How GenAI is transforming e-commerce catalogues | Neha Dixit

Hello everyone. I am Neha and I'm part of Target's data science team and today I'll walk you through how we are transforming item catalog management at scale using the right blend of computer vision, machine learning, and GenAI. What's our goal? Our goal is to have faster, smarter, and consistent item catalog management processes which will help our vendors, guests, as well as internal teams. Uh we refer to as customers as guest in Target. So, you will hear this term, you know, throughout this presentation. So, before jumping onto anything else, let's understand what is item data. In the world of retail or e-commerce, item data is everything. It defines how we describe, categorize, and present our products to our guests both online as well as in stores. So, I'm sure everybody is shopping these days, right? On e-commerce website. So, whenever you land onto a PDP or a public display page, you see this, right? You have title, images, lot of filters, these taxonomies, then uh product specific features. For example, in this case, you have dimensions, fabric names, style, so on so forth. All this is powered by item data. This is item data. Now, let's understand what's the impact of item data or basically why it matters. So, now let uh let's say you have a housewarming party to attend and you want to give rug to your dear friend. You log into a e-commerce website and you type rugs on the search bar. But, instead of rugs, you start to see results from furniture category. Or let's say bedsheet. You'll be like, "Ah, what's going on?" Or let's say you know that friend is your like dear dear friend, and based on the aesthetics, you want to specifically give blue rug. But again, instead of blue, you start to see yellow, red, green rugs. Your experience is ruined, right? Why is it happening? It is happening due to inaccuracies and inconsistencies in the item data. So, we all need to understand that high-quality item data powers search, SEO, guest search, and has a direct impact on our operational accuracies. So now, you all are with me, right? So now you know why I work for item in Target. So, yes. In this GenAI world, we are mindful and cognizant in picking up the right tech to solve all the challenges we see in the item world. >> [clears throat] >> Our decisions are solely based on whether we have enough data for training. What is the cost? What's the time to market? So on and so forth. So, just imagine you have enough data, training data, and you can ship a model in a week or two. What will you prefer, GenAI or deep learning or classical machine learning? You have the data. You can ship the model in week. So still GenAI or classical machine learning or deep learning? The faster one, right? So yes. So if I have data, I will just ship I can ship a model. I can host it on my infrastructure. I will definitely go for classical machine learning or deep learning based solutions rather than GenAI, right? Yes, now let's understand couple of capabilities which we are actively working on. Uh package dimensions are basically they impact our shipping cost. The pallet stacking, trailer space utilization, return rates, etc., etc. Now you can say that oops, Neha, if we have product dimensions, doesn't it translate to package dimensions? The answer is yes and no. So in this particular example, look at this particular mug. If it is made up of steel, the package dimension and product dimensions will be more or less similar. The moment we change the material of that particular mug to ceramic, you have to take care of the padding, right? Because it can break during shipping. And now let's take an example of this particular rug. Before shipping, either you will roll it or fold it based on the material it is made up of. Now if you get this dimension which is 4 ft by 6 ft and 6 in from from vendor, it's not right. This is an anomaly and this is what we are trying to detect. So for this, no GenAI, it's a classical machine learning solution. We have enough historic data. We have spent enough time in feature engineering where yes, with product product dimension, we have identified that yes, the material is equally important. The assembly details in case of furniture will help us decide the package dimensions. And we have trained XGBoost classifier for this and we train it quite frequently. Suggestive variation is another interesting capability which we have where we group item based on themes like colors, flavor counts. I'm sure you have seen, right? Whenever you try to shop for a t-shirt, you see that it comes in different colors, sizes, etc. But, see the complexity. The moment we change the item category from apparel to in this particular case milk, I have to club this milk based on these two things. Right? So, the benefit of this capability is all the guests can see the variations at one single place. So, to solve this, we have trained trained named entity recognition models on all these themes. We identify these themes from these titles. We remove it, and then the remaining the cleaner title are clubbed using normal fuzzy matching. Before this, this was a completely manual process. This solution has brought in efficiencies and reduced the speed to production. Or basically, reduce the item onboarding time. All the images on target.com, they have to adhere to strict guidelines. So, this solution, which is image decisioning, this brings in efficiency and quality. And we call it as image decisioning policy evaluator. So, here are are the glimpse of couple of rules which we have automated. Uh as per style guide guide violation, all the primary images cannot they have to have white background, and they cannot have