TinVec is an approach used by Tinder to provide personalized recommendations based on user embeddings. It maps each user to an embedded vector representing their characteristics based on who they swipe on. Vectors for similar users are clustered closely together. To recommend new users, it calculates a preference vector for a user based on who they liked and recommends those near the preference vector. TinVec achieves over 90% accuracy in predicting swipes and will be used to introduce new product experiences at Tinder to present users others are likely to like.
3. Overview
● Personalized Recommendations and why they matter
● The TinVec approach
○ Why choose + how to obtain user embedding?
○ How to leverage user embedding to provide match
recommendations?
○ Samples from TinVec results
● Evaluation
● Conclusion and Future Product Implementation
4. Personalized Recommendations
● Today we have personalized experiences using social networks , eCommerce
platforms or entertainment services
● Goal: to improve Tinder user’s experience
○ Each user has his/her own tastes (like, pass)
○ Personalized recommendations => users seeing relevant profiles
○ Better user experience: increased and improved matches and messages
5. Personalized Recommendations at Tinder
● Collaborative filtering
● Content-based filtering
○ Natural Language Processing - Bios
● TinVec
○ Utilizes swipe information
○ Users are represented as vectors in an embedding
space
○ Neural-network-based approach
7. TinVec Mechanics
● Users: swipers and swipees
● Each swipee is mapped to a vector
○ Embedded vector in an embedding space
● The embedded vector represents possible characteristics of
the swipee implicitly
○ Activities: playing football, surfing
○ Interests: whether they like pets
○ Environment: outdoors vs. indoors
○ Chosen career path: whether they are software
engineers or medical doctors
● Close proximity of two embedded vectors indicates
○ The swipees are similar => share common characteristics
● Goal: Recommendation
○ Identify more users whom you are likely to swipe right on
Sarah
8. TinVec and Word2Vec
● What is an embedding?
○ Vector representation of entities in the latent space
○ “Similar” entities are mapped to nearby points
● Why?
○ Represent entities more efficiently (~Tens or hundreds v.s. ~millions)
○ Useful for many tasks
■ NLP, recommendations
■ You can do calculations on them!
Goal (output) Property Training Training data
Word2Vec
(Mikolov et
al., 2013)
Word
embedding
Words share common contexts are
closer in the vector space
Neural
Networks
Large corpus of
texts
TinVec User (Swipee)
embedding
Swipees share common
characteristics are closer in the
vector space
Neural
Networks
Large amount
of co-swipes
12. How to Obtain The User Embeddings
INPUT PROJECTION OUTPUT
Target: Bob Context: Alex & Charlie
13. Clusters in the Embedding Space
A point:
A swipee’s embedded vector
in the latent embedding space
Close proximity:
Similar users
(who are co-swiped by many
swipers)
15. How Do We Recommend from the Embedding Space?
Preference vector
1. Josh’s preference is
represented by the mean
embedded vectors of his
likes
2. Users with close proximity
to the preference vector will
be recommended to him
Debbie
16. How Accurately Can You Predict a Swipe Left or Right?
● Area under ROC = 90%
● F1 = 85%
TinVec
● Receiver Operating Characteristic
Curve)
● TPR = Recall
● FPR
● Precision
#Correctly_Predicted_Likes
#Total_Real_Likes
#Incorrectly_Predicted_Likes
#Total_Real_Passes
#Correctly_Predicted_Likes
#Total_Predicted_Likes
18. TinVec + New Product Experiences
● Goal:
○ Use machine learning to present users that we are confident swipers will
like - in a fun, spontaneous and engaging way
● Will roll out slowly first to maximize quality
19. Conclusion
● Personalized Recommendation matter at Tinder
● TinVec: A new personalized recommendation approach
○ Based on the user embeddings
○ Simple input data: only swipes (no user profile data)
○ Training using neural networks
● Clusters show meaningful set of users that share common characteristics
● Swipe prediction achieved high accuracy
● Serves as the foundation for building new user experiences at Tinder