Next-best-action: ML-lel a jövő egyedi döntéseinek nyomában · 2019-11-27 · Next Best Action...
14
Next-best-action: ML-lel a jövő egyedi döntéseinek nyomában
Transcript of Next-best-action: ML-lel a jövő egyedi döntéseinek nyomában · 2019-11-27 · Next Best Action...
Next-best-action: ML-lel a jövő egyedi döntéseinek nyomában
Data - Based Data - Driven
It is Recognized
It is Known
It is Understood
No Idea
No Idea
No Idea
Data - Based Data - Driven
Scalable
Automated
Learning from Data
Low
Manual
Human Intelligence
Data - Based Data - Driven
Scalable
Learning from Data
Low
Manual
Human Intelligence
All the same?
2
Decision
Learning
#Decision Capacity
Preference
Decision Making
Channel
Customer Journey & Optimization with NBAIt recommends next actions that help users progress towards business goals as quickly and smoothly as possible
https://blog.datarobot.com/delivering-next-best-action-with-artificial-intelligence
All Customers Receive the Same Outbound Marketing Content
All Customers Within a Segment Receive the Same Marketing Content
With Next Best Action, Every Customer Has A Personalized Journey
Customer Lifecycle
Awareness Consideration Preference Activation UsePurchase Service Advocacy
4
Next Best Action Engine – ML based approach
5
• Optimization objective is typically complex• Historical feedback is typically incomplete – feedback for other
prediction are not available • Historical feedback is skewed towards a small set of actions• Actions are typically dynamic (seasonal changes or business strategies)
Key Challenges
Standard classification or regression modelso Only bandit feedback is availableo Sampling bias presented in historical data can not be
handled properly
Regular contextual multi-armed bandit model o It lacks the ability to model multiple objectives Stage-advancement policy (e.g. from login to complete profile)Goal-oriented policy (e.g. different purchase, becoming superfan)
Hybrid model
Offline Learning Online Learning
Propensity Scorer + Multi-armed bandit model
Intent Predictor Contextual Action EnginePolicy Selector
Problem Setting
6
User Current Context Action Reward Maximized:
Action 1
Action n
Action 2
.
.
.
.
Multi-armed Bandit Model
Advanced Approach
7
Action 1i
Action 1j
.
.
.
Input Intent Propensity Model - Profile affinities, Behavior Pattern
Purchase Decision Point (PDP)
obj 1
obj 2
obj n
Action 2i
Action 2j
.
.
.
Action ni
Action nj
.
.
.
.
.
.
Contextual Multi-armed Bandit Model
Output
0%
20%
40%
60%
80%
100%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
S ample Size
Cumulative Gains Chart
Random Sample Lead model
Point of purchase – algorithmic analytical approach to the scoring model
T + 1 day T + 2 days T + 3 days T + 4 daysToday (T)
0,710,66
0,94
0,78 0,84 0,89
0,77 0,790,54
0,85
0,79
Sales probability Purchase decision - No Purchase decision - Yes
3 mathematical rules and 0,8 the goodness of
probability
70%
% o
f Sal
es
8
0,460,550,69
0,52
Intent Propensity Model
9
• Different visited pagese.g. product, comparison, Check-out stages
• User activity – user selection on pages e.g. colour selector, tariff plan calculation, add to compare, summary details
• Cardinality and density - with dynamic parameterse.g. Total sum of visited product pages (x1, x2) between now and t1, t2 (e.g. t1 := 24 hours and t2 := 36 hours). There is a calculation in every 10 minutes : if x2/(x2-x1) >= p1 (e.g. p1 := 2) is true then visitor gets higher score.
• Events e.g. add to cart, personalized item click
• Solvency test / Willingness to pay - Customer and order check - Information from BE e.g. high risk personal data, backlist, debt
Combined scoring modelUser
behaviourClassic CRM
information* * if possible
Virtual features
e.g. actual tariffplans and
services, TF, HH
unsupervised learning ML
algorithms e.g. SVD
1
2
3
Point of purchase – fundamental elements of the scoring model Intent Propensity
Model
Sales funnel – boosted by physical product
SALES
8x
per
iod
Rateplan / Final Price
~ #10
~ #2
xp
erio
d
Purchase Decision Point (PDP)
The faster the more focused guidanceWin-win situationML Recommendation
+ Retargeting Recommendation+ Lead (in- and outbound)
Visitor Decision Process
Brand pairs – alternative brand sales potential
3 buckets of brand loyalty – the rate of the same brand’s views 15 days before purchase
Findings
10
Intent Propensity Model
WEB
SALES
Analytical lead generating process
11
CHAT
SALES
PDP
inbo
und
WEB
SALES
PDP
inbo
und
Online session w/ sales decision
Online session w/o sales decision
WEB
SALES CHAT
SALES
PDP
outb
ound
CC
SALES
Without online session
Today (T) T + 1 day T + 2 days T + 3 days T + 4 days
+41%
+28%
+13%+23%
Lead efficiency - Conversion Rates by time
Rates of return The scoring model’s painted profiles on the web
T - 2 days
T - 1 day
Today (T)
T + 1 days
T + 2 days
T + 3 days
T + 4 days
Additional 58%
Types of Model Driven Sales
12
Results of the analytical lead model
Benchmark Model Benchmark Model
Cross-session (between T and T+10 days)
In-session (T)
Online Chat – Conversion Rate
+ 58%+ 19%
+68%
Benchmark Model
CC (In-session) – Conversion Rate
Actual Chat Platform
New Chat Platform with
analytical lead module
12
+ 68%
Hybrid model - realization
Hybrid contextual multi-armed bandit model 1. More than regular: purchase decision points are in the model – plus
special features (user intent & interest) different actions are recommended to users that show different interests e.g. different purchase, skipping a few stages for fast shoppers
2. Reward strategy – we don’t reward zero or negative stage progression3. Both policies incorporate Thompson Sampling with Gaussian kernel for
better exploration
13
![The Indologist Indian. In Memory of Ervin Baktay · 3 Homo ludens: Emlékeim nyomában [“Homo Ludens. My Memories”] Bibliotheca Hungarica Artis Asiaticae Hopp Ferenc Kelet -ázsiai](https://static.fdocuments.us/doc/165x107/5e0ae190def69521871f281c/the-indologist-indian-in-memory-of-ervin-baktay-3-homo-ludens-emlkeim-nyomban.jpg)
