Customer Segmentation: Seeing the Trees and the Forest Simultaneously
- 1. Enhancing Customer Segmentation with Ranking Scores and RFM Insights Table of Contents: 1. Introduction - Seeing the Trees and the Forest Simultaneously: Limitations of traditional clustering and the need for ranking- based insights. 2. Traditional Segmentation via RFM: Overview of Recency, Frequency, Monetary, and its extended variants (RFM-L, RFM-P, RFM-S). 3. New Segmentation Method via Customer Ranking Scores: Introducing ranking scores within segments for influence measurement. 4. Overlaying Scores on Traditional Clustering: Enhancing cluster plots with ranking overlays for influence interpretation. 5. Interpretation of Segments Using Ranking Scores: Insights into segment personas based on ranking-derived traits. 6. Customer Segmentation Workflow: Modular steps including clustering, scoring, and interpretation. 7. Conclusion - The Definitive Edge of Ranking-Enhanced Segmentation: Strategic value, personalization, and model dependency of ranking scores.
- 2. Multiscale Customer Segmentation: Seeing the Trees and the Forest Simultaneously • Traditional clustering methods can't rank customers by importance, either individually or within segments. As a result, all segments (or all customers within each segment) are treated equally. • Also the traditional segmentation does not provide the analyst with insight into the underlying mathematical relationship between inter-cluster separability (the distance between cluster centers) and intra- cluster cohesion (the closeness of points within a cluster). • An example, the traditional clustering scatter plot on the right shows 3 segments (C1, C2, C3), but their relative influence isn’t known.
- 3. Traditional Segmentation via RFM (Recency, Frequency, and Monetary) • Recency: Measures how recently a customer purchased. Recent purchases indicate higher engagement. Helps identify active customers. • Frequency: Tracks how often customers make purchases. Higher frequency suggests loyalty. Identifies repeat buyers. • Monetary: Evaluates total spending by customers. Higher spending indicates greater value. Pinpoints high-value customers. • Other RFM Extensions: RFM-L, RFM-P, RFF-S, etc. • RFM-L: Includes ‘length’ of customer engagement, like time since the first purchase. This effectively distinguish between the short-/long-life customers. • RFM-P: Includes ‘product’ specific metrics, like variety of products purchased or category preferences. • RFM-S: Integrates customer ‘satisfaction’ metrics, like NPS or survey to evaluate customer experience.
- 4. New Segmentation Method via Customer Ranking Scores Members (customers) within each segment have different scores relative to one another, as shown by the upward trend of each segment’s score line in the navy blue curve below. The x-axis represents Customer IDs (CIDs). All CIDs within each yellow horizontal dashed line segment belong to that segment. For example, Segment 1 contains 39 members with a score of 89.58 corresponding to the CIDs beneath it. Each segment’s score can be calculated by taking the average of its members' scores, allowing their importance or influence to be ranked. The respective scores for Segments 1, 2, and 3 are 89.58, 102.32, and 140.37; therefore, Segment 3 is the most influential.
- 5. Overlaying Ranking Scores on Traditional Clustering Methods The scatter plot on the left displays 3 clusters (C1, C2 and C3) representing data segmentation. Clustering reveals groupings but not their influence, and the importance of individual cluster members remains unknown. The scatter plot on the right zooms in on clusters C1 and C3, overlaying their members’ ranking scores. With average scores of 92.07 for C1 and 101.49 for C3, it shows that C3 is more influential than C1. The customer segmentation via ranking scores allows one to see the importance of the segments and also of the customers within each segment simultaneously.
- 6. Customer Segmentation Interpretation via Ranking Scores: Segments 1, 2 and 3 Segment 1 Interpretation: “Long-term, high-value, and frequently purchasing loyal customers”. This interpretation highlights their longevity, significant spending, and consistent purchasing behavior, indicating a strong and valuable customer segment. Segment 2 Interpretation: "Highly frequent, valuable, and established, but less recently active customers“. These customers, with a long history and high spending, purchase infrequently and not recently. As dormant but valuable past customers, they offer a significant opportunity for re-engagement due to their high monetary and length scores. Segment 3 Interpretation: "Long-term, highly valuable, but infrequent and dormant subscribers.“ This segment of long-term, high-spending customers who rarely purchase and haven't bought recently represents a valuable opportunity for re-engagement due to their past high value, despite low recency and frequency.
- 7. Customer Segmentation Workflow The 'Customer Clustering' step (indicated by the green lines) is optional. It is only used if the clustering number 'k' needs to be automatically determined. However, if 'k' is user-selected, this step can be skipped. The workflow supports RFM (or its variants) or non-RFM tabular data with mixed-variate (heterogeneous) features. The 3 outputs produced are: Customer Individual Scores, Segment Individual Scores, and Segment Interpretations. The Segmentation Interpretation (i.e., Output 3) phase can use an LLM by inputting the feature importance or ranking scores for automatic interpretation; otherwise, the user has to perform the interpretation manually.
- 8. Conclusion: The Definitive Edge of Customer Segmentation with RFM and Ranking Scores Goes Beyond Basic Segmentation: Combining RFM with ranking scores quantifies customer and segment importance simultaneously, addressing the "unknown weight" in traditional clustering. Actionable, Targeted Strategies: Pinpointing each customer's influence and value allows for highly personalized and effective marketing, from re- engagement to loyalty programs. Scalable Workflow: The approach is scalable and can be frequently retrained to adapt to rapidly changing customer behaviors. Ranking Score Dependency: The ranking scores depend on the models (and their respective parameters) used in the pipeline, i.e., data imputation, normalization, dimensional reduction, matrix ranking, sorted score segmentation, etc. If the models (or parameters) change, the relative scores will slightly change.