Optimize data

In the previous section, we used AI to recommend products based on the straightforward customer inquiry, "I'm looking for light blue women's jeans".

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

The results were promising:

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

But what happens when we try a more complex inquiry like "I'm looking for women's jeans for a summer party"?

Let's first think about what kind of jeans we, as humans, would recommend knowing only this. Perhaps, light colors, lightweight fabrics, and perhaps wide-leg jeans for a flowy feel might come to mind.

With this in mind, let's look at our data again. Here's a subset:

Product Name	Gender	Year	Season
Peter England Men Party Blue Jeans	Men	2012	Summer
Jealous 21 Women Black Jeans	Women	2012	Summer
Jealous 21 Women Black Jegging	Women	2011	Fall
Tokyo Talkies Women Navy Slim Fit Jeans	Women	2012	Summer
Locomotive Men Washed Blue Jeans	Men	2011	Fall

Enriching data

In our previous example, we only vectorized the product description, but that leaves out both the season and year.

If you had to recommend jeans for a summer party, knowing only the color isn't enough. By enriching the product descriptions with additional data, we can provide more context for the AI model.

Here's how we could enrich our product descriptions by adding data from the other columns:

Column	How to enrich vector
Color	Visual matching and a common search attribute
Season and Year	Provide temporal context, hinting at the style period

Enrichment example

Here's an example of how we could enrich our original product descriptions with season and year:

Before	After
Jealous 21 Women Black Jeans	Fall 2011 casual black jeans for women by Jealous 21, ideal for everyday wear

Reasoning behind adding seasons

Consider a customer inquiry like:

"I'm looking for women's jeans for a summer party"

Suppose the season is not mentioned in the product description that we've vectorized. In that case, you're solely relying on AI to draw its own conclusions about what is considered suitable for a summer party.

Perhaps, the AI model might focus on colors, and recommend jeans in light colors, for instance.

By adding seasons to the product description, we provide the AI with additional context, reducing the reliance on AI inference alone.

AI product recommendations

Let's test our new customer inquiry on our original vectorized data to get an idea of how the AI model reasoned:

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

This one is tricker. Let's also look at the season for each recommended pair of jeans:

Comparing seasons of different vectors

Choose between different vectorization strategies

Text vectorsImage vectorsText + image vectors

This comparison shows that while the text vector recommendations include darker jeans, most of the recommendations are labeled as summer jeans.

Let's look at how we could optimize the product description by enriching it with season and year.

Enrich product descriptions

Let's enrich our product descriptions by adding season and year to each product description, here's a subset of how our new jeans descriptions will look like:

Before	After
Peter England Men Party Blue Jeans	Summer 2012 Peter England Men Party Blue Jeans
Jealous 21 Women Black Jeans	Summer 2012 Jealous 21 Women Black Jeans
Jealous 21 Women Black Jegging	Fall 2011 Jealous 21 Women Black Jegging
Tokyo Talkies Women Navy Slim Fit Jeans	Summer 2012 Tokyo Talkies Women Navy Slim Fit Jeans
Locomotive Men Washed Blue Jeans	Fall 2011 Locomotive Men Washed Blue Jeans

AI recommendations optimized data

Let's look at the AI product recommendations on the optimized data, starting with the text vectors:

Comparing seasons of different vectors

Choose between different vectorization strategies

Text vectors originalText vectors optimizedImage vectors originalText + image vec. originalText + image vec. optimized

The optimized product data shows that the new product recommendations are all labeled as summer for the text vectors and all but one for the combined text and image vectors.

Pure image vector recommendations stay the same

The pure image vectors remain unchanged since the optimization only affects the product text.

Product recommendations

Let's also look at the actual jeans recommendations, here are the product recommendations visually:

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

Comparison: original vs. optimized Data

Here's an overview of the original product recommendations compared side-by-side with the recommendations based on the optimized data:

Text vec. original		Text vec. optimized		Image vectors		Text+image vec. original		Text+image vec. ooptimized
Rank: 1		Rank: 1		Rank: 1		Rank: 1		Rank: 1
Rank: 2		Rank: 2		Rank: 2		Rank: 2		Rank: 2
Rank: 3		Rank: 3		Rank: 3		Rank: 3		Rank: 3
Rank: 4		Rank: 4		Rank: 4		Rank: 4		Rank: 4
Rank: 5		Rank: 5		Rank: 5		Rank: 5		Rank: 5

We've seen how naive vectorization of original product descriptions can be improved by enriching descriptions with additional data.

Next, we'll explore another optimization technique: hybrid search.