Skip to content

77.8% Conversion Surge: How Lumière Jewel Eliminated Allergy Returns with Private ESSM

/images/LumièreJewel-hero.jpg?x-oss-process=image/resize,m_fixed,m_lfit,w_300

The Allergy Crisis Crippling Returns

Context: For Lumière Jewel, scaling to $2.5M monthly GMV brought a critical challenge: Material Allergy Concerns: Customers frequently report allergic reactions to nickel and other base metals, leading to high return rates and negative reviews despite using 'hypoallergenic' marketing claims.

At this stage, standard rules failed. The Fashion-forward women aged 18-35 who seek trendy, affordable jewelry pieces demanded relevance. Their self-developed system couldn’t track material preferences, causing 18% return rates and crippling customer trust despite their “hypoallergenic” claims.

From Self-Built Limits to Private AI Power

To solve this, Lumière Jewel deployed WooRec. Note: Depending on their scale, they leveraged WooRec Private Deployment for complete data sovereignty and algorithmic control.

The transformation wasn’t instant. We architected a phased evolution of their recommendation engine:

Phase 1: Expanding the Candidate Pool with Vector Recall

Powered by WooRec Strategy Module

We needed to move beyond simple keyword matching. We implemented a Hybrid Recall strategy:

  • Foundation: We started by ensuring popular items were visible via Hot Retrieval, solving the Cold Start problem.
  • Advanced: Vector Retrieval (Embedding)
    • The Logic: We mapped users and items into a high-dimensional vector space using FAISS/Graph Embedding. This allowed us to capture “latent interests”—finding items that are semantically related, not just textually similar, and crucially, to factor in material properties (like nickel-free) as part of the embedding.
    • The Result: This allowed us to surface items that matched the user’s style and material needs, reducing the chance of allergic reactions.

Phase 2: The Model Evolution (LR to DeepFM to ESSM)

Powered by WooRec Model Serving

This is the core of the engine. To achieve the target Conversion Rate, we iterated through three stages:

  1. The Baseline (Logistic Regression): Initially, we used linear models. While fast, they failed to capture complex feature interactions, such as how material preferences interact with style and price.
  2. The Upgrade (DeepFM): We introduced Deep Factorization Machines to learn high-order feature interactions, significantly improving accuracy on sparse data. This helped in better predicting which items a user would click on (CTR).
  3. The Final State (ESSM - Entire Space Multi-Task Model):
    • Why this model?: To solve the CVR estimation bias and to balance the dual objectives of CTR and CVR, we deployed ESSM. This model shares representations across tasks (CTR and CVR) and corrects for selection bias, leading to more accurate conversion predictions. This was critical for increasing AOV and repurchase rate.

Phase 3: Traffic Control & Business Logic

Powered by WooRec Rule Engine

Raw scores are just probability predictions. To align with business goals, we applied a Traffic Control Layer:

  • Diversity (Scatter/Shuffle): We implemented a sliding window rule—no more than 2 items from the same category in a row—to prevent visual fatigue and encourage exploration of different styles.
  • Business Injection (Hard Insertion): Specific slots (e.g., Position 4 and 10) were reserved for Lumière Jewel’s strategic partners or high-margin house brands.
  • Dynamic Weighting: We boosted items based on Inventory Depth and Profit Margin, ensuring the AI drives not just clicks, but sustainable revenue and avoids stock-outs.

The Seamless Frontend Experience

Here is how these intelligent recommendations appear on the Lumière Jewel storefront:

dashboard *Figure 1: The result of WooRec's engine—hyper-relevant product recommendations displayed to the user, with material properties highlighted.*

The Impact: 77.8% Conversion Growth

The speed of deployment meant faster results. By toggling on these strategies, Lumière Jewel achieved:

> *Interactive Chart: The rapid growth curve following WooRec configuration.*
  • Conversion Rate: Increased by 77.8%.
  • Return Rate: Reduced by 61.1% (from 18% to 7%).
  • Repurchase Rate: Increased by 59.1% (from 22% to 35%).
  • Average Order Value: Increased by 31.8% (from $85 to $112).
  • Customer Satisfaction Score: Increased by 22.2% (from 3.6 to 4.4).

Customer Voice

“Moving from manual rules to ESSM was a turning point. The system now balances user intent with our business inventory logic perfectly. The 77.8% lift in Conversion Rate speaks for itself.” — Sarah Chen, Head of Data Science at Lumière Jewel

Ready to Configure Your Growth?

You don’t need a team of data scientists to build a world-class recommendation engine. With WooRec, it’s just a matter of configuration.

Launch Your Strategy with WooRec