Interview

Tomohiro Furusawa (<a href="https://twitter.com/rilmayer_jp">@furufuru</a>), a member of Mercari’s AI Team, has held a variety of positions in his three years at Mercari, including PM, engineer, and scrum master. Despite shifting among these roles, he has always been involved in improving the discovery experience at Mercari. We sat down with him to talk about how Mercari’s Recommendation Team has developed since its establishment, and some of the things he wants to work on in the future.
<h2>Profile</h2>
Tomohiro Furusawa (<a href="https://twitter.com/rilmayer_jp">@furufuru</a>) | Product Manager 
Tomohiro discovered the joys of data science and software engineering, and before he knew it, he made the leap to the world of IT. After finishing graduate school, he used his background in library and information science to develop products for information retrieval and recommendations. He joined Mercari Inc. in 2019 to work on search evaluations as a member of the Data Science Team. He then established the Recommendation Team and is currently handling product improvement as a Product Manager.
<h2>“Since my days as a student, I’ve always been figuring out how to organize information and make it easier to use.”</h2>
 ──Can you tell us a little about what you’re working on right now? 
Right now, I’m on a team focused on improving the core Mercari experience, which is the buying and selling experience. My own particular focus on that team is improving the buying experience.
When you think about customers purchasing items, and how exactly they find an item they want, the most common way is that they search for it. Searching is a process where a customer comes up with an idea of something they want, so they look for the item on Mercari, find it, and then they buy it.
Beyond searches, though, we’re also working to improve the “discovery experience.” This involves finding new purchasing needs or finding items based on purchasing needs without making any special effort to find it—in other words, without needing to search. You might open Mercari, and items you’re interested in or thinking about buying are right there. Or, while you’re searching for one item, you find another item you’re interested in, and you buy that one. That sort of thing is how we define the “discovery experience.”
Recommendations are a powerful tool in improving the discovery experience. 
Mercari currently gives customers recommendations by controlling three units. One is “content,” which is about introducing items of interest. Then there are “topics,” AKA search criteria, which are item groupings that can be interpreted as such. Finally, there are “items,” which are the items customers actually purchase.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/219d7b80-image_1-1024x683.jpg" alt="" />
<center>
Tomohiro Furusawa (@furufuru) 
Furusawa has consistently been involved in recommendations since joining Mercari in 2019
</center>
When I was in college, I worked on library and information science. Through academics I studied how to collect and store large amounts of knowledge and information and how to make it easily accessible to a wide range of people. Looking back at my career since my days as a student, I feel like I’ve always been involved in figuring out how to organize large amounts of information and make it easier to use.
When I was a student, I thought about a career as a librarian, but when I started interning as an engineer and data scientist, I found myself really interested in manufacturing and data usage. So, after I graduated, I decided to become a data scientist at Recruit. Recruit, which is where I worked before Mercari, has a lot of contact points with customers through all sorts of life events, so it was interesting just to look at the data. That said, I was also almost painfully aware that data is meaningless just by itself, and that its real significance lies in how it’s used. That’s when I started digging deeper into the technology of searches and recommendations.
 What’s the difference between a “search” and a “recommendation?”
When we want a certain type of information, we use keywords and other things to ask questions through some sort of app, right? Then we look at all the results and get the information we want. That series of actions is what we call a “search,” and a search system is the mechanism that makes that experience possible. 
However, if we utilize customer data we already have, we can provide information from the service side without customers needing to enter any kind of keywords or other information. A system that can provide information without a customer having to actively ask questions is a recommendation system. A “recommendation” is different from a “search” in that it refers not to an experience, but rather to a means used by the service provider. 
Building good search and recommendation systems makes the service easier to use, and I think it ultimately bring us closer to the world I want to create, which is a world where people from all walks of life can access knowledge and information with ease.
<h2>Improving the recommendation system while shifting from position to position</h2>
──Why were you interested in a job at Mercari?
It started when a former colleague of mine was working on improving the search experience at Mercari and invited me to join. Mercari’s user base was growing fast, and the company had a lot of momentum, so I thought changing jobs would give me the chance to dig deeper into search technologies that use a wide range of data.
