Craft Code-Free Personalized Recommendations with AI - AI PM Community Session #35

Key Takeaways

[Music] welcome everyone for this session I'm super excited to be presenting um the topic today um for me about me I'm a product manager working in retail for a while now and uh doubling into recommendation systems so the topic today is on recommendations um so um if you could share the slide deck in the meantime I'll share my screen okay sure I'll be I'll be sending the link to the uh slides in a second right let me find the link there you go here's the link I just shared and I'll be sharing it with folks on LinkedIn as well all right so I'm believe you should be seeing um the presentation deck slide and not my speaker notes right that is correct all right wonderful so let's get started so you know again welcome everyone for this session wonderful to see so many of us attend this live session today and as a fellow product manager doubling in AIML recommendation systems in retail I've seen firsthand how they can solve real customer problem s and fuel growth in this session we will explore the recommendation systems we will look at different types um and then we will run a Hands-On lab setup in algolia recommend so algolia is so we'll talk about what algolia is uh later on and towards the end um the takeaway for you from this session is you would have a very simple but but powerful personalized recommendation system up and running all right so what are recommendations imagine walking into a store with endless items what if there's a friendly assistant who gave you suggestions based on your interests and that's the magic of recommendation systems they act like your personal guides recommending products and content that you will enjoy based on your preferences your choices and making your experience most fulfilling these systems are all around us in a world where consumers are bombarded with choices and can effortlessly switch Brands standing out by creating unique personalized experiences is no longer an edge for companies it's it is a necessity and lastly personalizing recommendations need a product mindset and that's you know essentially the essence of this today's session where I'll provide insights as we go by on why and how this is done you know it's about bridging gap between what customers want the right Tech to deliver it and the business goals which are driving it which means you're crafting experience es that resonate to your customer on a very personal level um in I have posted a link in this U you know slide which says read you could definitely take a look at it and know go deeper on this topic all right now let's dive into the different types of recommendation systems now recommendations is not something that know was brought in by IML recommendation systems or recommendations can be achieved even with traditional rule-based systems but what does machine learning give is this ability to avoid needing to Brute Force fine tune recommendations and handling data at scale yeah the systems the AIML systems help identify patterns and Trends in user data and which allows you to serve your customers better now let's look at the three types of recommendation systems which essentially use supervised learning and these topics are already covered by Mahesh the other Mahesh essentially who has been teaching so aesome on this on this cohort so he has already covered this uh and I've linked or mentioned the community sessions where you can deep dive on each of those so we'll just skim through those first is the content based filtering where recommendations are based on the users interactions like purchases or intentional kick clicks that they are making so what we are doing here is we're calculating the simil similarity of items using different attributes in the product like a product title description of it's a movie recommendation genre the main assumption here is that users who express interest in an item is likely to prefer similar items for example if I watch a John Vick movie or like I'm a big fan then I will enjoy other movies in the sequel yeah next is the collaborative filtering where recommendations are based on actions from other people yeah this uses the collaborative power of ratings provided by multiple users now these ratings could be implicit like like a thumbs up or you know a share or a forward or you know explicit to now for example you know consider two users Julia and John they are friends having similar tastes if the ratings with both of them have specify are very similar then the underlying algorithm can identify their similarity and even in collaborative filtering there are two types one is user based you know as I just gave an example of Julia and John right the main idea is that people with similar characteristics share similar tastes and then second is item based where a similarity between items is calculated using people's rating so think about it is like you know a movie which is watched more so often by people it bumps up on the popularity rating and the third type is the hybrid filtering strategy for recommendations here both content-based and collaborative filtering are often used together and they complement each other each other's shortcomings so the content based filtering its key shortcoming is over special ization those recommends tend to be similar to other items that the user has seen and on the other hand with collaborative systems they cannot be effective for recommending let's say new items or new people who don't have build any history yeah so these two approaches of content based and collaborator are often combined together in the hybrid system and you know what in our lab we will be looking at all three of them so just hang on all right challenges uh as product managers I think this is one thing that you know we need to really be looking at um and essentially this is where your product sense will have to be you know U ized so there are these four main challenges with recommendation