Understand your customers deeply ... Engage with a personal touch! #ItsPersonal Debdoot Mukherjee Personalization & Customer Insights @
Aug 17, 2015
Understand your customers deeply ... Engage with a personal touch! #ItsPersonal
Debdoot Mukherjee
Personalization & Customer Insights @
Modeling
Personalized Customer Engagement
Data Driven Retail Functions
Product Listings Handpicked For Me
Notifications Fashion Feed
Offers & Promotions
Re-targeting
Marketing Research
Campaign Targeting Audience Monetization
Category Planning
Brand Benchmarking
Content Design
• Ephemeral and non-identifiable items unlike Books and Movies – Extremely sparse user/item matrix –Bias of products with higher inventory – Exploration versus Exploitation trade-off
• Diversity and Serendipity • Closest domain: News Articles
Fashion – What s diffe e t?
• Recommend based on user profiles stored as preference / weight vectors on item features, learnt from relevance feedback on items.
• Good vector representation for items? – Bag of product attributes does not work! Too many features,
s a e o se histo y fo a si gle use , so p ofiles do t generalize.
– Learning preferences along latent factor / topical dimensions or product groups (clusters) helps.
• Maintain two user profiles: long term (taste), short term (intent) – Incorporate time decay into browse history – Degree of personalization depend on the amount of browse data
• BUT, att i ute ele a e does t suffi e. The e is so ethi g e a t aptu e ia att i utes a out so e fashio ite s
that make them popular, others not. May be aesthetics.
Feature / Content based Approach
• Simple user-user, item-ite CF te h i ues do t o k ell e ause of extremely sparse user-item matrix
• Matrix factorization:
• Regularization is tricky and severe cold start. In practice, models are trained specific to each category of product, so maintaining separate models for cold start and warm start becomes difficult.
• Recent advances in Feature based Matrix Factorization address this - SVDFeature, Factorization Machines, RLFM, FOBM, fLDA, UFSM …
• Train model on snapshot of active products for recommendations
Collaborative Filtering Approaches
• A good vector representation for items would make si ila ite s eigh o s i the e to spa e. #di e sio s
should be not very high. • Co-browse of items in a session is (weakly) indicative of
si ila ity . “u h a sig al ei fo ed a oss a y sessio s becomes strong.
• Inferring substitutable and complementary products – Leskovic et al. KDD15 – Train a logistic regressor with features defined on the similarity
of item vectors represented as topics to predict whether two products are similar. LDA using the analogy (Item Æ Document, Item Attribute Æ Word)
– Core Idea: Joint training of logistic regressor and item topics by simultaneously optimizing both topic distributions and logistic parameters to maximize the joint likelihood of topic memberships and product similarity.
Vector Representation for Items
• We use this analogy so that existing models for finding representations in text / IR become applicable: – Browse Session Æ Document, Items Clicked Æ Sentences, Item
Attributes Æ Words • Evaluate LDA, Word2Vec, GloVe …
– Yields varying levels of topical and functional similarities along dimensions of the item vector • “ea h fo si ila te s fo nike :
– Topical Similarity: adidas, puma, sports, dry-fit, polyeste … spo ts elated te s – Functional Similarity: adidas, puma, fila, merrell, hrx …
– Mining interesting relationships between entities of interest viz. brand, price band, pattern, item collection etc.
• Spherical clustering to create product groups – a better unit of analyses than individual products.
• Create user profiles by aggregating their preferences on such item dimensions and product groups across all browsing sessions.
Vector Representation for Items (2)
• Explore/Exploit trade off – Popularity scoring of items (normalized for each
category / product group) • Use Thomson Sampling in a context free bandit formulation
that assumes Gaussian reward (CTR) • Adjust CTR with rank to formulate reward
– Contextual bandits can help in choosing the right recommendation strategy given page and session context
• Use of LSH to ensure diversity of recommendations
Explore / Exploit, Diversity ..
Handpicked For Me
Personalized page with different kinds
of recommendations: • Taste, Intent based • Cross sell based on
last purchase
Customer Insights Platform A platform to slice / dice mined customer profiles has over 50K
different dimensions
Create a segment of loyal customers in Delhi who wear heels
Affinity toward heels
Highly loyal
From Delhi
Delhi women have a greater affinity for taller heel heights than Chennai women
A woman from Delhi is 2x more likely to be interested in stilettos than someone from Chennai
Brand A
Brand C Brand B
48.5% - 26 yr
27% - 28 yr 15% - 26 yr
1.8 % - 29 y
Loyalist Distribution
0.4% - 28 yr
Comparing 3 Men Shirt Brands and their loyalists
Compared to an average Myntra customer
What else do the loyalists buy?
Less likely
More likely
Personalization Services Myntra.com
Event Distributions
Event Processor
S3 Event Storage
Cassandra (Clickstream Aggregates)
Model Training
Serving Caches
Near-line Personalization
Online Personalization
Customer Profile ETL
Customer-wise Event Aggregator
Mongo (Customer Insights)
Insights Platform
Architecture
Transactional DW
Product Knowledge Graph
Ariana Grande
M perso al st le is a i ture of, like, girl , throwback, like retro '50s pin-ups, floral, like hippies, like a thi g fe i i e, a d like flirt .
Perso al “t le is about having a sense of yourself what you believe i ever da
Ralph Lauren
Ever o e looks at your watch and it represents who you are, your values and your personal st le
Kobe Bryant
And You Still Think I Would Know About Personal Style ?!!
Read more at http://sartorialscience.myntrablogs.com