Composable eCommerce Platform | RETISIO

Webinar Transcript – AI Optimized for eCommerce

Learn with RETISIO

Webinar transcripts from the experts in retail technology

Webinar Transcript

AI Optimized for eCommerce: Why it Matters to Your Business

RETISIO Host: [00:00:00] Well, good morning, good afternoon, or good evening, depending where in the world you are today. Welcome to our webinar, AI optimized for e-commerce, what it means to your business. We’re really excited to have you here today. Today, you will learn more about artificial intelligence machine learning and how you can leverage them in your own online business.

We’re going to get started in just a couple minutes. First. Let’s cover a few housekeeping items. If you’re having any problems with your sound or video, use the chat box on your go to webinar console, and we’ll attempt to resolve your issue. If you have any questions that occur to you during the webinar you can type you in the questions box on your console.

We will be having a brief Q and session following the webinar today. Okay, now that that’s out of the way I’ll introduce our presenters to you. Up first is Raj Yarlagadda. Say hi Raj.

Raj Yarlagadda: Hey guys.

Raj has more than 25 years of experience in building leading retail organizations, including serving as VP of eCommerce technology for Kohls for more than a decade.

Raj is currently a co-founder [00:01:00] and chief digital officer for RETISIO.

Next up is Tony Moores. Hi, Tony.

Tony is a veteran veteran digital commerce practitioner who’s been developing software teaching and consulting for more than 20 years. Tony has deep industry knowledge acquired by working closely with many brands across many verticals.

Tony is currently the chief technical officer at RETISIO.

Now I’ll go ahead and turn it over to Raj Yarlagadda:
Raj: Brett, thank you for the introduction. Welcome everyone to the webinar, artificial intelligence and machine learning as health retailers deliver engaging customer experience provide real time personalization and improve operations across the board. And ML have health companies like Amazon significantly and become an industry.

Our goal is to help use the power of ML across your organization to improve your top line sales gain market share and reduce cost.

In, in today’s session, we’ll provide an overview of the value that [00:03:00] AI ML can bring to your online business. Discuss about some AI ML use cases in marketing, product discovery, sales, and operations.

Show you some examples of AI, well driven product discovery and KPIs they achieve.

Let’s start with marketing retailers, spend a lot of money on marketing. If the campaigns and segmentations are not optimized, you might get low converting traffic, clustering models and propensity models will optimize your segments and campaigns. One example is behavior cluster. Behavior clustering informs how customers behave will buy.

Do they use discounts? How frequently do they buy? How much do they spend this algorithm help you select the right segmentation and campaign for your customers?

Propensity models and clustering models will drive high quality traffic, which will [00:04:00] convert better on your digital channel. Once a customers come to your digital platform, you need an engaging, personalized product discovery experience, an ML based search engine, product recommendations, visual search and NLP classifications are core components in creating a faster, relevant, engaging product discovery experience.

Some of the ML models used in search are recurrent neural networks, logistic regression, random forests, and learn to. Some of the models using product recommendations are collaborative filtering and XG boost regression. Since it’s a short session we don’t have time to discuss ML models and sales and operations.

Let’s take a deep dive into ML driven product discovery roadmap.

Here’s a snapshot of a good roadmap of product discovery. Using ML ML models should be trained to understand long and complex [00:05:00] search querie. The model should be trained to understand context, be backed by deep neural networks for price categories, brand affinity, and other key dimensions. The results should be personalized using effective relevance, ranking models.

Some of the features that should be built for smart product findability are search query prediction, understanding and predicting user intent, self learning, and personalized search results. Smart recommendations, help improve customer engagement in each stage of product discovery, the model should dynamically recommend products based upon their customer journey.

It could be trending products frequently bought together smart product bonding or best sellers. The key is to personalize their product discovery process by understanding the current context, current customer journey and historical context.

Let’s take a look at a [00:06:00] typical customer journey, the pain points they face and how AI ML can create an amazing product discovery process. Let’s take an example. Washer under thousand looks very simple, but here, when a customer’s looking for a washer under thousand, they might misspell something or not provide space between words because they’re using a mobile device or typed in.

It will be a bad experience. If the customer gets no results, the model should be smart enough to understand the query intent and produce relevant results. Let’s take a look at another example, Sony camera with SD card here, the customer is very specific about the brand product and features. The model should understand the search query intent and produce relevant results.