systems the at least the AI ml based ones one is the cold start and this is especially true for the collaborative filtering technique what's the problem here recommendation engines struggle when they lack data new users there's no browsing history or purchases what are you going to recommend or new items yeah no ratings or feedback from those users how do you judge their appeal so that's the first problem second is sparity data sparity now e-commerce users rarely interact with each and every action on your site yeah like cards adding to cart purchases yeah this lack of data which is data sparity makes it hard for recommendation engines to group similar users where most ratings are missing leaving empty spaces yeah so that's data sparity now third one bias and fairness I think this something that we are pretty well uh you know in The Knowing for a imlb systems when the algorithms are biased they may favor certain groups over others limiting their access to a opportunities or information yeah or even under representation some groups may not see relevant content if their preferences are not well represented and last is explainability the as I called it the blackbox problem yeah often you don't know why an algorithm recommends something now frustrating this is frustrating for the users users might not understand why they see certain recommendations or you as a product Builder it's hard to identify and fix issues if you don't understand how the algorithm is working so these problems are something to really look at we will not go into approaches to solving these problems in this session given the limited time instead I have posted a few Links at the in the last slide of my deck that will allow you to Deep dive into this so I'll just pause here to see if there are any questions comments uh before moving on I be good oh let me quickly see if there are any questions popping up if anybody have any questions please raise your hand uh I don't see any questions somebody just actually can you give more examples of biased algorithm that's a great question actually yeah so uh in terms of biased algorithm there's this one example um which came through uh about um well I don't have uh one at the top of my mind if anyone from the group what would want to share uh please do so okay anybody wants to share any examples okay um all right cool so let me just see um more questions coming in one question is like can could algorithms from one country be used for other or their ethnicity do you see any problems with that well no I don't think see any problems with that but based on my experience uh what is most important is understanding your data and uh this is something we will talk about uh in the next slides but you really need to understand your customers one and number two understand the data which is powering these recommendations so you kind of like not only slice and dice it for analytics and to understand patterns but also understand if inherently there are any biases coming up and this would come up if the data is based on one or generated on one set of users you know and then that is the algorithm is used for a generic audience so that's definitely U you know something that as PMS you should be looking at a quick question sorry this is Raj again very 10 seconds question so most of this U the biased algorithms you are giving for your store decision makers or you are giving for the end user decision making or like a customer who is for this the end uh consumer for this are users of your website who are logging in and purchasing products for best example is Amazon right as you know as users you go to Amazon you look up a product and when you see a product you see prodct more collaborative analytics so you're you're recommending collaborative analytics based on the the user patterns yeah so it's both collaborative and content based uh is what I would say got it enough we we'll look at in the models yeah good question so um I can see that um there are a couple more uh questions that coming up um so let's see so I think uh uh bias recommendations based on gender ethnicity yes that's that's one of those things that uh has been observed a lot it really goes back to the data that you've used to train and sometimes U that could cause some very significant issues and I believe um so we do have a case study on this actually um in our cohort probably remembers this uh the Microsoft example right where it was it was basically there was like some bias that caused that particular product uh to run into issue so I won't get into too much just because we got a few things feel freee to say a couple words there uh but then I think uh let me see one more question uh if there's anything that um you know would be relevant I really interesting point Monica brought up right like said potential biased for recommendations based on gender and ethnicity resulting in impacting your credits risk uh that's such an interesting uh use case that you find out so thanks so much for that uh let me see if anybody has posted anything on LinkedIn real quick as we're and I'm waiting um yes content based recommendations. true uh that is true as well and that could also cause uh you to kind of go down a particular path for a user um all right sounds good so um you can continue from here on I think at the end we'll have more time for more questions as well it would be great to kind of get to the lab section of the session got it let's do that yeah and I'm sure there should be would be more questions on the lab that we look at so all right so we talked about these different types of recommendation systems we looked at the common challenges with using these systems you and as a role of product management manager what you would need to be doing here now despite all of their potential recommendation systems having reached there now or rather despite all their potential recommendation systems haven't reached their full potential due to various hurdles and these are common hurdles across