How AI Works in Cluster Finding

When you select parameters for your clusterization in the Cluster Designer and set the necessary limits, Karsa Labelizer's artificial intelligence (AI) takes the initiative. Unlike simple rule-based systems that segment products based on fixed threshold values, Karsa Labelizer uses advanced machine learning (ML) algorithms to identify optimal and natural groupings (clusters) in your product data.

Basic Principles of AI in Karsa Labelizer

1. Multi-dimensional analysis:

   • AI can analyze products based on multiple parameters (dimensions) of your choice simultaneously. This means it doesn't evaluate products in isolation according to one criterion (e.g., just ROAS), but looks for complex similarities across the entire spectrum of selected metrics (e.g., ROAS, Conversion Value, CTR, Product Price, etc.).

2. Goal: Consistency and Predictability:

   • The main goal of AI is not just to divide products in any way, but to create highly consistent clusters. A consistent cluster contains products that behave as similarly as possible in terms of the chosen parameters.

   • This consistency leads to higher predictability of future campaign performance, which is key for effective optimization by Google Ads algorithms.

3. Hybrid algorithm:

   • Karsa Labelizer uses a sophisticated hybrid machine learning algorithm. This approach combines the advantages of various models and is designed to best suit the specifics of product data in e-commerce.

   • Key features of our algorithm:

     • Working with uncertainty in data: Provides probabilistic assignment of products to clusters, which better reflects the natural variability and uncertainty in performance data.

     • Automatic determination of the appropriate number of clusters: Within the limits you set for the minimum and maximum number of clusters, AI identifies the number of clusters that best corresponds to the natural structure and distribution of your specific product data. It is therefore not necessary to "guess" the optimal number.

     • Robustness against noise and outliers: The algorithm is designed to better handle outliers (extremely high or low performing products) and noise in the data without unduly distorting the resulting cluster structure.

     • Interpretability of results: Despite being an advanced ML model, we strive to make the results as interpretable as possible. Probabilistic characteristics can help better understand the strength of a product's assignment to a given cluster.

     • Natural processing of missing values: The algorithm can effectively work even with products that may not have a complete history for all chosen parameters (e.g., new products, if a strategy for their separate processing is not chosen).

What Happens "Under the Hood"? (Simplified)

1. Data preparation: Your chosen parameters and product data are normalized and prepared for analysis.

2. Structure search: The AI algorithm iteratively searches through the data and looks for "clusters" of products that are similar to each other in the chosen dimensions (parameters).

3. Optimization: The process is optimized to:

   • Make products within one cluster as similar as possible (maximizing intra-cluster homogeneity / consistency).

   • Make individual clusters as different from each other as possible (minimizing inter-cluster similarity).

   • Maintain the minimum number of conversions per cluster.

   • Find the optimal number of clusters within your limits.

4. Product assignment: Each product is (with a certain probability) assigned to the most suitable cluster.

The result is a data-based segmentation proposal that is ready for your evaluation and subsequent deployment to Google Ads through Custom Labels.

---

Understanding the basic principles of how AI works will help you better interpret the results of clusterization and more effectively use Karsa Labelizer to optimize your campaigns. In the next section, we will look at practical [Custom Label Setup](../custom-label-setup.md).