Unsupervised Learning Analysis for E-Commerce Orders Segmentation¶

Building on the cleaned e-commerce dataset established in our data preprocessing phase, this report applies unsupervised learning techniques to uncover latent structure within order behavior. Specifically, we pursue two complementary objectives:

1. Customer Order Segmentation via K-Means Clustering — grouping orders by their numerical characteristics (basket value, quantity, freight proportion, and product diversity) to identify distinct purchasing profiles that can inform targeted marketing and logistics strategies.
2. Market Basket Analysis via Association Rule Mining — discovering statistically significant co-occurrence patterns between product categories, payment methods, and order sizes to surface actionable cross-sell and promotional opportunities.

Together, these analyses translate raw transactional data into strategic business intelligence.

Setup¶

The following libraries underpin the full analytical pipeline. pandas and numpy handle data manipulation; matplotlib and seaborn drive visualization; sklearn provides preprocessing and clustering utilities; and mlxtend supports the Apriori-based association rule mining workflow.

Part 1: Data Ingestion¶

We load the consolidated, cleaned dataset produced during the data preprocessing phase — a single order-level file (ecommerce_orders_cleaned.csv) representing the merged and validated output of the original orders, order_items, order_shipping, and payments tables. This serves as the authoritative source of truth for all downstream analysis.

Part 2: Unsupervised Learning¶

Objective¶

With a clean order-level dataset in hand, we apply unsupervised learning to surface hidden structure in purchasing behavior. The goal is to group orders by shared numerical characteristics — order value, item quantity, freight proportion, and product variety — into distinct customer segments. These segments form the basis for targeted promotions, differentiated logistics planning, and personalized recommendation strategies.

Data Preprocessing for Unsupervised Learning¶

Constructing the Order-Level Feature Matrix (all_orders_df)¶

Before clustering can be applied, the dataset must be structured so that each row represents exactly one order. From the cleaned dataset, we extract and derive the following features:

• basket_size — total monetary value of the order
• basket_qty — total number of items
• freight_ratio — the share of the order value attributable to shipping (total_freight_value / order_total_value), a proxy for logistics intensity
• product_diversity — the breadth of distinct product types relative to basket size (num_unique_products / num_items), capturing whether customers shop narrowly or broadly
• top_product_category, payment_type, order_size_category — categorical dimensions used later in association mining

Division-by-zero cases (zero-value or empty orders) are handled explicitly. Remaining missing values in categorical columns are replaced with "Unknown", and numerical columns are zero-filled.

Sampling for Cluster Exploration (orders_analysis_df)¶

To tune and validate the clustering configuration (specifically, selecting the optimal number of clusters) on a representative but computationally efficient subset, we hold out 30% of all_orders_df as an analysis sample using a stratified random split with random_state=37 for reproducibility. The final cluster model is subsequently re-fitted on the full dataset.

Determining the Optimal Number of Clusters¶

K-Means Inertia Sweep (k = 3 to 9)¶

Before committing to a cluster count, we systematically evaluate K-Means across a range of values (k = 3 to 9) on the analysis sample. Only the four numerical features — basket_size, basket_qty, freight_ratio, and product_diversity — are used, as clustering on raw numerical scales would bias results toward high-magnitude features. All features are therefore standardized using StandardScaler prior to fitting.

For each k, we record the inertia — the sum of squared distances between each order and its assigned cluster centroid. Inertia decreases monotonically as k increases; the objective is to identify the point at which additional clusters yield diminishing returns in compactness.

Elbow Method — Selecting the Optimal k¶

The Elbow Method plots inertia against k to visually identify the inflection point at which the marginal gain in cluster compactness begins to plateau. This inflection — the "elbow" — represents the most parsimonious cluster count: enough structure to be meaningful, without overfitting spurious micro-segments.

Based on the plot below, we select k = 4 as the optimal configuration, as this marks a clear change in the rate of inertia reduction before the curve smooths out further.

Cluster Analysis and Visualization¶

Assigning Clusters to the Full Dataset¶

With the optimal cluster count established at k = 4, we re-fit K-Means on the full all_orders_df using the same four standardized numerical features. Each order is assigned a cluster label, appended as a cluster column. Fitting on the complete dataset ensures that all orders — not just the analysis sample — carry a meaningful segment assignment for downstream profiling and recommendation logic.

Cluster Profiling and Visualization¶

Two complementary visualizations are produced to characterize and communicate the resulting segments:

Cluster Profile Heatmap — Computes the mean of each numerical feature per cluster, then re-standardizes these means as Z-scores so that features measured in different units (currency vs. counts vs. ratios) can be directly compared. A diverging colormap distinguishes clusters that score above average (positive Z-score, red) from those below average (negative Z-score, blue) on each dimension, providing an at-a-glance segment fingerprint.

Cluster Size Distribution — A bar chart showing order volume per cluster. This reveals whether the segmentation produces balanced groups or whether certain segments dominate, which has direct implications for how broadly any cluster-specific strategy can be deployed.