different AIML systems you know especially or even in the geni space a building and maintaining these systems can be expensive and technically demanding and also it requires specialized skills additionally data privacy concerns might d some companies right especially with you know when we talking about customer data but hey there's good news that there are low code solutions that are breaking down these barriers platforms like algolia recomend which is the you know topic uh that we will look at and the product line we look at they offer userfriendly tools to implement recommendation now alol is a company which is the brains behind search experiences and recommendations on some very popular apps know eBay twitch Airbnb so these this company has been in the market for a while uh and is is one of the best uh you know low Cod solution providers out there and Amazon personalize uh you know is another product from Amazon I didn't you know get to check out check it out a whole lot but looks very promising and what we'll be doing today is we'll be looking at alola's recommend product offering this offering provides uh you know ingesting recommendation in users customer Journey at the point of where they meet the users Journey yeah for example like homepage on the homepage you will have trending items popular items you or let's say on the product page given a if you're seeing a shoe on Amazon you would essentially see other products other related similar uh shoes you know available or search results or even checkout page like in the checkout you you may be uh offered other products which are similar to or go with what you're buying let's say you're buying a shirt then you might have like you know trousers belts and other accessories available yeah okay cool now the fun step yeah so let's dive into looking at algolia recommend product yeah so here's the link so you can copy let me paste the link here and open the link out here thank you I'm going to share the link with on the LinkedIn Cloud as well and we can so while ban is doing that essentially this lab will involve a three-step process one we'll prepare our data two train the model and three preview the recommendations okay so let's start from here right on this page you essentially would kick on click on get started for free to get you onto this registration page essentially know you basically register and sign up uh I'll use my GitHub account to log in now I'll just wait for folks to you know complete the registration process and you should be looking a page something like this it says my application and let's get you started and requesting you to create your first index I'll just pause here uh give folks some time and then you know we can continue other the way one thing disclaimer we're not making any money or have any sponsorships or anything from aolia just putting it out there um we're just trying to kind of use the business tools that uh people find helpful as product managers we are going to live in a world where all these different local tools become kind of part of our daily job uh whether it's U building a agent on Amazon bedrocks or a gbt store by open AI or something else but uh we have to be well aware of all these new tools that are becoming more and more relevant to us so quick disclaimer we don't have any sponsorships making any money or anything they're not even a word of we doing this so true neither do I work at Alia so all right okay so this is very interesting it just to morning I believe they change the interface we'll still go through it anyway so are we good uh um have people been able to log in should we move ahead ban yeah go ahead uh continue and anybody who's like kind of uh stuck somewhere I think you can always watch the recording afterwards and kind of take it from there cool all right so essentially you will be asked to create um an index so you just provide a name Fubar any company name say create index X and then just so the interface has changed and kind of like uh you know going back and forth between these two and when you say create index It'll ask you to either import your records or use sample records so what you should be using is a second option of using sample records yeah so essentially uh what we did is I selected uh a sample set which algolia already provides it's a record of movie you know and essentially what they provide is all right I think you can directly jump jump on recommendations so one second if I select the sample data set it's asking me a drop- down box of e-commerce s media so you choose either one of them e-commerce media yeah you could use that one too yeah or you could use movies both of them are good but you know talking about media because I don't see a movie as an option got it if you see did you see this particular page yeah ask I I gave a name and I created an index and then it it says import your records it say upload records or select sample data set and when I click on the select sample data set it's a drop down box correct it's saying again uh it's saying again choose between e-commerce SAS or media yeah so pick the first one e-commerce okay yeah as I said I think looks like they have personalized my experience also because I've been using this for a while now that the previous screens are not there but yeah when you choose use data set it actually gave me the option to choose movie or eon and MOB ah so I think they're doing AB testing because for some we are seeing this for the other that's a whole new different experience but nonetheless I think eventually you should land where you have that use your sample index it will have a bunch of records and then you should be able to see you know the top few records from that sample so you know for this movie index as you can see uh there are a bunch of movies here are listed and when you look at attributes you should be able to look at all of the different product uh you know movie attributes available here yeah so uh okay I have a question it's it's asking me