Typical search engines might have difficulty in understanding long keywords and complex search queries. In few minutes, I’ll show you video of a bad and a good customer journey. [00:07:00] Using these examples in two different websites, one is powered by our company. Another one is a top leading retailer. The key capabilities are smart.

Product findability are such query production, understanding, query intent, learning self self learning models and personalizing the customer journey based on real time interactions and historical data.

The capabilities, the key capabilities of smart recommendations are collaborative filtering, smart product bonding, optimize, cross sell, and upsell product. The goal here is to personalize the product discovery process by understanding the current context, the customer journey and their historical interactions.

So the key is real time interactions and also the historical interactions, smart search and smart recommendations will [00:08:00] help customers find relevant products, faster, improve customer engagement, ready, ease, balance rate, and improve conversions, AOI and revenue.

Raj Yarlagadda: Why don’t, [00:10:00] why don’t I go through the KPIs and then hand it over to Tony who can walk you through the technical aspects of how these algorithms work and Brett, maybe you can play the video at the end if that’s okay.

Brett: Sure. That’s fine. Yep. Yeah. That’s. Sure let’s go to the next slide.

Raj: So based upon what we have done for our customer, right we were able to increase the KPIs. For them. And here are the certain KPIs where we achieved for one of our clients, which is brand smart. They were able to achieve five X uptick in their conversions.

The search revenue jumped from like $7.5 to $29 person. And this is just the starting as algorithms learn, they get better and better. And their search re lead revenue will increase up to 50%. Right. And this is a typical journey that you can get using a well. A great ML based search engine. So with that, I’ll pass it on to Tony who can walk you through the [00:11:00] technical aspects of how the algorithm works and then Brett can come back and then play the video.

Tony Moores: Thanks Raj. I appreciate that. So we’re gonna start off by taking a look at different types of data that you’ll have in your e-commerce site and some different types of learning that You’ll hear, you’ll hear folks talk about. So first off then we’ll talk about unsupervised learning, which is essentially about uncovering artifacts in your data, whether it be categories or subtypes or, or trends that are apparent, just.

By just through the structure of the data itself. And in contrast to something like supervised learning. Where you have the ability to provide either an example of something in a category or not in a category, or have some [00:12:00] sort, sort of Oracle that can tell you true or false this thing qualifies or doesn’t qualify for whatever, for whatever test you are.

You’re trying to perform. And then a third type of learning. Which is called reinforcement learning is, is a bit more like a trial and error mechanism where you see if something works, you take a guess, you get some sort of feedback. Either positive or negative, which will, which will you’ll use next time when you make the next guess.

The idea being reinforcement learning should refine over time. And when it’s used in conjunction with either supervised or unsupervised learning you can essentially. Bootstrap your whatever it is you’re trying to, to figure out and then get better over time. Typically you’ll [00:13:00] have the most, and, and most structured data around your catalog, which is, which is where folks will have their, their biggest opportunity to leverage AI or ML.

And then of course, You would, you would like to do better over time dealing with outcomes and, and in that sense, you would, you would hope to be able to increase the O the overall amount of data and structure in these other areas as. Now, when you talk about the types of things that you need in order to make AI and ML work well for you usually start with data and analytics, as we said before, you tend to have most the greatest amount upfront with your catalog.

But over time, you’d like to make these things grow across your user [00:14:00] content and outcomes. In fact, it’s the outcomes where folks start making that transition from data and analytics into business insights. These are going to, these are the analytics that allow you to actually operate your business better or to make decisions that really, that really count.

Then, of course those being the things you’re ultimately trying to optimize, these are the things you want to apply AI and ML to in order to, in order to increase these things over time, you want to be able to do that with to, to, to lower the overall time and effort. It takes you. To, to optimize these outcomes by applying some machine learning.

So in order to do that, [00:15:00] one of the things you’ll have to you’ll have to sort out is this notion of bias and infinity, both between your users and the catalog and your users in the content. What we’re talking about here is. Do certain types of content appeal more to certain types of users, do certain types of product tend to be purchased more or, or more favored by certain types of users.

Right? So this notion of bias and, and affinity goes a long way into learning how we can present the right things to the right folks. To go one step further. We want to tie this back with outcomes, right? So when we start talking about cohort analysis or behavioral models that really tie back to which users are likely.[00:16:00]

To experience which outcomes as a function of the content or the catalog items they’re experiencing. Right. So now what we’re talking about really is measuring things like how often people click, whether or not they go deeper down a. Whether or not they stay on a page or a screen whether or not they convert, for example, a search or, or a product purchase.

