Artem Kozhevnikov, lead Data Scientist, will present some quality metrics commonly followed @ tinyclues in order to evaluate the model predictive power. Those metrics are going beyond well known technical metrics like AUC or RMSE and seem to be important in the context of CRM campaigns targeting.
2. SAAS
SOLUTION
USING
FIRST PARTY DATA
COMPATIBLE WITH ANY
MARKETING STACK
DESIGNED FOR
MARKETERS
TINYCLUES IN A FEW WORDS
DEEP AI FOR
CUSTOMER ACTIVATION
DRIVES REVENUE AND
ENGAGEMENT
ACROSS ALL
CHANNELS
UP IN 2 WEEKS
60 employees // 30 R&D // Hiring 20
3. YOU DON’T KNOW WHAT YOUR CUSTOMERS ARE INTERESTED IN TODAY
INTENT-DRIVEN MARKETING
YOUR STRATEGIC OPPORTUNITY
TO DRIVE REVENUE FROM ALL YOUR CUSTOMERS
IS UNDERSERVED BY YOUR MASS TARGETED CAMPAIGNS
1%
§ ONSITE RECOMMENDATIONS
§ REMARKETING MESSAGES
§ DISPLAY RETARGETING
Works well for recent visitors, but is rapidly
repetitive and inefficient
99%
?
6. Comments :
• This is topic centric formulation (unlikely for onsite recommendation system)
• Need to score all users, in particular, those without recent activity
• In reality, our goals are more complex, we follow various campaign related ROI metrics :
• CTR, CR, Opt out, Attributed Revenue, ...
Predictive Problem
Given a Topic, for each user u ∈ Users we want to build a predictive score
such that users with higher score will have higher probability of conversion
(buying) after receiving a communication about Topic through a given
channel (like Email, Notifications, Facebook Custom Audience, ...).
8. • A. has only Implicit Feedback (only positive) information
• To set a (binary) classification problem for A. we need to define a contrast, or negative
response.
• There is no canonical definition for negative response :
• u ∈ User at random ?
• user u ∈ User that bought some other products ?
• …
• Scores for A. problem are not calibrated
• For B. and C. we can use Explicit Feedback, so their scores are calibrated
• Collecting robust feedback takes several days (delayed response)
• You need to implement explore/exploit strategy to have more efficient learning for C.
3 models
9. Data Base
Impression Click
Channel Data
Relational DataBase
(Simplified Schema)
User
Campaign
Optout
Product
Purchase
Campaign
Attributes
Product
AttributesPageview
WebSearch
Add To
Cart
Attribution
Rules
12. Unsupervised Feature propagation
Multi-layer Unsupervised
Module
Multidimensional sparse
tensor (DataBase)
Asynchronous, daily updates
Raw sparse features
Scoring
A.
Scoring
B.
Scoring
C.
Latent Features
Bank
Cold Start
Features
Warm Start & Channel
Specific Features
Scoring Micro
Services
13. • Train/Test AUCs at different points of pipeline
• Robustness
• Aggregations (average, min, max) of AUCs over most frequently used Topics
• Pre/Post campaign evaluation (“in time” generalization robustness)
• Accuracy/Recall at extract point
• NLL, RMSE
Predictive Metrics : AUC
14. • Calibration ratio = sum(observed) / sum(predicted)
• Works only for calibrated scores
• Monitor calibration ratio over different Topics
• Independent of features engineering
• Predictive Debug : simple method to see how well a feature is taken into account by model
• In Pre/Post campaign shoot
Predictive Metrics : Calibration
Age >=
40
nb_message nb_clickers CTR sum(proba_is_clicker) mean(proba_is_clicker) Calibration
ratio
True 1304544 29350 2.25% 20645.54 1.58% 1.42
False 701544 21747 3.1% 30904.78 4.4% 0.7
All 2006088 51097 2.55% 51550.32 2.57% 0.99
20. • Long term CRM planning optimization
• Automatic Predictive Setup
• Large scale industrialization of the predictive modules and tools
Next challenges