it's asking me to actually choose um data to display like which you choose like maybe we can have the same um you know I can I can send my screenshot uhuh yeah so I think uh what you could be looking at uh is this is this is an earlier screen so as U you know I think vij selected right so when you selected the e-commerce index you essentially should see these records um showing up y I can see that the mobile apps product lines okay good good yeah so Valentina yeah so check it out uh you know uh see uh do you want to look at or you know do you want to describe what you are seeing right now it's okay I just send a screenshot I I think I you can move on just I'll you know got it all right so let's do that right so for few you should be able to see the e-commerce index of essentially mobile product with their name brand description category price popularity so this is essentially your product catalog which you are giving to algolia yeah now creating this product catalog is probably one of the most important and U most engaging experience as a product manager you know this catalog one most essentially needs to have clean data um so this is one of the key step in preparing this data is removing R rened items incomplete uh products products that are irrelevant or duplicate products this is to this is very important to ensure uh that your model is you know generating accurate recommendations and as you see here the algolia recommend provides you an ability to look at your content to see what are the different attributes available what are the different products yeah all right so from here um we'll be getting to the next step on the left hand you should see um a button called recommend it has the algolia recommend logo on it so clicking that you get it into the the recommend features so this is where you will train and review the recommendation models now there are a bunch of models out here and then we will we'll definitely talk about these we'll Deep dive into these so uh these models uh help generate recommendations for different use cases for example the first one the compliment recommendations this will allow you you know this model you can deploy on your checkout page where it could display frequently purchased items together right so in Amazon as you you know as the experience is so this is essentially the uh alternate recommendations model so this you could display on your product detail page so let's say if you're showing a sneaker you could probably recommend other sneakers of of from different brands different types trending items this is probably what you would hit uh on the homepage for any website that you visit and cly what are the most popular items or for Netflix let's say what are the most popular movies that are being watched and then something called trending facet value this is essentially could be used in a category page think about is like you are you have a category say sweaters yeah so what are the popular sweaters that uh that are being um you know trending um for for the season so what we going to do next is select the alternate recommend model and I'll explain you why in a short time so so this model uh allows you as I said helps you to maximize conversions expose your catalog so give if the user is seeing one product they can see other products that you have there in the catalog using this model yeah so what we will do here is Select our source data source which is our index so I just created a full index uh oh I've already train this but essentially you will select your index here then come to uh defining uh content based attributes now let me go back and open the one that I already have so this is how you will see this option of defining key attributes now this is something to definit Ely look at these attributes can help generate accurate recommendations now the selection of key object attributes here is a very crucial function um you know as a uh AIML team the attributes you choose can significantly impact the user experience and the effectiveness so for this e-commerce um site the catalog of you know mobile devices I have picked three attributes name description and brand so essentially what we doing here is letting the model know that hey I want similarity to be built based on these product attributes so let's say wherever the name is similar or the product description is similar uh so these would provide an Insight in terms of how the similarity function will identify which items are similar yeah so once you do that uh at the bottom let me go back you would click on start training everyone here so far so this will take a few minutes so what we'll do is so this is how it will look so you'll have your model you know your Source index updated you know when it was last updated and the status would should be showing here you know in progress essentially so this model I have kind of put together a while back so it's already trained but let's have your model kind of get fully bit and meanwhile let's go and explore these model types that we have here when you click training what what data are you using to train the model your catalog data the whatever the index that you the sample index that you chose at the start after registration is what it is using for training doesn't it need more than just product data catalog yeah right you need user data the r or purchase data that kind of stuff right yes absolutely that's a great point point that so the reason why I you know Jerry pick this particular model is this model uses uh the hybrid uh scheme of using collaborative and content based filtering which means that when collaborative filing is not there you know essentially the data that vatt was talking about right you know the purchase history the conversion history of users since we just created this new index that data is not available with us so what this model will do is it will fall back on content based U you know recommendation system to calculate the similarities so what option you have is essentially