ct M a n a grerovid
volorectae e r , elluptate
G r a quunditatia
ph Data S aut
rae c i quae
enc e quam harcia quibus doluptas mollorio
nobiti
nonet ventiorro vitatempos si ommodi offic tem volore con- vellaborepra vella doles dem doluptaspel eossita tendiae
sent, que explant fugit maiosae et arum, omnia quae non- estios magnatet liquibus autatiis exerrum dolupta voluptasi

Dr. Phani Dathar

Data Science Solution Architect

2
Helping the world
make sense of data

Problem them around on the web.

Given the intricacies of recommender systems and

Recommender systems are ubiquitous and essential the risks of serving bad recommendations, it is
for increasing customer satisfaction and accelerating necessary to leverage advanced analytics to ensure
business growth. Recommendations powered by that customers receive the most relevant and timely
machine learning have come a long way since recommendations. With Neo4j’s Graph Data Science
collaborative filtering, the likes of which Netflix utilized (GDS) framework, you can use the connections in
in the 2000s. Now, companies across industries filter your data to power recommendation engines that
based on content, demographics, products, and more. fuel customer satisfaction and business growth,
Moving beyond generic recommendations based on gaining otherwise unattainable insights from the
macro level insights, personalized recommendations relationships that exist in the data you already have.
are becoming especially relevant to ecommerce and
retail businesses with large customer bases and
product portfolios in the millions.
Data Model
Companies must be concerned not only about the
quality of their recommendations (are we suggesting One common use of recommendations is product
the right thing?) but also about how quickly they can recommendations based on user behavior and
derive relevant recommendations and serve them to product ontologies. Users could take multiple actions
their users (are we suggesting the most up-to-date on a website: browse through navigation for product
thing?). No one likes seeing an advertisement for a categories, view products of interest, and add
pair of shoes they bought three weeks ago following products to their cart.

A graph data model for recommendations

2
Helping the world
make sense of data

This data can be coupled with product information • Community detection algorithms, like
based on ontologies: products with similar uses, of Louvain, for customer segmentation. Rather
the same type, or complementary products. While than profiling users based on demographic
this example applies to ecommerce, the same data, identify relevant peer groups based on
could be said for a media publisher that wants to buying behaviors and actions. These same
recommend content or a SaaS company that wants algorithms can identify commonly co-purchased
to recommend additional product features.
items across multiple transactions or users,
In the data model shown, a business could use query- improving standard market basket techniques.
based logic, use graph algorithms, or implement
• Centrality algorithms, like PageRank, use network
supervised machine learning techniques to serve a
structure to identify important items, so you can
promising recommendation to the user.
understand which product is most relevant – or
Graph-based approaches enable you to power which items to promote to increase total sales
recommendations that take into account not only
domain-specific knowledge of your product offering • Similarity algorithms, like Node Similarity and
but also user-driven actions and behavior, leading KNN, use graph structure to find similar items
to more accurate recommendations with higher – exploiting the network structure of actions
conversion rates. across a dataset to make recommendations.
Together, these techniques yield powerful
Solution recommendations that consider both a product’s
characteristics and user buying patterns.
Neo4j’s Graph Data Science framework offers a
variety of analytical approaches to make relevant Supervised Machine Learning
recommendations based on the relationships within
your data, ranging from localized query patterns to Machine learning on a graph encompasses two
machine learning–based insights. aspects: how you represent your data, and the
predictions you want to make with it.
Queries To characterize your data, graph embeddings learn
to encode your graph in a simple, machine-readable
Cypher is a powerful, intuitive, graph-optimized query
format – specifically tuned to solve your problem.
language. For example, a content provider could They simplify all the rich information in your full graph
conclude that users who are interested in content into a simple descriptor for each item or customer.
about business are also likely to be interested in That description then feeds into downstream
economics. A simple recommender system could machine learning tasks.
therefore include a Cypher query that serves articles
from the economics section over content from less Techniques like link prediction and node classifica-
relevant sections. tion can use graph embeddings to predict changes
to the structure of your graph. You can train a model
to predict which item a customer will buy (future re-
Graph Algorithms lationships forming), or to identify which customers
Neo4j Graph Data Science offers out-of-the-box graph might churn (future labels).
algorithms for similarity and community detection.
These algorithms are useful on larger datasets
where it’s hard to know exactly what you’re looking
Results
for. These techniques often identify candidates for Graphs enable recommendation engines to leverage
recommendations that are not obvious with queries the most powerful predictor of future behavior:
alone. Some examples of useful algorithms include: relationships. Better recommendations mean a more
personalized experience, higher conversion rates,

3
 Helping the world
make sense of data

and even the ability to forecast trends to manage interactions with product taxonomies, products,
inventory. and search queries. Identifying users with similar
behavior solves for users with no historical data,
While graph-based recommendations created using and identifying similar products increases coverage
manual Cypher queries may be a huge step forward for new products. They also use node embeddings
in user experience for many, large volumes of data that account for complex relationships between
and uncertainty around user behavior mean they users, sessions, products, and taxonomies to make
may not be the final solution. Graph algorithms and recommendations for unusual and ambiguous
graph-based machine learning let you identify the search terms. Embeddings are critical to improving
overall search relevance, ranking, and quality, even
important patterns in big data, and predict unusual
compared to state-of-the-art recommendation
or unexpected future behaviors. Accurate, timely
systems.
recommendations are crucial in driving conversions
and, ultimately, business growth. With Neo4j GDS, they are able to incorporate the
connections in their big data that are crucial to
understanding their customers’ needs, and provide
novel and relevant recommendations where even
Customer Spotlight: Top the best recommendation systems fall short.
Retailer As a result, customers receive timely and helpful
recommendations, which leads to more purchases
and return visits.
A top retailer with over $100 billion in yearly
revenue relies on Neo4j to power their product
recommendations. The retailer’s portfolio has millions
of products that fall into complex, overlapping Conclusion
product taxonomies, and their customers are
increasingly expecting near natural language Neo4j’s Graph Data Science framework enables you
search capabilities. It was essential to go beyond a to make sense of your data –- at scale.
traditional recommendation engine.
No matter whether your data challenge is
The main challenges were the volume and variety of streamlining your recommendations strategy,
product and user data – which were not connected keeping readers onsite, or bolstering your search
in the relational world – and the limited ability to experience, using Neo4j Graph Data Science for
improve search relevance and product ranking recommendations provides fast and accurate
based on sparse and heterogeneous data. Their results. Recommendations is just one of the many
previous solution failed when there was not enough use cases enabled by graph data science.
data about the user or when new products were
introduced into the portfolio. Learn more about Neo4j’s Graph Data Science
framework at neo4j.com/graph-data-science or get
To solve this, the retailer uses Neo4j’s Graph Data started right away with a Neo4j GDS Sandbox.
Science framework to derive similarity relationships
on a graph that connects two years of customer

Neo4j is the world’s leading graph data platform. We help organizations –

including Comcast, ICIJ, NASA, UBS, and Volvo Cars – capture the rich context Questions about Neo4j?
of the real world that exists in their data to solve challenges of any size and
scale. Our customers transform their industries by curbing financial fraud and Contact us around the globe:
cybercrime, optimizing global networks, accelerating breakthrough research, [email protected]
and providing better recommendations. Neo4j delivers real-time transaction neo4j.com/contact-us
processing, advanced AI/ML, intuitive data visualization, and more. Find us at
neo4j.com and follow us at @Neo4j.
4
© 2021 Neo4j. All rights reserved.