Lucidworks im Gartner®️ Magic Quadrant™️ 2024 als Leader – WEITERE INFORMATIONEN

Close Message

Video: AI-Powered Recommenders: Self-Driving Mode and Human Intervention

Presented at Activate Product Discovery 2021. Cross-sell, upsell, or straight up just sell. Lucidworks recommender capability is far beyond that of any other product discovery platform or recommender specific vendor. Whether you want to leverage preconfigured models, build your own, or post process with domain specific knowledge, Lucidworks gives you the keys and puts you in the driver seat. We’ll review the different technologies, differentiators, and use cases that can power self-driving recommenders and also highlight the value of balancing AI with your business intel.

Speaker: Garrett Schwegler, Lucidworks Program Manager, Digital Commerce

Transcript:

Garrett Schwegler: Hi, everyone. Welcome to my session on AI-Powered Recommenders. I am Garrett Schwegler, the program manager for Digital Commerce here at Lucidworks. I really wish we could all be doing this in person together, but considering the circumstances ,this will do. There’s really nothing more that I enjoy and love than meeting all of you and having these conversations in person. It’s different times, but like many of you, my wife and I are now full-time work from home. And so over the last year, we haven’t really gone on any trips.

It’s been mostly on lockdown. With that we’ve had some opportunities to do some pretty cool things, one of which is this office that I’m in. We built this, so that was quite a fun project. So this is one that I work out of, And then my wife she’s in one of the rooms in the house,
so we spent some time decorating that.

When I say decorate, what I really mean is we put a dog yoga calendar up in her office.And so with Shortbread and Tinsel, looking at these guys every day, you can only imagine after a few months of sheltering in place, what happened next? So we got a dog and really so much more than a dog. We got Colin the dog. So he’s maybe around five or six years old. He’s an Airedale terrier mix. Yhose dogs tend to be pretty hardy, they’re fun. And he’s been quite an addition to the family, fits right in and really fits in so well, because not only do I have a sensitivity to gluten, he does as well, which is quite bizarre, so we knew it was a match made.

So like any new pet owner, and first dog owner, I went to where I grew up going for goldfish, right? I created an account and just started on the shopping. We had to get food, we had to get the toys, all the good stuff. So we go to the dog food, and luckily there’s a filter that we can filter this by grain-free. I scrolled down and I ended up clicking on this one. After reading through the details and the sides of the package, it looks good. It was gluten-free, a little bit pricey.

This is probably more expensive than some of the food that I eat. And so I wanted to explore a little bit more, what other options there are, and being in commerce, I feel like I know how to get around sites pretty efficiently. And so I head down to the recommendations, which are generally below some of the long description and started to take a look at those to see what my options are. They should be able to help me get around pretty quick.

So I started reviewing these, and I could tell that the first one is totally irrelevant, the melting ice that’s not necessary in the part of the world that I live in. Then there’s a couple options that are all from the same manufacturer that I’m looking at. However, there’s an option there that has whole grain and that is not what I’m looking for, so also totally irrelevant. So two of the five products here are just not giving this company an opportunity to have me convert and add to cart with these. So at the end of the day, I ended up buying the dog food for the PDP that I landed on and all was good.

After I bought it, I go to Facebook, right. I wanna see what my wife and what she’s posted about Colin and I lately. Colin and I are the models, and she’s the photographer, and right away, like after a couple of scrolls, Petco, there they are again. So, great stitching, they know who I am and where I’m going. Also pulling in the Honest Kitchen, knowing what I was shopping for. So that was the brand that I purchased.And this time there are two of the four products listed are whole grain. So, now we’re down to 50% of relevant products in the recommender.

Lastly, I get the email thanking me for joining and at the bottom of this, there’s another series of recommenders of which many of these, or both of these rather, are off base too, right? Colin is not a large dog and he doesn’t have dry skin. So I’m just highlighting here, the opportunity that occurs across so many websites and we have today, the technology, the sophistication, the machine learning and the AI and the human intel to build better experiences quite easily.

So let’s jump in for one more example. Many of you probably recognize this brand just by this statement. And I love these guys, lots of shoes from them, especially through this work from home era, which is why I came back to the site, was to get some fresh slippers. So search for slippers, filter by men’s, and I see here, the Berkeley ones, pretty stylish for me, and click into those to go to the PDP. And as you can see here, when you look at the recommenders at the bottom of this page, all four of these are women’s shoes, right? I explicitly provided the filter of men’s, and I’m on a men’s shoe PDP. Another reminder of the indicators that we should be picking up on, the signals that shoppers are expressing, and how we should be able to leverage those.