at the start when you you know created your model you would import your records in terms of both not only your catalog but uh the user interactions and what algolia also provides is a way to hook up algolia to analytic systems like Heap analytics or Google analytics which will continue to provide algolia the details in terms of user you know purchases their activities which will help build your collaboration model right we good does it answer your question yeah yeah so so what you're saying is that uh you import additional user data right if you want to do collaborative filtering correct yes okay got it yeah so if you see here you know step number two talks about send events using inside API or by uploading your historic events so algolia allows you two ways one is you could just upload a CSV file of user activities or plug in your analytics to the API which will then started start to stream the events data in real time which provides for the collaborative experience okay so while the model is still getting ready let's kind of take a look at what these models are yeah so each of these serve a different purpose so let's look at the first one which basically the first one is you know recommend frequently bought together now this model recommends items that are often purchased together by customers yeah it's like when you go to a store and see products displayed together because people tend to buy them at the same time so this frequently bought together feature that where you are you know deploying your model uses collaborative filtering to make these recommendations this means like it looks at the behavior of many users to find patterns in their purchases for example if a lot of people buy a phone case when they purchase a new phone this model will pick up on that and start recommending the phone case to anyone who buys a phone right so let's say if you buying a particular model on iPhone on Amazon there's a likely chance that you might find uh a phone case recommended to you and that is happening through the collaborating filtering technique which this modu use yeah so essentially the model is trained on users conversion events right so when we say you know what are the users collaborating on is essentially the events right and what kind of events a user clicking on an item in the search box to view its product page so that's a high indication that the user has interest in in a product yeah another example is where a user adds the item to their shopping cart so that's uh another example so all of these conversion events are used for this filtering technique uh in order to recommend products and um yeah so to train this model we need at least 1,000 conversion events where two or more items were purchased together in the last 30 days this ensures that the model has enough data to learn and make accurate recommendations and that is for I mean that is one reason why if you go try this model now you won't be able to do it because you essentially don't have the events data so either you you know can curate one and upload at a CSV or hook it up to one of your apis to train this all right so let's look at the next one the related products I mean this is the one that we are kind of baking right now so this model recommends items related to one another yeah as you can see in the image here these are these are great for product uh detail Pages where you know it helps to increase um your catalog discovery so this model employ or uses collaborative filtering or and hybrid filtering both so here what happens is uh essentially collaborating filtering is used uh to learn user interactions it observes patterns in clicks and conversion events and understand what users like for you know as an example if a user often clicks on sport shoes the model learns to recommend similar shoes yeah however there are situations where we don't have enough user interaction yeah remember the cold start problem that we start we talked about no brand new user coming to your system there are no interactions so what this model does is essentially it fall backs on content based filtering and for content base what we look at is is essentially users past purchases and recommend items that are similar in nature now for a new user there won't be any purchases either so the way this is problem is solved is there are you know a curated set of uh items that are provided to the user or you get a user experience similar to Netflix so let's say when you are a brand new user login Netflix will ask you for the genres that you are most interested to watch and those genres are then use to create content based filtering of items that you can recommend or are recommended and over time essentially you you know the engine becomes smart to improve the accuracy of your recommendation and yeah so this hybrid model for collaboration filtering require around 10,000 events created in the last 30 days you know and this ensures that the model has enough data to learn and make accurate recommendations and you know in with inference it right now recommends 30 or makes 30 recommendations per customer okay so we'll spend uh a very short time here so trending items essentially is this model is a feature that recommends popular items from the product catalog you as well as popular facets also it uses the collaborative filtering technique and is trained again on users conversion event to train this model you need 500 conversion events in last 30 days I'm not sure how they came up with this number but I believe they would have uh you know identified uh the best possible uh data chunk size to make you know to have this balance of you know accurate recommendations and just enough data all right so now let's hop back and see our model which was getting trained yeah so now given the number of records are small in this uh index you should have the model trained by now so what you do next is click on view to preview the results here so this is where you will see the power of Alia's AI in action so