Fashion Similarity for planning and analysis

Blog post
Smart Data Services
Ian Shulman
06
.
02
.
2024
a hand holding a shirt on a hanger

Recognizing similarities with neural networks

In the fashion industry, it often happens that two items have completely different names, descriptions and characteristics and therefore appear to have nothing in common. However, when it comes to sales patterns and turnover figures, they end up being practically identical. This is often because, outside of the classic database entries, there is a fundamental similarity between these two products that only the customer can recognize. Or a neural network. In our projects, we therefore use neural networks to plan new articles in the fashion industry, for example, in order to recognize visual similarities between article images.

But why is such an analysis important? We can think of at least two areas of application: Sales planning and recommendation systems.

Use in sales planning

Applied to sales planning, a company can predict the sales of a new item based on the sales of an existing, similar-looking item. For example, perhaps striped blue men's shirts are particularly trendy this season. However, a company launching a new shirt may not be aware of this particular trend, may not consider the stripe pattern to be an important feature, and may not have listed the stripe pattern feature in its database. Consequently, it is also difficult to make a suitable sales forecast.

However, using a neural network that can recognize the pattern, we can identify similar-looking shirts and predict sales of the new shirt based on historical sales of visually similar items. The prediction becomes even more accurate when we combine these findings with existing master data such as price or available sizes. The price alone would not be enough to predict the sales of the shirt, but in combination with the pattern information it can give us a good idea of how shirts of the particular design and price category tend to sell.

The addition of visual insights to the database can have many other useful effects - for example, the tracking of predecessor/successor articles. Assuming that a successor item often looks very similar to its predecessor, the neural network can make suggestions about item relationships. Storing this information can prove to be very helpful for sales analysis and forecasting.

For incorrectly labeled items, the system can indicate a possible mislabeling if an item with high similarity belongs to a different product category. For example, if a shirt is incorrectly labeled as pants, the system easily recognizes that most similar looking items are classified as shirts and most pants look different than the specified item. In the case of incomplete master data, the system can also provide labeling suggestions based on item similarities.

Finally, if the company wants to categorize items into groups, for example for sales or marketing purposes, visual similarity can be a very important grouping factor.

Use for recommendation systems

Recommendation systems are another implementation. The more detailed the master data, the better the item recommendations and the higher the purchase probability. By suggesting items based on visual similarity, customers receive more precise suggestions that go beyond the content of the master data. In addition, sometimes there is no way to ensure that the master data actually contains the particular feature that makes two items similar. For the customer, a particular item may be almost perfect, but there is still something missing that is difficult to describe. A very similar item may then fulfill all the criteria.

‍Whyneural networks?

Using neural networks to solve this task has many advantages. Human processing takes a lot of time and even the human eye can be biased and disregard some features that are actually relevant for sales prediction. Pixel-by-pixel comparison is too simple for this task - even the same object photographed from a different angle differs at the pixel level, but is identified as the same by a neural network. This is because neural networks are able to recognize patterns regardless of colour scheme, angle or other factors. Such a solution can be integrated into a planning platform and thus utilize the power of image comparison for planning and forecasting tasks.

The way it works can be explained as follows: an image of the respective product is matched in pairs with each image from the customer's image database. A neural network performs a similarity analysis between each pair of images, extracting different image properties in the background that create similarities or differences between the images - such as color, pattern or clothing type when it comes to clothing images. Based on these properties, an overall similarity is calculated for each image pair in the form of a "distance" to the original image. The following graphic illustrates this approach:

Fashion images as a starting point

Our current implementation is based on fashion images. This is because fashion items typically have a wide variety of color options, patterns and other features that are relevant to the customer's choice, but are not necessarily fully reflected in the product description. However, this approach can also be used in practically any other industry where visually striking consumer goods are involved, as long as high quality images are provided and it can be assumed that the appearance of the product has some influence on the purchase decision.

Discover the real similarities of your products

Would you like to understand the similarities between your products at a deeper level and use the power of neural networks to achieve better results in sales planning and analysis? Feel free to contact us and do the practical test with your own images.

,

Blog post author

Ian Shulman
Ian Shulman
Senior Data Scientist
celver AG

Ian Shulman is a Data Scientist at celver with a particular focus on text generation using Large Language Models. He focuses on simplifying complex and cumbersome processes through machine learning approaches and using data science methods to extract new insights from the data.

Case Study on the topic

Our news provides you with the latest insights into smart planning, smart analytics, smart data and smart cloud.

Register now