💻 computer science

An Analytical Multiple Criteria Framework for Temporal and Dynamic Business-to-Business Customer Segmentation in Manufacturing

This paper proposes and validates a dynamic multi-criteria decision-making framework that extends traditional RFM modeling with stability and growth dimensions, integrated with adaptive analytical processes and graph-based consensus models, to enable robust, time-sensitive B2B customer segmentation in the manufacturing sector.

Original authors: Muhammad Raees, Konstantinos Papangelis, Vassilis Javed Khan

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muhammad Raees, Konstantinos Papangelis, Vassilis Javed Khan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a massive warehouse that supplies parts to thousands of different factories. You have a giant ledger listing every single order these factories have ever placed.

The Problem: The "Snapshot" Mistake
Most companies look at this ledger like a photographer taking a single snapshot. They ask: "Who bought the most? Who bought the most recently? Who spent the most money?" This is called the RFM model (Recency, Frequency, Monetary).

In a simple store selling to regular people (B2C), this works fine. But in the business world (B2B), it's like judging a marathon runner by only looking at where they are right now. It misses the whole story. A factory might buy a little bit today but has a contract to buy millions next year. Another might buy a lot today but is about to go bankrupt. The "snapshot" misses the stability and the growth potential.

The Solution: A "Movie" Instead of a Photo
The authors of this paper propose a new way to sort these customers. Instead of a photo, they want to watch a movie of the customer's relationship over time.

They built a framework that does three main things:

Expanding the Lens (The Criteria):
They didn't just look at money and dates. They added two new "lenses":
- Growth: Is the customer buying more products over time? Are they expanding their order volume?
- Stability: Is the customer reliable? Do they buy consistently, or do they vanish for months and then return?
- Analogy: Think of it like dating. You don't just look at how much someone spent on dinner last night (Money). You look at how often they call (Frequency), how long you've been together (Recency), and whether they are becoming more committed over time (Growth/Stability).
The "Expert Panel" (The Weights):
Not all factories are the same. For one company, "Growth" might be the most important thing. For another, "Stability" is key.
The authors used a method called AHP (Analytic Hierarchy Process). Imagine a panel of expert sales managers sitting around a table. They argue and vote on what matters most. "Okay, for this strategy, we care 40% about growth and only 20% about stability." The system then adjusts the math to match their specific goals.
The "Double-Check" (The Consensus):
The system runs two different types of analysis:
- Time-Series Clustering: It looks at the movie of the data to see how customers move over time.
- Stability Modeling: It checks how "wobbly" a customer is. Do they jump between groups, or do they stay put?
Then, it uses a Graph-Based Consensus Model. Imagine two different detectives investigating the same case. One says, "This guy is in Group A." The other says, "No, he's in Group B." The system looks at their notes, weighs their opinions, and finds the "truth" that satisfies both. This ensures the final groups are solid and not just a fluke of one specific math formula.

The Results: Sorting the Warehouse
They tested this on a real manufacturing company with 3,458 customers.

The Old Way (Standard RFM): Grouped customers based on simple math.
The New Way (Their Framework): Grouped customers based on the "movie" of their behavior and the expert panel's priorities.

The new method found that the old way was missing the nuance. It successfully sorted the customers into four distinct "personas":

The High-Frequency Strategists: They buy often but in small amounts. They are active but not super loyal.
The High-Spend Giants: They buy less often, but when they do, it's huge. They are very loyal and profitable.
The Dabblers: They buy a little bit of everything but rarely. They are hard to pin down.
The Volume Workers: They buy massive amounts but don't make the company much profit. They are reliable but low-margin.

Why It Matters
The paper claims this method gives business leaders a "superpower." Instead of guessing which customers to treat like VIPs, they can use this dynamic, time-based map to see who is actually growing and who is stable. It helps them decide who gets a dedicated account manager, who gets a discount, and who needs a nudge to buy more.

In short, they moved from taking a still photo of a customer to watching a full movie of their relationship, allowing the business to make smarter, more stable decisions.

Technical Summary: An Analytical Multiple Criteria Framework for Temporal and Dynamic B2B Customer Segmentation

1. Problem Statement

Customer segmentation is a critical tool for optimizing business strategies and supply chain management. However, traditional implementations in Business-to-Business (B2B) manufacturing contexts often rely on static, limited criteria, primarily the Recency, Frequency, and Monetary (RFM) model. While effective in Business-to-Consumer (B2C) settings, standard RFM models frequently fail to capture the complex, long-term, and dynamic nature of B2B relationships. Existing temporal segmentation approaches often treat customer dynamics as static snapshots or rely on aggregated features, overlooking variability, stability, and growth trajectories over time. Furthermore, there is a lack of frameworks that integrate Multi-Criteria Decision Making (MCDM) with time-series analysis to allow domain experts to adapt segmentation strategies based on specific business objectives and preferences.

2. Methodology

The authors propose a dynamic, adaptive MCDM framework designed to segment B2B customers by integrating RFM with dimensions of growth and stability. The methodology follows a five-step process:

2.1. MCDM Data Model and Feature Engineering

The framework extends the traditional RFM model by incorporating:

RFM Dimensions: Recency, Frequency, and Monetary value.
Growth Dimensions: Product diversity (mix), volume/quantity, and monetary growth (profits).
Stability Dimensions: Loyalty (years since first order), inter-purchase volume, and profit margins.
These features are derived from transactional data to capture long-term commitment and behavioral fluctuations.