on the left hand side this is the item that you have viewed and on the right hand side these are the products that are related to this item so these are the items possibly you could show on your web page when a user is is browsing for this particular product so again highlighting the point about uh using the right attributes right the importance of selecting the right attributes for your specific use case is is is very important so we can play around with this uh so like an iPhone 7 you have you know the iPhone selected you have a bunch of recommendations made you can look at you know more options Here and Now what You observe is essentially the recommendation score right so this is one of the features that algolia recommendation product line provides is the recommendation score this indicates the confidence of the model in its recommendation and this can be a very useful metric for evaluating the per performance of your recommendation system the high score indicates that the model is confident in its recommendations and this can lead to improved user engagement and fulfillment I I noticed a flaw in the recommendation uhuh on the first row at the end it says uh Sprint you you you that's not compatible with AT&T right it's a locked phone right yeah where did you see the Sprint oh oh the the first row at the end don't don't don't scroll up to your yeah yeah yes now that's a great observation now the one reason why this has happened that is when we gave the attributes for the model we just selected the the title the description if we had also given the the brand or I know the the carrier the carrier yeah it would have dropped the recommendation for this guy yeah but great point I me I'm glad you brought that up so hence the need you can go and essentially edit and let's say if we do have an action of carrier uh popularity price okay don't have an option here but yeah if we had the defined carrier then that should have taken care in terms of dropping the the model accuracy for that product yeah so yeah so do play around with this essentially you can you of go around clicking on a product it provides your recommendations um and you know there is this info section towards the end where it says that the all the 400 items which are there in the catalog uh have recommendation so at least they have one or more recommendations and the products that are shown as recommendation has a 100% coverage so this may or may not be true for let's say products which are very nich you know let's say or let's say a Nokia phone essentially right you know probably out of you know out of market so you may not essentially so see a lot of recommendations for that hey mahes quick question mhm how are the attributes derived uh is it based on the data that we upload is it derived automatically for AI no it is based on the data that we have upload so if you remember when you created the index you you know that index has a bunch of attributes which your index defined and these are the attributes that are used uh for doing the recommendation hey mahish I have a followup question sorry go ahead if you haven't finished yeah so this is the existing data store that we using right let's say if I have my own data if I have my own CS file that I upload then based on the data it will look for the relevant attributes or features right and know that you can select here well for this particular model that we selected it specifically asks you to select attributes uh because you know in terms of the problem statement that you're seeing that oh okay why doesn't it just look at the attributes identify the patterns identify which of them will make sense and then you know use those ones to recommend so what problem statement that you're talking about leans more in terms of the you know neural network side of things where you know the algorithm is intelligent enough to do these additional complex U uh you know assessments but here you're using machine learning with a very simple uh Paradigm that let me provide the attributes that I want to be used for recommending if you remember mahesh's uh first session on content based filtering you know uh what happens if you remember that in this technique uh the data the catalog data is used and for the attributes that you select there is something as embedding that happens where it takes those words essentially converts in in machine talk and then for those it tries to find similarities right and then based on the similar is it figures out which items are similar so that here in this case is happening based on which on these attributes that you're selecting so Mah to add to that question just to clarify I'm very new to this um just a rudimentary question so first we pick the database s we can pick a sample database and then after that after it's loaded we can Define the index by selecting the attributes and then we generate the index is that correct no no not in those order so the index is already there it's generated okay you either upload the index or use one of the so is the the database is the index or no I I think we are collecting the cluster of data once you collect the clust of data the index from this product creates a natural key okay that natural key will give the the strategy is a non-cluster index or the cluster index ordered by class or whatever it is based on that you are uploading your data the behind the screens the algorithms have captured the patterns it is giving you predictive analysis oh okay yeah thank you Raj all right question this uh my question is that since we chose three different attributes there could be actually you know in real world scenario there could be a lot more attributes in your e-commerce site how do you make sure that which attribute is more important I mean who is doing this job uh I mean in this we defined it but in a real I I'm asking for a real world like when you are working on your products how do you design do you choose a certain model for that purpose and can that be influenced by