But essentially tying back this cohort analysis or this behavior. Back to these behavioral trends back to the content and catalog available to your users during their experience. Now this is all very academic. So I wanna pick up and use the same example that that Raj was talking about in the search experience to show you how some of these [00:17:00] things can actually be used together.

We’ll work out the mechanics of these various types of learning and approaches in that intelligence search example that we, that we talked about earlier in the presentation. So please note very typically. When folks do a search on a, on a website or any other digital channel, obviously it starts with the user.

They enter some sort of search term, either spoken or typed the search engine executes that search. Against some sort of index that was taken beforehand of the catalog or content that, that index and, and the relationship between the search term and the index will come up with some sort of relevance model for the results that come [00:18:00] back.

We typically display those results. And then wait for some call to action. This is typically where you’ll see somebody else click deeper or, or maybe research because they didn’t find what they wanted. So to make that search more intelligent, we’re gonna start by adding some natural language processing.

So same idea in that you’re going to index these things ahead of time. But what NLP is going to do is it’s gonna bring structure. It’s going to being structured in the terms of parts of speech stemming it’ll take care of, you know, to pants. Does pants mean trousers or trousers mean pants or, you know, what, when does a boot mean something you wear versus some part of your car?

But essentially it adds more data for us to, for our search engine. To to [00:19:00] operate with that, of course there’ll be a, a new relevance model. We might find the same set of results, but some some of those results might be ordered differently or more relevant at all. The, we can also use NLP up front.

Right? So what we were just talking about was how do we use in this case, NLP to make our search engine better able to find the search term that we’re looking for. Now, we’re talking about adding NLP to make the question. Better the idea being, can I get better search results? If I ask the question a better way, we believe you can.

So we, we also use NLP to try to discover the intent of the, of the user by, by doing some analysis on [00:20:00] that search. and then applying what we call facet affinity. So facet affinity is essentially I’ll think of it this way. If I can, if I can break down that question into adjectives and, and a verbs and say, oh, under a hundred, is.

Probably a price or, you know, you know, is 10 80 a price or is 10 80 a resolution. What’s the context. And, and, and if I could. If I could structure the query before I gave it to the search engine to make use of the facets available, could I get a better, could I get a better search result? So I do those, I do those things together.

And then of course we’re using reinforcement learning in NLP all the [00:21:00] time. so that over time, as, as we experience more questions and we experience more catalog items, we can, we can refine our, our ability or get smarter about what things mean. Now we’ve introduced in NLP, a new, a new relevance model.

But again, relevance might not just ha just be a function of the search term. You might have. You might decide that some things are more relevant based on their popularity. It might be based on their inventory. It might be based on how often these things are clicked. It could be a recency or frequency Mo based model.

So the idea of getting better search results for the users’ intent [00:22:00] might mean that one of these models, depending on their intent serves them better. So you can think of essentially having a vote across these different models. Maybe a one to say you get a vote and a zero to say you don’t, or taking sort of a weighted balance across all of these different relevance models to decide for, for.

You know, a search at large, what’s the, what’s the best combination of these different models that will give our everyday user, our standard user, our 80% of users. If you’re thinking more along the 80 20 rule, what’s the, what’s the best balance of these different models to provide. A relevancy in other, in order to get the most relevant item at the top, based on the users in tap, [00:23:00] of course, the, the learn to rank which is the mechanism we’re talking about.

When we talk about balancing these these different relevance models and. The facet, the facet affinity we talked about earlier are also examples of things that will get smarter over time as, as they’re used more often so more reinforcement learning in these areas as well. And then of course the, the real trick is, is once you sort of settle on

What’s your, what’s your best model for, for your general? For the general public, you can start looking at micro segments or just segments in general, but another another, if you, if you have the data one of the things you might be able to do is use some either. Supervised or unsupervised learning to segment your [00:24:00] users into these micro segments.

And what that allows you to do is everything we just talked about. But to do it in parallel, not only for it, for the entire populous of users, but for each micro. Each micro segment at the same time. So for example, the facet affinity there might be a different balance or a different priority a ranking.

So for example, would it be. A price sensitivity. Should we always assume that the facets that have to do with quality and pricing should take preference and of, and of course, when we do the results when we display the search results would the, would things like pricing again, be something that we use to drive relevance.

[00:25:00] So essentially using these classification models. To find which segment or micro sentiment your user belongs to allows us to essentially pick a, a different a different formula, a different vector to apply to these different ranking models so that we can do a, a just in time personalized

Ranking for our for our search results. And of course We’re refining these with feedback. Most of that feedback of course, is taken from the, the calls to action made by the users. You know, whether they drill deeper into the item, into the search results how many pages they go down, whether or not they research whether or not they refine.