this border. So, primary images are the first image you see when you land on a product page. So, for this, what are you doing? We have a deep learning model, SAM, segment anything model. We use that to extract all these product images. Sorry, the product from the image. And then computer vision, like averaging, edge detection techniques to catch these violations. Now, let's go to this does not match item description and text on image. For this, we are using Paddle Ocluzia, which will extract extract the text from these images, and we will compare it with what seller has provided us. Now, for example, here, okay, it is not that clear, but it says peanut butter sandwich. So, now, uh from my image, we'll extract this text, and we'll check, "Okay, seller is saying veg sandwich." It's not matching. Something is wrong. So, these violation we catch, and we prevent the images from flowing to our website. An action is generated for business team, and they will communicate to the vendor about this mismatch. Now, let's go to this rule, marketing message not allowed. Marketing messages, they keep evolving over time. They vary from brand to brand, and vendor to vendor. Is it possible, like in this GenAI era, to keep track of how the marketing messages are evolving, and build training data over time, so that we can train a model? No, right? So, that is where for this rule, we are relying on GenAI. So, now you can understand that for this particular solution, is the right solution where it's mix of vision, we have deep learning, as well as GenAI. >> [snorts] >> Um infographics is again a hybrid solution, where we are relying on same computer vision, deep learning, and a pinch of GenAI. When I say pinch of GenAI, all these words you see here, we call it as run-ons, right? So, we are generating these run-ons using GenAI. So, what infographics does is, uh these are, you know, basically three templates, which is given to us by our business teams, our creative teams, rather. And using infographics, we identify unique feature about a product. So, in this particular case, we have these storage boxes. So, how can you define these storage boxes? That they are stackable and they have latch closer, etc. So, we identify these unique feature and we create this. So, one of my colleague, he gave a talk in 2025 fifth elephant edition in Bangalore. And if you're interested in learning more about us, you can go to this link. It's a very interesting talk, by the way. We [clears throat] have built scalable integration between all our data science capabilities and engineering, digital, and item teams. We support both API as well as Kafka integration. >> [clears throat] >> We give flexibility to all our tenants to trigger these uh you know, modeling requests. So, they can train the models as per their uh what do you say? Requirements or SLAs, etc. So, and yes, to ensure 99.9% uh uptime, we have built observability and reliability dashboards, which can generate real-time alerts, too. And here, we monitor during uh training, we monitor our model metrics like accuracy, recall, based on what uh model we are training. And based on that, we decide whether we want to replace the older model or not. All these models are hosted on our Target data centers. And we have a feedback loop, which will ensure or basically track our model performance in production. And it will help us, you know, create golden data sets for validation and training. For all our GenAI capability, we have our in-house again platform, which is called as Think Tank. And we use we work with them uh quite often whenever we generate our basically design a GenAI solution. >> [clears throat] >> Here you can see that we have we cover the all ML stack. We have solution based on classical ML, deep learning, and GenAI. We work primarily on Python, PyTorch to be precise, and yes, our training pipelines are built on Vertex AI. >> [snorts] >> Okay. Now, let's move on to one of our biggest impact area, derived attribution. I'll keep referring it to as DA, acronym. We Target is known for acronyms. Okay. Uh so, I'll walk you through couple of interesting issues which DA was able to identify and rectify. So, you see this image of ottoman, right? How will you describe it? You'll say a square ottoman, sprinted, maybe cream or blue color. But, what if it is tagged to a round value? Now, if you're search if you search on that search bar, I'm looking for square ottoman, will it come? No, because it is tagged to round. Now, if you try to apply shape filter, and specifically you say, "No, no, I want this. It's square ottoman. I want square ottoman." Will it come? No, right? DA was able to identify, and we have rectified it. And this is the live example from our website. Now, let's say you're interested in collecting trading cards, and you are specifically looking for an ultra rare card. But, unfortunately, that attributes So, remember I showed you list of attributes which describes a product? That particular attribute is missing from Target taxonomy. Like, rarity card rarity does not exist. No matter what you search, this will not come up. You cannot even create a filter because we create filters based on our attributes. Because of DA, DA was able to identify this missing attribute, we were able to add this in our target taxonomy for all the trading cards, and now we were able to create a filter on top of it. Derived attribution is a two-step process. In first step, we are trying to enrich attributes and its value. Right? So, how we do it? We learn from our competition. We take competition data. We take target data, whatever we have in our ecosystem, at every category level. Then, we take the search term, whatever guests are searching