Vectorize data

Now that we've examined the data, let's proceed to the next step: vectorizing the product data.

All images/dataset used throughout this guide are from: Aggarwal, P. (2022). Fashion Product Images (Small). Available online: https://www.kaggle.com/datasets/paramaggarwal/fashion-product-images-dataset

Product columns

In our dataset, we have 4 columns:

Product data columns
Color
Season
Year
Description

Product images

In addition to these columns, we also have product images.

Here's a subset of our jeans collection:

Select the number of images to display

Show 5 jeansShow 10 jeansShow 15 jeans

• 
  Peter England Men Party Blue Jeans
• 
  Jealous 21 Women Black Jeans
• 
  Jealous 21 Women Black Jegging
• 
  Tokyo Talkies Women Navy Slim Fit Jeans
• 
  Locomotive Men Washed Blue Jeans

Vectorizing data

Different strategies

When vectorizing our product data, you can use three different strategies:

1. Vectorize product texts
2. Vectorize product images
3. Vectorize both product texts and product images

Let's examine how the AI recommendations vary with different vectorization strategies.

1. Vectorizing only product texts

We start with vectorizing only the product texts. This approach can be useful if you either don't have images or if the product images don't contain visual features relevant to customer inquiries.

For instance, here's how the product text vector looks for one pair of jeans:

Product text to vectorize	Product text vector
Jealous 21 Women Black Jeans	[0.021333912387490273, -0.01840313896536827, ....]

After vectorizing all the jeans descriptions, we can run a customer inquiry like "I'm looking for light blue women's jeans". This straightforward inquiry mentions gender and color. Let's see the AI model's performance with just the text vectors.

Product recommendations

Based on this customer inquiry, the AI model provides these jeans recommendations in order:

AI product recommendations for customer inquiry

Show top 5Show top 10Show top 15

• 
  Elle Women Light Blue Jeans

  Rank: 1
• 
  ONLY Women Blue Jeans

  Rank: 2
• 
  ONLY Women Blue Jeans

  Rank: 3
• 
  Lee Womens Blue Jeans

  Rank: 4
• 
  Lee Womens Blue Jeans

  Rank: 5

The top 5 recommendations are definitely light blue, but the top 10 and top 15 include some darker blue.

Let's see how the AI model performs using only the images.

2. Vectorizing only product images

Next, we vectorize only the product images.
Here's how the product image vector looks for one pair of jeans:

Product image to vectorize	Product image vector
	[0.0037902002, 0.018807068, -0.0027826785,....]

Product recommendations

Based on the customer inquiry, the AI model now provides these jeans recommendations:

AI product recommendations for customer inquiry

Show top 5Show top 10Show top 15

• 
  Lee Women SS Blue Jeans

  Rank: 1
• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 2
• 
  Jealous 21 Women Washed Blue Jeans

  Rank: 3
• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 4
• 
  Puma Women Blue Jeans

  Rank: 5

Compared to the text vector recommendations, all the jeans recommended by the AI model are in light blue colors.

3. Vectorize both text and images

In this experiment, we'll vectorize both the product text and product images.
Here's how the combined vector looks for one pair of jeans:

Product text to vectorize	Product image to vectorize	Product text-image vector
Jealous 21 Women Black Jeans		[0.0037902002, 0.018807068, -0.0027826785,....]

Product recommendations

Based on the customer inquiry, the AI model now provides these jeans recommendations:

AI product recommendations for customer inquiry

Show top 5Show top 10Show top 15

• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 1
• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 2
• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 3
• 
  Jealous 21 Women Washed Light Blue Jeans

  Rank: 4
• 
  Lee Womens Blue Jeans

  Rank: 5

Observations

The top 5 recommendations are light blue jeans, followed by a range of darker jeans. Compared to text-only vector recommendations, these recommendations include more light-colored jeans but fewer than the image-only vector recommendations.

Comparisons side by side

Here's a table showing the top 6 product recommendations for the customer inquiry "I'm looking for light blue women's jeans" for comparison:

Vectorization	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5
Text vectors
Image vectors
Text+image vectors

While this approach to vectorizing product data is quite basic, optimizing the product descriptions can further enhance the AI model's product recommendations.

Let's discuss how we can enrich our product descriptions in the next section.