So again, this is just one example. Of how these different types of algorithms [00:26:00] or types of learning are applied to something like search. They can be applied to many other aspects of your eCommerce journey, things like assortment optimization guided selling and, and, and so forth. So I hope this example is helpful.

I wanted to go back to the, the roadmap of AI at and ML and in product discovery that that rod introduced. If you walk around this wheel, You will see lots of different areas where you can find some some actionable items to, to enhance your current retail commerce site. Many, there are many point solutions that will.

Add to these areas as well. Whether it’s in the analytics space or, or in the recommendation space, [00:27:00] average issue, we believe that the, the total power available is, is kind of a function of. Using these things together. So we make, we like to make sure that when we design something, we have we bake AI and ML and analytics into everything we do.

So that rather than talking about a single view of the customer, single view of the catalog, we believe with a single, with a single view of the entire system. We have the best opportunity to leverage AI and ML along the way. We have the best opportunity to use these new algorithms as they evolve.

And and in, in general, just bring that, bring the value of AI and ML to our customer. [00:28:00] I’m going to pause here for a moment and see if we have any questions from the group. Brad,

Back on. So I did have a couple questions that came in. Let’s see. What’s the best one to start with?

What is the easiest way to get started with AI ML for a, a medium size online retailer? Either one of you guys wanna take that one.

Raj: Easiest would be to, to sign on to when we can, we can take care of it for you. See the models are available with, with everybody, right? It is about training the models properly using the data properly and, and having the models. Self flown is the. And that requires key data science expertise, ML expertise that is required.

Right. So if you have that in-house, you know, you should go ahead and do it. It’s, it’s, it’s, it’s, it’s not that easy. So that’s where comes in. We, we can help you build that you know, leverage our framework that they’ve built and get you started off and you should be [00:30:00] able to, we will train those models using your data.

In six to eight weeks and you should be able to have an ML powered search and product recommendations in less than two months. Okay.

Host: Thanks, Raj. Here’s another one. Maybe Tony, you can take this one with machine learning based search. Do I still have a daily review of not found results and tune to reduce them?

Tony: Ah So that’s two separate questions. I think like one is, will I still get a bunch of it

Host: is two.

Tony: So two, two separate questions. Yes. Yeah. So, so there that’s, that’s a good one. Cause technically speaking as your search becomes smarter, the number of Empty searches will reduce over time. Also what’ll happen is the, the number of empty [00:31:00] searches reduces for two reasons.

One, because you are your finding better items enable to turn some of those. Empty searches into non empty searches, but also because rather than saying, oh, I have no searches. You’re better able to guess at what’s likely being asked for as opposed to, to showing up an empty search. So you’ll still, you’ll still get them your ability to make better guesses will give a better experience to your, to your customers.

But if you’re using, if you’re using some good reinforcement techniques the human factor of spending that time. Kinda curating, what are those gaps should go way down? So the human cost of, [00:32:00] of dealing with empty or gap items should go down. Now again, this isn’t, this isn’t a magical beast here.

You’re, you’re going to also be. Leaning heavily on your, your analytic system. It may be very well that you are, you’re getting these, these gaps because there’s a, there’s a hole if you will, in your data. Right. So there, there, there are things that, that just simply require you to go back and beef up your data that You know, we, we had spoke about spoken about using NLP is probably a great way to beef up your data without actually adding new data to it.

But there are also plenty of techniques for enriching the data you have by creating more relationships between that data and external data. These two will will reduce [00:33:00] your, your zero search your yeah. Empty search results. Okay. Thank you,

Host: Tony. While we’re, while we take this next question, Tony, can you share your slot to slides the ending slides again?

I don’t think you’re sharing. No worries. We do have a, a question here. Hang on just a second.

Okay. I think this one’s gonna be for you two, Tony. What is the lead time before the ML engine can really start to surface relevant results and recommendations? So how long does it start before it? Ah, good,

Tony: good question. And of course the answer is, it depends. It depends on how well you bootstrap, which is again, If you’re using NLP which will get you a good base.

Many of these other algorithms like learned rank and what have you will take anywhere between [00:34:00] three months and six months with standard traffic. Obviously. The more traffic and the more buried the traffic the faster impulse work, but the nice part about these algorithms is you know, they’re the they’re boosting something that people are already using naturally.