So aside from being frustrating, and introducing friction into the experience, what’s the big deal? Why are we here for this? So you don’t wanna raise eyebrows with your recommenders. You wanna raise AOV. This is a study in which they took a look at the engagement of recommenders and the impact that they had with average order value, with a no engagement and the recommenders starting around $45. And then if you go to just one engagement, one click with a recommender, the lift is almost 370%, right? So almost four X, and then it continues to climb until about five clicks for this particular study. The significance here is not the nominal dollar value specifically, going from $40 to $200, right? It’s really the relative increase. And so plug in your AOV numbers there and you’ll see it as well.

So what else? How about conversion? It’s nearly the same story. Conversion can nearly triple once shoppers start engaging with recommenders. So what 208 yeah, so almost three times for conversion. What’s the big deal, right? Well recommenders can not only reduce friction, but they can make you boatloads of money when it’s done well. And so that’s what I wanna focus on today, right? Everybody has recommenders but clearly there’s an opportunity to do this better. Clearly there’s an opportunity to make more money.

So let’s dive in, and nerd out a little bit about these recommenders. Now, when we take a look at this, this is kind of gonna be the cadence that we go through here. There’s not really one model or approach or application that’s gonna be the answer. And so if I look at these and I see that there’s training and post-processing under human and AI synergies that can really also apply to the popular and trending models or the product content based models. And we can merchandise all of these, right. But for the context and for our conversation, what I want to do is just kind of compartmentalize these and we can run through them and then show a couple of use cases as well.

So self-driving mode. I thought this was relatively appropriate for how we’re gonna talk about recommenders, right? This approach is taking into account all kinds of pieces of content and context as well, right? Not only is the algorithm looking forward and down the road, it’s looking in the rear view mirror, it’s looking all around. It understands how fast you’re going. And all of these different components get plugged in to really provide an autonomous experience where it’s learning, and the output of it is gonna be predictable.

All right, so let’s dive into a couple of these self-driving mode approaches. So we have content-based recommenders, which is really a neat way to train a model without signals. We consider this a good cold start option in which you’re essentially feeding the model your catalog, and all of the the attributes about the products. And from that, it’ll determine the relationships amongst the products for a good item to item recommender to start.

And so this is a good way to get up and running pretty quickly, if you don’t have any signals.

The next one here, we can power a trending now, popular items, popular queries with our trending algorithms approach. And these can also be for, and all of these really, can be used for both products and content. And in this context, when I refer to content, what I mean is branded content, and other documents of interests that you might want to surface to the customer and their product discovery experience.

Many of these can really be deployed without any data science background, right? So we’ve done all the legwork to get these into a position where most teams can just jump in and train the model and be up and running pretty quickly. However, if you have a data science team or resources capable of building their own models, we have a data science toolkit and we framed this as bring your own model (BYOM), where teams can build and then train models and integrate them into index and query pipelines with the machine learning from your own team.

If you are saying, well now there’s not really a solution here, or we have resources that we really wanna get creative with, and build models and experiences that are specific for our business.

This is the perfect way, the perfect tool to do that.

So next up we have human and AI synergies. I love this picture for a couple of reasons. One, the girl on the right, that’s probably the face many of us make when we’re scrolling down the page and preparing for what may or may not be a relevant recommender on our own site. Also, these toys, if we think back to playing around with them and how they can walk for a certain distance and then they need to be picked up, repositioned and there’s this human interaction with it. And it’s a pretty good synergy, so let’s dive into this next topic.

So when we think about some of the algorithms that we have listed here, I see right ALS and BPR, those tend to be pretty common. And I think most people are familiar and aware of how those ones work. But the one I wanna focus on briefly is the semantic vector.

And this, if you really wanna geek out, I would go see Eric Redman’s talk, and he will dive into this really deep. So for our purposes here, the way that I would convey this is that graphic above it is a high density of vector space. And that probably makes everybody scratch their heads. But what this model does is, it’s looking at the relationship of products between each other and across the catalog and grouping them. So these concentrations of colors and dots are similar products, and the relationship or distance between the dots indicates how similar to the other ones they are. For a quick analogy, if we think about maybe a grocery store, and if we’re at the fish counter, we may have the fish laid out there, like the salmon filets, then some of the rockfish, some of the shellfish, and each of those groups of seafood are compartmentalized.

Then you might notice that there’s lemons, or ingredients for tartar sauce, or some Old Bay. And so those generally would be further across the grocery store, but are brought closer because there’s a relationship there. So the lemons have relationships to other produce but also in this instance.

In the grand scheme of things, we’re using models like this to understand not only like a semantic similarity between products, right, just really beyond their lexical definition, but really the goal of the shopper between queries and products as well, so that we can make sure we’re generating products that can meet the goal of the shopper that may not be just on a keyword match or a lexical relationship. So it tends to get pretty deep into the recommendations and very valuable. That’s a very sophisticated approach that we’ve packaged up so that it can be picked up, trained, and deployed relatively easily.

