VisualML

VisualML is an Experience OS extension, and part of the AdaptML^® self-training, deep-learning AI system. Contact your Customer Success Manager to learn more. 

VisualML is a deep learning recommendation algorithm that identifies and suggests items that are visually similar to the currently viewed item. 

Using an advanced visual transformer model, VisualML finds and recommends items that match visual attributes not easily described or tagged in metadata. It's available for both product page recommendations and email campaigns. 

The model's performance is influenced by the type of images it's trained on:  

• Wild Images: Product images featuring models in atmospheric or inspirational backgrounds, often including additional products. For example, a picture of jeans that also clearly shows sunglasses, a shirt, and shoes. In this case, the model focuses on recommending items that match the overall style and look represented in the image ("You might also like").
  If you use wild images, we suggest applying recommendation filters to align with the specific category of the item in context. 
• Studio Images: Clear images of the product against a clean background. If a human is presenting the product, the product in question is the only fully visible item. The model prioritizes recommending similar products based on attributes like color, texture, material, and style. Studio images typically yield better results in a "Similar Items" recommendation widget. 

Limitations

• The model requires access to your image CDN to download and process your image. Have your technical team add the following IP addresses to your allowed list:
  • • 8.197.234.165
    • 3.127.189.219
    • 18.158.108.19

  • • 18.192.189.87
    • 18.195.163.252
    • 3.92.38.28

  • • 3.227.221.249
    • 54.145.37.112
    • 75.101.154.167
• If the CDN does not respond within 5 seconds for a specific image, that image is skipped. 
• Maximum image size is 10MB
• The image path must be explicit.
  For example:
  Correct: https://mydomain.com/myimage.jpg 
  Incorrect: //mydomain.com.myimage.jpg or //myimage.jpg 

FAQ

How long is the training before I can use the model? 

The initial training can take anywhere from several hours to a couple of days, depending on the size of your catalog. Then the model is ready for use. After the initial training, the model processes only new SKUs. This typically takes seconds and up to a few minutes. 

Which industries can benefit the most from using VisualML?

The algorithm is versatile and capable of functioning in any sector where images contain pertinent visual data for comparing different products. It is particularly well-suited for industries such as fashion, furniture & home decor, art & decor, jewelry, and related subsets.