AI/ML to achieve Customer Segmentation

Modern customer segmentation strategies/techniques

• Personalization in Recommendation
• Using patterns of activity to define segments
• Tying multiple traditional methodologies together for better customer segmentation
• Building a customer persona
• Online customer segmentation

Data sources

Limitation of traditional data sources:

• Survey based: 
  • Limited Data
  • Limited contextual understanding
• Siloed Data

Newer data sources:

• Internal Data Warehouses
• Social Media Platforms
  • Social Media Advertising
  • Social Media Analytics
• External Web Tracking Tools (Pendo and similar tools)
• Service/Product Usage (Internal database/warehouses)
• Web Scraping (Post Segmentation – eg. Spotify)
• Freely available data (Government provided et. al.):
  • Demographics
  • Profession
  • Industry Type 
  • (Helpful in the starting phase of company to understand the market)

Clustering:

• When we talk about segmentation, the very first thing that comes to our minds is AI based clustering
• AI models enable clustering based on similar traits/similar characteristics but with softer boundaries as opposed to the rule based models
• Clustering forms the first step to personalization 
• Ecommerce example:
  • Most commonly used technique:
    • RFM clustering ( recency/frequency/monetary value based segments created)
  • Limitations of RFM:
    • RFM technique is used mostly on transactional data
    • Parameters such as  age, gender, region etc. of the customer are not tied to the transactional data
    • Customer based data points like Usage / trend / seasonality / calendar and impulsive buys not taken into consideration
  • Churn/Engagement scores as an additional input to the clustering algorithm
    • Identify accounts with high propensity of churning (score them)
    • Behavior (usage) – understand the engagement of a customer wrt platform and provide scores
    • These Propensities/probabilities can be used as features for Segmentation
    • Usage – data source examples GA, Pendo, Internal Data warehouses

ML Techniques used for clustering

Unsupervised learning is generally used for clustering

• K-means 

• • • • Define the number of clusters (K)
      • Forms clusters after computing sum of mean of squared distances between centroids and the data points
      • Optimize the position of centroids to get this sum to minimum
      • Computationally less expensive and easier to build

• DBscan

• • • • Density based clustering
      • No need to define the number of clusters (unlike K-means)
      • Handles outliers quite well
      • Computationally expensive; Takes a lot of time for execution
      • If the density varies across the data, then the clusters may not form properly
      • Can’t handle data with higher dimensionality

• GMM

• • • Based on 
      • Mean – defines the centre
      • Covariance – defines the width
      • Mixing probability – defines how big or small the cluster will be
    • Can form clusters of different shapes, and not just circular like K means
• SOM/PCA
• • • • Techniques used for dimensionality reduction
      • The output from these techniques can be subsequently used as input to the clustering algorithms

Use Cases for clustering

• • Once clusters are created, you can devise growth/marketing strategies around each cluster based on their common characteristics 
  • Types: Retention /Growth/ Upsell/ Xsell/ Ad-targeting
    • Retention – High value customers
    • Growth – Low level engagement + prospects – push them to the other side of the fence – eg: Ad targeting
    • Upsell/Xsell – Customers (medium engagement but scope for increased revenue) with better recommendation (Recommendation System)

Recommendation Systems – Various techniques/approaches

• Collaborative filtering (Cross referencing – pick favourites from one user and recommend to another user belonging to the same cluster)
  • Some of the useful data points for collaborative filtering for a platform such as Spotify:
    • Genre
    • Artist
    • Preferred Language
• Content Based filtering using  NLP for a platform like Quora
  • Word embeddings – convert words to vectors
  • Similarity between words (distance between the vectors)
  • Doc2vec (Finding similarity between documents)
  • Topic modelling on the document
  • Contextual similarity – similarity of the content falling within a certain topic

Case Study:Spotify B2C case case study 

Tune in to our podcast here.