for on our website, we take we feed into our GenAI model. Our output will be recommended attribute list, which is divided into two buckets, guest-facing attributes and internal attributes. Guest-facing attributes go sit on PDP as part of specification, we create filters on top of it. Uh then, internal attributes are used by our search teams. The second step is Okay. Oh, I'm not sure whether you can see it or not, but it is attribute enrichment or attribute derivation. So, now once we have this attribute list list is frozen, and we have lot of products on target for cards. Now, what we have to do is we will take those active assortment one by one, and we will extract the values of it. How we'll do it? We will take the metadata. When I say metadata, I'm talking about title, images, description, etc., whatever seller is giving us or whatever you see on.com will take that feed into another gen AI model and extract the value. So for example, color is yellow, material is cotton, so on and so forth. Okay, if it is getting con- too confusing, let me simply simplify it for you. You see this card? We discussed about this card in the previous slide. Now, I want to enrich the uh what do you say, attributes for it. What we will do is we will take all the target attributes whatever we have for this particular card. We will learn from our competitors. We will extract the attributes from our competitors. Then we will rely on gen AI retail knowledge as well. And we will take these search terms. We will feed it to a attribute collator. Now, attribute collator is say, "Oops, card rarity is missing in target taxonomy. Would you like to add it?" And we added it. Now, once my attribute list is frozen, we'll go to the next step. We will take all the cards which exist on target and we will enrich the values. So this new attribute is already added now for this particular card. And the other attributes like package type, it says box set, color is golden, so on and so forth. So this we will do it for the all the active assortment we have. With me so far? Perfect. So for this we followed test, launch, and adapt strategy. We tested it with five categories, home categories to start with. But we it took us three months to launch this. Or basically, whatever we enriched, it took three months to go in production. Why? What happened? Because our business teams were manually validating the attribute values. So just imagine there is some human in the loop and they're saying, "Okay, whatever your AI is giving, I want to validate it." Okay, color attribute, accuracy looks good, you can derive it using AI. Material, I'm not sure, you cannot derive it with the AI. So, just imagine for five, it took us 3 months to go in production. That is where we designed an accuracy framework. And this enabled us to go live for 95 item categories and reduced our time to production by 70 to 80%. What do we do? Or basically, what is accuracy framework? For our Just a second. Yes. We identify items within Target's ecosystem which has comp present. So, let's Okay, let's look at this particular dress. What we will do is, if I have this dress, I will ensure that I can get the same information from two or three different sources. So, I'm trying to create a match. So, if I'm going after dresses, I need 100 or 200 dresses which are present with someone else also. Once I have that, I will extract attributes for all those separately. So, from source one, I will extract color, material, etc. From source two, all this I will extract using derived attribution. That's step two. Then, we will create a synthetic ground truth using weighted average or max voting, etc. We will create a synthetic ground truth, which was missing earlier. Now, once I have synthetic ground truth, I will use it to come up with the confidence score for every attribute. Let's dig deeper. I know it's a little confusing. See, I've come prepared. So, what we do is, let's take example of this red dress. We will ensure that we get information from two or more sources. Then we will take the metadata from all these sources. So here like title, long copy, bullets, etc. This dress exists in Target ecosystem. So we have title, long copy, bullet, etc. from Target ecosystem also. Plus seller has also provided the value to me because it already exists, right? I've not I've not started enriching this category with DA yet. I'm still trying to identify the right what do you say? The attributes which DA can derive. So now we will run the GenAI model on these three separately. And I already have seller value. So let's consider the consider case one where AI from on source one is saying the color of this dress is red. Red, red, red. Perfect. The ground truth is red. Now in this particular case out of four sources, three are still saying red. My ground truth is red. Wherever there is a tie we give little weightage to our seller. Because seller for us is kind of ground truth. We assume, although it's not 100% they're not 100% right, but still we give little weightage to them. Hence the ground truth is blue. So we do this we repeat this for all the attributes. So if it is a dress, what is the material, color, collar, so on so forth. And then once I have this synthetic the ground truth for every source, I will come up with an accuracy score or a confidence score. Now if for any confidence score above any threshold, we will say go ahead and derive using GenAI. >> [clears throat] >> So derived attribution help has helped us mitigate the data quality issues. It has helped with enrichment of our attributes. Now once we able to do this, it will have a direct impact on our guest experience center because now if you're looking for rugs, you will find