Product Recommendation Engine using DCO
Dynamic Product Recommendation is a smart & powerful technique to show relevant products or services to customers who have shown interest in an advertiser’s website or app through a Dynamic Creative Optimization(DCO) Engine. Showing ads of relevant products to users increases user engagement and the likelihood of purchase.
Dynamic Creative Optimization(DCO) is an ad technology that uses user data to show personalized display ads in real-time. As DCO creatives are more relevant and personalized for each user, they are known to outperform generic creatives.
DCO can be used for both prospecting and retargeting campaigns, one of the most common uses of DCO campaigns is product retargeting based on ‘Last Seen Products’ or ‘Add To Cart’ products.
We implemented a multi-product recommendation engine using our client’s product details data and their purchase data, to recommend multiple products to users who have shown interest in a product on our client's website. The recommended products are selected from the same product category, price range, and highly purchased in the category that the user is interested in. As the creatives will show only the relevant products that the user is looking to purchase, it will create more personalization and promote higher user engagement.
Example of Dynamic Product Ads by Facebook —
Data Sources Used —
- Product Details Data has details of each product — the SKU id, Product Category, Product Description, Price, Quantity, etc.
- Purchase Data has the transaction level product purchase details like — Order Number, Products Purchased, Quantity, Product SKUs, etc.
Model Building Steps —
We used a product similarity-based approach which can recommend multiple products for a last seen product
- The keywords from the product category and description in the product details data were used to generate word embeddings in vector space using word2vec model
- The cosine similarity distance between was computed for each product combination
Cosine Similarity is the cosine of the angle between two vectors and determines whether two vectors are pointing in roughly the same direction
- The price difference score was computed for each product combination which was min(price)/ max(price). The price difference score is close to one when the prices of the products are close to each other and it is close to zero if product prices are far apart
- The frequency of purchase was computed for each secondary product SKU using the capture pixel data for the client
- The three metrics are then ranked and the top three products for a last seen products is selected for a recommendation based on optimized threshold values
The recommended products were appended to the product details data and was used to target users in a dynamic ad creative for the last seen product through our DCO partner.
Recommended Products Example —
We were able to achieve a considerable increase in CTR and CVR by implementing dynamic multi-product recommended creatives through DCO and this method is a highly efficient Machine learning approach to optimize dynamic product targeting through DCO.