So for example, if you’re using a learn to rank. To make better guesses and you don’t mind looking SI silly. You can, you can simply just turn it on and, and it’ll get better over time. However, if you don’t wanna look silly you can, you can simply use a confidence a confidence measure.

And do something where, you know, you only turn it on when the confidence hits a certain threshold. And then of course you don’t even have to be that technical. You can kind of watch it by eye. And let it, you know, [00:35:00] when it, when it seems to work in a, in, in a mechanism that you can observe you can turn it on at that point, but, but typically it doesn’t take very long, certainly less than a full retail season to get these things really optimized.

Raj: One thing that I would like to add on to that is that not all of that training required to be in production environment. Right? So this can be based upon a historical data. The training of the models can start there and, and once it goes, live into production and it can be further optimized. So it, it’s not that it has to be in production

We just start learning historical data comes into play here. And the real time context of learning from that can happen from the production. So it can be split between your in an offline environment, the learning, and then in, in a real time

Tony: production environment, that’s an excellent point. Raj, thank you for that.

No problem. Okay. Thanks [00:36:00] guys.

Host: We have another one. We’ll try. I think we have two more, so we’ll try to work through those real quick. We’re coming up on time here, but so I think this one is for you too, Tony. If we have a multilingual site, then we will the ML be configured separately per local, or will it be at site level?

Meaning can we leverage the data and browsing behavior trends between local.

Tony Moores: Yeah. So, so that’s a, that’s a brilliant question. Typically I would say when you’re dealing with things like natural language where your bootstrap is a dictionary it’s. Most search engines will give you the ability to either use single dictionary or multiple dictionaries.

Language is a tough one because it used to be, you know, let’s go back to what Raj was talking about. Right? Historical data matters. If you break up your model [00:37:00] Accord across languages, then all of the data, all of the observations, all of the interactions you’re, you’re developing them sort of one at a time for each, for each language.

When in fact. Maybe the language doesn’t matter quite that much for the majority of the things that count. So separating your NLP makes kind of makes sense. But for the other models, like the the examples we were giving the, the facet affinity. To learn to rank for the, for the search results, you’re probably depriving yourself of observations by separating or breaking that into smaller groups.

Right. It it’s much easier for [00:38:00] machine systems to filter out noise. Than it is for them to, to create new connections when they’re, when they, when they don’t have the data, the observances to build upon. If that makes sense. I think, yeah. So I would, I would say do them together if, if you. Okay.

Host: Let’s get through this one.

One more for you, Raj. And it relates back to your presentation and maybe some of the video that you guys will see later, but have you direct sales impact you know, among clients or whatever improvement in search

Raj: using using ML? Yes. 100%. The simple answer is that a ML based search engine will always do better than a non ML based search engine.

So [00:39:00] understanding customer intent, understanding your products, catalog. And what are the assortments that we need to show to a customer? All of that are done really well with an ML based search engine and the search revenues, the engagement, the AOB, all of them will go off significantly with an ML based search engine.

And we have seen it with all our customers in the past.

Host: Okay. Awesome. Well, that we’re right at 1215. So we’ll go ahead and end this, this session. Tell me if you can advance to the final slide there. Thanks again to, to rise and Tony for for a great presentation and the great information. Thanks you all for the taking time to join us.

We really do appreciate it. We do have a special offer for you. You can get your site, your website, search and recommendations engines evaluated for free. If you go to this website here you should be seeing a link in your chat box here in a minute. That’ll take you right to it. If. If, if, or I, we’re also gonna send you the slides, so you can get to these links, but you can get a search and [00:40:00] recommendations, evaluation, where you can see how you stand and how your website’s performing.

And some of those things that we talked about, and you can also see how your competitors are performing are performing. So you can kind of see where you stand and if you’re losing ground or not just with your search and that sort of thing. So if you’re interested, please sign up for that. See the final thing.

If you’d like to learn more about RETISIO and our AI ML driven, intelligent e-commerce platform you can schedule a demo anytime by dropping an email to sales at retisio.com or by clicking here.  So just FYI again, we’ll be sharing the recording and slides with everybody via email, and we’ll also include that video that you can watch.

So we appreciate you being here again Have a great rest of your day and be again on looking out for that email that I will have some of that information for you. Thanks everybody.

Other resources

Read a case study

Learn how a large electronics and appliances retailer improved sales conversions by 500%.

Read our blog

Our blog is a great place to get the latest news and stay informed on eCommerce tech and trends.