rugs, right? You'll not see furniture, etc. And the same capability can be used during vendor or item setup. Because now I don't have to ask if you want to sell something with us, I'll say I have a capability, you just give me your unstructured data, I will run it and extract the value of all the attributes. If we do that, we will be able to reduce our speed to market by 50%. Right? So, our biggest takeaway from this journey is that we all techie are super smart, you know, we can solve for any challenges we face. But sometimes, apart from your day-to-day deliverables or work, right? Try to unblock whatever roadblocks are there for scale as well as, you know, for any production cases. So, in this particular case, accuracy framework, we all were sitting and we were chat chit-chatting and we're like, how can we enable this? You know, how can we enable our business teams? So, we came up with this solution. Now, the biggest takeaway is choose the right tech for all the challenges you face in your domain. Remember, GenAI is not the hammer for every nail. So, be wise and choose the right tech. And keep learning and keep growing. Now, I'm open for questions. >> [applause] >> Hope it was not boring, right? Good. >> [snorts] >> Hi. I had a question. Um how do you come up with the attributes in the first place? So, I was able to follow your example with the dress. Mhm. There were six uh potential attributes and you identified Which one? Okay, let's go here. I'll go this one. >> in the slide where you showed the dress colors, like you did like the voting kind of thing. Okay. So, in this particular Okay, in this particular case, what we will do is let's say this dress is being sold at competitor one also and we are also selling it. We have title, highlight, description, and images. So, this particular information that the the color of this partic- the dress is red is there in my unstructured data. Now, when we feed this information to our GenAI model and prompt it that give me the right color for this particular dress. So, we have so many attributes, right? So, for every attribute and we have probable list of values also. But, color is an open-ended attribute or something like that that can take n numbers of values, etc. So, you just prompt GenAI and tell what is the right color based on the information I'm passing it to you. >> Yeah, so my question was that you said there are so many different attributes for a particular product and it looked different for every product. How do you figure out what those set of attributes should be? Okay. So, at every item category, so if I'm talking about dress, let's say I can define my dresses with 30 attributes, that's fixed. So, now if it's a yellow dress, pink dress, or a long dress, short dress, attributes are fixed. So, the difference between long, uh mid dress, etc. is length. So, length will be an attribute. So, this attribute list at a category level is always fixed. And that the first step is, okay, let's say I have 30. But, GenAI is saying, "No, no, no, no. You need two more. Add two more." Or you have one redundant, remove it. So, in attribute enrichment, we are trying to expand or shrink those predefined attributes for every category and then we extract. Got it. That makes it clear. Thanks. Um thanks for making it so simple for everyone. Uh since you've already like worked upon a lot of this uh attribute enrichment and uh derived attribute, so >> [snorts] >> if you can shed some light on how the recommendation system will work because uh looks like this will also help uh cross-sell and upsell and uh the recommendations that we see in a typical e-commerce platform and uh possible. I am not an expert of recommendation though, but you need to understand that this is the foundational data. Right? So, if the recommendation system are interested, let's say search for that matter. Forget about recommendation, search. If I'm searching for something, I need the right data tagged with that particular uh item. If anything goes wrong there, if your recommendations systems are recommending, "Okay, Neha prefers black." I need to show her, you know, black color outfits, something like that. But what if my uh apparel data 50% is not correct. So, you will start showing me uh what do you say? Red instead of maybe black. So, think of item data as that you are trying to enrich and rectify the base level data. Search, recommendation, personalization will be built on top of it. If something goes wrong here, it will have a cascading effect on everything. Yeah, makes sense. So, this is definitely like a starting point for This is the starting point. And think about guess that you are interested in black and I'm showing you, "Yeah, this dress is black." What if it's red and I ended up shipping you red? Then you'll be like, "On your What do you say? Website it's red, and why are you giving me black?" Return rates, right? We'll have to bear those return rates, etc. And then guest satisfaction will go down. So, item data, just understand it's the base for everything. No, it's just the What do you say? Base data for recommendation engine. Yes, you're right. We are still not there yet, but we don't know for how long we're not there. For example, this is the interesting talk we two were having sometime back that vision uh an year or maybe 1 and 1/2 year back we weren't confident about vision models, but now just look at all these stable diffusion models, etc. What all magic GenAI can do. And recently I was looking at one article where it has mentioned that for bank analytics reports, now you have GenAI models, right? So, it is Yeah, it is un pre- predictable, but yes, so far I agree with you it's not possible. Thank you, Neha. Thank you.