A Powerful Team: PIM and AI
AI-based classification of products in e-commerce
Artificial Intelligence (AI) in e-commerce - this is no longer pie in the sky. In interaction with product information management systems (PIM), AI shows its strength especially in the classification of various articles. The six tips from Adrian Gasch, Manager PIM Technical Solutions at Arvato Systems, show what is important here.
Tip 1: Trust the Power of Images
A picture is worth a thousand words. Methods of AI-based and thus automated image recognition have great potential in e-commerce. With a well-photographed image, it is not only immediately apparent whether the item is a fashion item such as shoes, a consumer good such as a coffee machine, or an industrial product such as a machine from a series. It is also possible to recognize the color of the article. Even beyond such use cases, AI-based neural networks are able to distinguish between articles in the same product group.
Tip 2: Train the AI
With the right training, a neural network delivers accurate results (deep learning). To do this, you first need to teach it to classify items independently. A large number of product images stored in your PIM system and provided with meaningful metadata serve as training material (Big Data). Properly trained, the AI-based system recognizes an item of clothing or a machine, for example, even if it has been twisted, partially obscured or taken in unfavorable lighting conditions. It thus relieves you of time-consuming repetitive tasks.
Tip 3: Set Different Levels of Complexity
You should set different levels of complexity for training the AI:
- Type of the article
- Color of the article
- Brand of the article
- Size of the article
- Distinctive features (patterns and such)
- Material of the article
- ... and so on.
You should keep in mind: At the first two levels, an AI delivers very convincing results. From level three, at the latest four, a neural network currently still reaches its limits. As long as the design is clear or a logo is recognizable, it can distinguish between brands and detect counterfeits. In this way, you can effectively prevent trademark infringements. To classify an item even further, the image quality is often insufficient. Also, an AI needs a lot of training for such complex tasks.
Tip 4: Define Thresholds for Image Recognition
Thanks to their outstanding abstraction and classification capabilities, artificial neural networks can evaluate images in milliseconds and categorize products in real time - regardless of lighting conditions, viewing angle and background. They don't even have to match the reference object one hundred percent. A sufficiently high probability is sufficient. That's why you should set the thresholds wisely: The AI recognizes an item as just that when a predefined threshold is reached, for example, "90 percent match with the reference object." If a clear classification is not possible, you can manually select the matching product from automatically generated suggestions.
Tip 5: Check the Results
Especially at the beginning of the training phase, you should continuously check whether the AI delivers the desired results. If this is the case, you can start working with the AI-based system. An image flows into the PIM system, where the AI automatically classifies the pictured item and assigns probabilities to it: For example, 95 percent of the photo shows a shoe or a household appliance, and 100 percent shows it sideways. In such unambiguous cases, you can accept the suggested classification along with matching keywords (tags) without hesitation and store the product image in the PIM system. Likewise, you can specify that the photo showing the item from the side automatically appears as the first image in the online store. Without the manual selection, the system may select a less descriptive image, such as of the sole of the shoe or the back of the device. To avoid this, you should specify the order in which the product images are to be played out.
If the system does not detect a sufficiently large match - for example, if the probability is below a previously set threshold of 80 percent match with the reference object - you are well advised to check the pre-classification and adjust the proposed tags in case of doubt. You should also determine the order of the images.
Tip 6: Choose the Right Service Provider
The technology has the potential to automate manual processes in product data maintenance to a high degree - provided the metadata is maintained in the PIM system and thus forms an optimal basis for meaningful tags, which in turn are indispensable for training the AI. In order for an AI to do its job, it has proven useful to bring a competent service provider on board who checks use cases for their feasibility and points out technological limits. He also analyzes whether the required metadata is available in the PIM system and how a specific use case should be designed. Last but not least, he trains the neural network. Due to his great understanding of data and interfaces, such a service provider is able to professionally accompany projects in the field of AI-supported image recognition and lead them to success.