2.2. Preferential Feature Weighting (AHP)

To align the model with specific business goals, the authors employ the Analytic Hierarchy Process (AHP). This allows domain experts (e.g., sales professionals) to assign relative importance weights to the various criteria through pairwise comparisons. An inconsistency check (Consistency Index < 0.1) ensures the reliability of these weights. This step makes the segmentation adaptive to decision-maker preferences rather than relying on fixed, static weights.

2.3. Time-Series Clustering

The framework utilizes multivariate time-series clustering to group customers based on their behavioral trajectories over time.

Algorithms: Hierarchical (agglomerative) and Spectral clustering are used, chosen for their efficiency and ability to handle diverse variable types compared to K-Means.
Distance Measures: Specialized time-series distance metrics are employed to estimate dissimilarity between customer trajectories:
- Complexity-Invariant Dissimilarity (CIDS)
- Dynamic Time Warping (DTW)
- Temporal Correlation Coefficient (CORT)
Validation: Clustering quality is evaluated using Silhouette and Calinski-Harabasz indices.

2.4. Temporal Stability Modeling

Beyond initial clustering, the model quantifies customer stability over time by analyzing:

Transitions: The frequency of customers switching segments.
Volatility: Defined as the number of times a customer changes segments divided by total time steps.
Continuity: The duration a customer remains in the same segment.
A "Segment Score" is calculated using preferential weights ( $\alpha, \beta, \gamma$ ) to balance continuity, volatility, and transitions, identifying stable customers versus those with high drift.

2.5. Graph-Based Consensus Modeling

To synthesize the results from the time-series clustering and the stability modeling, the authors apply a consensus clustering approach.

A similarity graph is constructed where nodes represent customers and edge weights encode the agreement between the two clustering methods (weighted by user preference).
The Louvain-Leiden community detection algorithm is used to identify dense subgraphs (communities), resulting in a final, robust segmentation that minimizes the bias of any single method.

3. Key Contributions

Extension of RFM to B2B: The paper bridges the gap between standard RFM and B2B requirements by explicitly adding stability and growth dimensions to the feature set.
Integration of MCDM and Time-Series: It introduces a novel framework that combines preferential MCDM (via AHP) with dynamic time-series clustering, allowing for adaptive segmentation that responds to changing business objectives.
Stability and Volatility Metrics: The work contributes specific metrics for tracking customer transitions and volatility over time, addressing the underexplored area of temporal customer stability in segmentation.
Consensus Mechanism: The application of a graph-based consensus model (Louvain-Leiden) to merge time-series and stability perspectives provides a more robust final segmentation than single-method approaches.

4. Experimental Results

The framework was evaluated using a real-world dataset from a global manufacturing company, comprising 3,458 B2B customers and over 5 million transactions spanning five years.

Clustering Performance: Spectral clustering with the CID distance measure and 4 clusters ( $k=4$ ) yielded the highest Silhouette score (0.712), while CORT was optimal for the Calinski-Harabasz index.
Comparison with Baselines: The proposed Adaptive MCDM approach significantly outperformed standard RFM and Fixed MCDM models across all clustering algorithms (Hierarchical, Spectral, K-Means). For instance, the Adaptive MCDM with Spectral clustering achieved a Silhouette score of 0.712, compared to 0.365 for standard RFM.
Consensus Analysis: The consensus model successfully balanced the time-series and stability perspectives. While the time-series model was weighted higher (0.6), the stability model induced notable shifts, particularly in segments C1 and C3 (approx. 11% reallocation), demonstrating the value of incorporating stability metrics.
Segment Profiles: The final segmentation identified four distinct groups:
- C1: High-frequency, recent customers with moderate profits but low loyalty (transactional).
- C2: High-spend, loyal customers with strong profit margins and product mix.
- C3: Low engagement, low value, but high product diversity.
- C4: High volume, low profitability, high loyalty (legacy buyers).
Computational Overhead: The consensus modeling added manageable computational costs (approx. 22.63 seconds for 3,458 customers on a standard CPU), though the authors note that graph construction for very large datasets may require optimization (e.g., sparsification).

5. Significance and Claims

The paper claims that its primary contribution is providing a robust, adaptive toolkit for B2B customer segmentation that moves beyond static snapshots. By integrating MCDM with time-series and stability modeling, the framework enables domain experts to:

Customize Strategies: Adjust feature weights based on specific market conditions and business goals (e.g., prioritizing growth vs. stability).
Capture Dynamics: Gain deeper insights into customer behavior trends, stability, and transitions that static RFM models miss.
Support Decision-Making: Offer actionable insights for sales and marketing teams to formulate targeted strategies, such as loyalty programs for high-value segments or reactivation campaigns for low-engagement groups.

The authors conclude that while the method requires expert input for weighting and feature selection, it offers a significant analytical advantage over traditional, fixed approaches, particularly for manufacturing organizations with diverse and dynamic customer bases. They acknowledge limitations regarding the reliance on transactional data (excluding qualitative insights) and the computational cost of graph construction for massive datasets, suggesting these as areas for future refinement.