the selection of the model so how do we decide what's the most important attribute from a relevancy point of view great question yeah so uh my understanding of that is essentially it's your AIML team essentially you as a product manager working with your data Engineers you would have to kind of do you know so what we call first feature engineering where you normalize the data that you have across the organization and once it's normalized you look at the potential of each attribute to see you know what is the potency it has to you know create recommendations that's number one so it's you know to a greater extent uh I wouldn't say it's a trial and error but it's more about the product Discovery uh that you do as part of using and implementing the recommendation models number one number two is in terms of the attributes you definitely want to keep them as minimal as possible so if you see algolia limits you to only three attributes and the reason being more attributes you have the complicated the model is going to be um to run for data storage and also for inferences you know so uh you know in this particular case the models are in are updating every 24 hours but if you want something is a near realtime system then you know having more attributes is uh be or would be counterproductive hopefully that helps answer the question yeah yeah thanks uh hi M I have a question actually just just note um we only have uh three minutes left happy to say a little bit um longer if needed but um hopefully this would be the last question so that mahes can continue go ahead Andes go ahead and continue thank you so uh I think we are towards the end this is the last okay sure go ahead okay yeah as you were mentioning about the embeddings right so the question is for the B2B recommendations uh with collaborative filtering uh what happens is the customers are not ready to share their data right but they wanted to get the insights uh so in case uh you have a cloudbased single denant customer in one and the other customers wanted to know the insides out of the data so in that case how the B2B space takes care of this is it the embedding something like that uh I wouldn't say embedding so essentially what you're looking at is uh conversion events of of one uh you know customer using those for other or even for the same customer so the way I would see it is it would be a very similar problem to how you get analytics from platforms like Google analytics or Heap you know so essentially the data is available but all of the pi information like personal information is masked out so you would have something called a user ID for a session but you wouldn't know what that user is but you will have all of the different conversation events conversion events that the user has done so by masking it in such a manner you could potentially uh use those events for collaborative filtering now when you talk about embedding this is more a function of the content based filtering so that would be a separate uh you know problem statement okay okay thank you got it all right yeah so here on the last slide um you know um by the way I will quickly show you there is another product that these guys have launched which is looking similar so looking similar is a very interesting model uh that they very recently launched which finds related items um there so it finds related items based on images in your index navigating through similar looking products can be a very uh you know great way of discovering new products which users may have not searched for so yeah so you could try give this a try uh and definitely play around with this and then if you are kind of like go a step further you can import few user click events using the CSV and then play up with the other models too and last but the least these are some of the areas that you can go into and one last uh you know thing that I want to leave you with is while recommendations model have been there um for a very long time there is a lot of innovation happening around this for example Amazon just few weeks back introduced rofus uh shopping assistance now this Amazon is broadly using um this to incorporate generative AI in all of its products I'll drop a link in the chat here but essentially you know this space is is evolving using you know multimodel recommendation systems uh graph neural networks and you know you know lot of very exciting areas uh being used thank you so much this was great uh go ahead sorry to interrupt you squ you no I'm good I think uh this kind of like concludes what I wanted to talk about I see we are a minute over so yeah definitely over to you ban awesome thank you so much mahes and um I think the link that you shared about um the Amazon Rus what I'm going to do is also share that on LinkedIn uh so link to amaz okay perfect great um I think this was a really awesome session and uh we can see how as product managers we have all these new tools available to us and understanding the data what it takes to really produce meaningful recommendations uh that's really amazing thank you so much I thought this was a really great session um so link to the deck yes um somebody provided it I've also shared it earlier um so feel free to scroll up or get it on LinkedIn we'll also have a recording of it um available to uh any body who follows Us on YouTube um LinkedIn will also have this uh thank you thank you so much Mahesh um I also want to one more time welcome uh the great people at our cohort 5 that have just joined our program I we look forward to having you all in the program this is very exciting and uh looking forward to uh going through this journey together so thank you so much again I'll be reaching out to some of the cohort members that we just want the class today uh shortly so you'll expect to hear from me briefly and uh looking forward to getting to know you all and again hi to everybody and bye to everybody from outside the cohort community and Thea thank you so much again and have a good day bye thanks my