Cluster Interpretation¶

1. The heatmap shows moderate-to-strong separation: Cluster 3 is clearly high in basket_size and basket_qty with low freight_ratio, while Cluster 0 shows the opposite pattern (low value/quantity but high freight burden). As for Cluster 2, its samples are less differentiated based on the moderate Z-scores compared to other clusters.
2. basket_size and freight_ratio are the strongest defining features because they have the largest absolute Z-score contrasts compared to other clusters, with basket_qty as a secondary differentiator and product_diversity mainly distinguishing Cluster 1. This is because monetary value and shipping costs represent the primary economic trade-offs in e-commerce; for example, high-value items often require specialized logistics, naturally separating "Premium/Bulky" orders from "Standard" ones.
3. The clusters represent distinct purchasing behaviors, including high-value buyers, low-diversity shoppers, bulk purchasers, and customers paying higher freight costs. These behavioral differences allow the e-commerce business to design targeted promotions and operational strategies tailored to each customer segment.

Market Basket Analysis (Association Rule Mining)¶

Methodology Overview¶

Market Basket Analysis is an unsupervised technique that quantifies co-occurrence patterns across discrete transaction attributes. Rather than predicting an outcome, it surfaces rules of the form "when X is present, Y is also present with high probability" — invaluable for understanding multi-dimensional customer behavior.

In the e-commerce context, this helps answer questions such as:

• Which payment methods are preferred for which product categories?
• Do certain order sizes correlate with specific product types?
• What multi-attribute patterns exist across category, payment channel, and basket size?

Association rules are evaluated along three key metrics:

We apply the Apriori algorithm to the categorical features of all_orders_df to discover frequent itemsets and generate high-lift rules.

Transaction Encoding¶

To apply the Apriori algorithm, each order must be represented as a "basket" of discrete items. We construct a transaction list where each order contributes exactly three tagged items—one per categorical dimension:

• cat_{top_product_category} — the dominant product category
• pay_{payment_type} — the payment method used
• size_{order_size_category} — the order size class (Small / Medium / Large)

The prefixes ensure items from different dimensions remain distinguishable after encoding.

Apriori Algorithm — Frequent Itemsets and Rule Generation¶

Using the mlxtend library, the workflow proceeds in three steps:

1. One-hot encoding via TransactionEncoder — converts the transaction list into a Boolean DataFrame where each column represents a unique item.
2. Frequent itemset mining via apriori() with min_support = 0.01 — retains only item combinations appearing in at least 1% of all orders, balancing pattern coverage against noise.
3. Rule generation via association_rules() with metric = 'lift' and min_threshold = 1.0 — produces all rules where co-occurrence is at least as likely as independent random chance.

Top 10 Rules by Lift¶

Rules are ranked by lift in descending order to surface the most statistically surprising — and therefore most actionable — associations. For reference, rules with lift > 1.5 are considered strong associations; lift > 2.0 indicates a very strong co-occurrence.

Example interpretation: A rule (cat_furniture_decor, size_Large) → (pay_credit_card) with support = 0.042, confidence = 0.756, lift = 1.84 reads: "Customers who purchase furniture in large orders have a 75.6% probability of paying by credit card — 1.84× more likely than the baseline credit card usage rate."

Rule Interpretation and Sensitivity Analysis¶

1. The top three rules suggest that large orders are strongly linked to furniture/decor, while medium-sized baskets (often with credit card payment) are more associated with bed/bath/table purchases, indicating distinct shopping missions by basket size.

2. The strongest association is the size_Large ↔ cat_furniture_decor combination (highest lift), meaning these items co-occur far more often than expected by chance.

3. This likely happens because furniture items are typically bought in bulk home purchases, and paid by credit card possibly by installments or credit card rebates. On the other hand, size_Med + bed/bath/table items are common household replenishment purchases where customers buy in moderate quantities and paid using credit card for convenience.

4. When min_support is reduced to 0.005, more niche and low-frequency rules appear (often higher lift but less generalizable), lowering support captures the "Long Tail" — we see rules for rare items (like musical instruments or art). While these have high Lift, they are less useful for broad business strategy because they represent such a tiny portion of total sales. Meanwhile increasing it to 0.3 keeps only very common patterns and removes many specific and useful associations. In conclusion, it might be ideal to lower min_support to extract more meaningful and insightful patterns from datasets compared to common patterns, depending on the business needs.

Business Recommendations¶

• Prioritize large furniture baskets with payment support: For customers in cat_furniture_decor with size_Large, offer installment/BNPL and targeted checkout promotions to reduce abandonment on high-ticket (expensive) orders to hit credit card rebates.

• Run medium-basket card promotions: Since medium baskets are strongly linked with card payment patterns, launch category-specific credit-card cashback bundles (especially home/bed-bath-related items).

• Deploy rule-based cross-sell recommendations: Use high-lift pairs in cart and checkout ("frequently bought together" by category + size) to increase average order value and conversion.