So talking about training, what do you need to train these models? A lot of times it just comes down to your product catalog, signals, which you’re likely collecting already. So these are the clicks, add to carts purchases. These can be signals from the store. These can be signals from customer service. Anything that can help train more relevancy, more sophistication in these models can be captured, if they’re not already, and then business intel. What I mean by this is often times because of the challenges that are faced and some of the legacy technologies for product discovery, there may be a lot of human intervention by way of creating relationships between either words, so like synonyms, but also like project oriented things. So if you’re gonna repaint a rusty railing, you may have manually built a project for that and have a relationship of products that you’ve put onto a product detail page. So things like that can be taken and put into models to help train them as well.

And then lastly, post-processing, which is really the bulk of this human synergy with AI where we can block boosts, we can put rules in. If you think back to those examples from the slippers, we can make sure that there’s gender rules that are put in to only recommend men’s products to men or grain-free dog food, if you’re shopping for grain-free dog food. So there’s different ways that we can make these models stronger by post-processing them.

So I wanna show an example of what that can look like. Here we are, this is just a demo site that we’ve assembled, but on this PDP, I can scroll down to find the recommenders and you can see that they’re all relevant, right? These are other jackets that are women’s. So we’ve done a good job of delivering that. And then when we think about post-processing, there are different ways that we can do this, right? So we can boost with signals, we can apply rules, there’s all kinds of different tweaks that you can do, right in Fusion to deliver a more enriched experience. So maybe taking a color boost of the color of the product on the PDP, making all of the products in the recommendation carousel show that color as well. So different tweaks and things that you can do.

So, lastly, I wanna take a look at human intervention. So what are the different tools that we have to really get the human involved into these experiences? And this is really what I love, having done some of the merchandising work in the past on both the business side, I’ve helped out on the technical side here, and merchandisers know best there’s always times where it just makes sense to have the merchandiser put their touch on it. Whether it’s for campaigns or emails, product detail page, or the homepage, there’s always gonna be a need for the flexibility to curate these. And so with our Predictive Merchandiser tool, we can not only block, but also pin. There’s also a breadth of functionality in there that expands beyond recommenders. And I think if you go over to Tom and Katie’s talk, they’ll dive more into that as well. So that’s the human intervention side of it.

Now, when I thought about what use cases I wanted to show here, I don’t want this to be so much of a Recommenders 101 – this is where the recommenders go – I want everybody to come out of this thinking, okay, we can try something new, we’re in these unprecedented times, where the evolution of product discovery is accelerated drastically. And at this point, it seems like this is the natural next step.

So let’s take a look, this virtual shopping experience, I’ve seen these popping up across all kinds of brands. And the value that you can provide is extraordinary by having a human individual participate in the shopping journey, when we can’t do that at the stores anymore.
So what if there was a human on the other end, but the experience was still being driven by Machine Learning and AI, it’s kind of the best of both worlds.

So when you click into this, the first thing you see, product recommendations. So to me, this sounds like a tremendous use case for gifting. If you go a little bit deeper to schedule an appointment, they will ask for a bunch of different information. At this point, they know everything about me. They know what I’ve bought in the past, what size and colors fit me, what my order history is from both in-store and online, and what my interests are for this conversation. So there is a ton of rich data in this that can really drive at the end of the day, a lucrative experience.

Maybe we start out with what I came here for. I got my slippers now, so now I need some joggers. So this model could recommend the joggers, and then how about maybe a propensity type model to say, “Hey, Garrett we know that you’ve bought all these things in the past, and when we look at that against all the other data, we think that you might be a good candidate to start getting into our polo short line of business,” and to make a recommendation there. But there’s really endless opportunities here, and it’s pretty exciting.

Lastly, from use case perspective, the in-store experience, right? So we’re gonna go back into stores, and when we go back into stores, it’s probably gonna feel a little bit different, and I’m hoping what’s been happening over this last year is we’re figuring out creative and innovative ways to enhance that experience to get people to come back. When you look at the situation here with the tablet and recommendations on it, whether it’s products that are in stock or out of stock, or knowing who that shopper is and being able to recommend products based on their past purchases or their painting, what are other products that should be recommended for that in the event that associate doesn’t have that knowledge. So again, another huge opportunity here.

As far as our recommenders go, I think I showed a fair amount of apparel, but between the dog food and the paint, I just wanna make it clear that these go across B2C across B2B, D2C, life sciences, apparel, grocery, these models are ready to roll and we can cover the spectrum.

So before I wrap up, I just wanna let you all know that this is Lucidworks and we’re not just a search company. I hope that this was a compelling and convincing demonstration of where we’re at with recommenders and that we can compete in this space. You don’t need an army of data scientists to be successful with these tools. They’re ready to roll. So thank you, appreciate your time and your attention. Please join me, a couple of my team members as well for a little bit of Q+A right after this. Thank you.

Play Video