After joining Mercari, I started by evaluating the existing search system. All services—not just search functions—of course need to be built first, but they also then need to be improved through evaluation cycles.
With our service, a search is considered to have gone well when a series of experiences is born—for example, when you enter a keyword for something you want, you find the perfect item, and then make your purchase. On the flip side, a search doesn’t go well when, say, a customer is looking for an item but can’t find it using keywords, or a lot of items with little relevance show up in the search results.
Another issue unique to Mercari is that the search results sometimes show only items in good condition that can be purchased right away, which means the customer can’t compare and consider other items as well. The Search Team was using trial and error, repeatedly defining and evaluating what makes a search a success or a failure, to figure how to improve the search process.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/c184be5d-image_2-1024x683.jpg" alt="" />
We established Mercari’s Recommendation Team after that. I’d actually been quietly working on a prototype of the recommendation system before the team was formed, but the company also decided to focus on the recommendation system, so I ended up supporting the establishment of the team as a PM. After a while, Aki Saarinen (@akis) joined the team as PM and I shifted back to being an engineer.
Moving through the development process as an engineer, I found that a good development system hadn’t been created at the team level, so I acted as scrum master to fix the group’s development system while handling my engineer duties as well. After a while, things came together as an organization and the Recommendation Team started producing results, but there was also a shortage of PMs at the time, so I went back to being a PM again. So, although I’ve always been involved in the recommendation system, my job title has gone from PM, to engineer, to scrum master, and then back again.
<h2>Thoughts on getting recommendation features to contribute to GMV</h2>
──Have you hit any walls in your career so far?
Actually, it took some time before the Recommendation Team was able to successfully contribute to sales. Before the current system, items on the home screen were recommended to customers based on their past activity, but appearance wise, it was just a timeline listing one item after another. From there, we increased the types of components in the recommendation system to display recommended keywords and categories based on a history of customers’ recent activity. That was when we first saw a big boost in sales. Instead of just recommending items, we also display why an item is recommended and define functions for components that allow for deeper and easier exploration, and I think that has made things easier to understand for our customers.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/5fb7f5d4-image_3-1024x679.png" alt="" />
<center> 
メルカリのホーム画面の変遷 
</center>
Related Article：<a href="https://mercan.mercari.com/articles/27899/">No Two Home Screens are Alike – The Reasons Behind Mercari’s “Uniquely Difficult” Personalization | mercan ＊This article is only available in Japanese</a> 
The home screen went through big changes in terms of UI appearance during this period, but we also revamped background systems at the same time. To handle the labor-intensive process of developing home screen functions in a short period of time, we reworked most of the content displayed by the client into a form we can manipulate on the back end. We had always thought it would be great if the system could be modularized so that we could just change the back end to switch between home screens, but it took until around 2020 before we could actually develop and deploy it.
──Are there any other things you want to try or challenge yourself with at Mercari in the future?
The recommendation system is still using only relatively simple logic, so I’d like to try applying a broader range of technologies to solve advanced problems. We still need to address a lot of things, like developing data utilization infrastructure for the recommendation system. If you’re someone who loves approaching things from a variety of logical perspectives, definitely get in touch with us!
Mercari is a high-traffic service used by a huge number of customers, so someone with experience in operating a large-scale service will probably have an easier time grasping the essential points intuitively. That said, even if there are some things they don’t know, we love working with people who enjoy meeting challenges head-on!
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/3e6f9193-image_4-1024x683.jpg" alt="" />
<a href="https://ai.mercari.com/en/articles/ai/furufuru_interview_2/">In the second part of the interview</a>, Tomohiro talks about the things he focused on while working at Mercari, and Mercari’s unique organizational culture, as someone with experience in a range of positions, including PM, engineer, and scrum master.
<ul>
<li>Photo : Shun Nakayama (<a href="https://shunnakayama.myportfolio.com/">Web</a>/<a href="https://twitter.com/shunnakayama_">Twitter</a>)</li>
<li>Text : Ayumi Iga (<a href="https://twitter.com/aymmm_i">Twitter</a>)</li>
</ul>

The Past and Future of Mercari’s Discovery Experience, from the PM who Launched the Recommendation Team

Tomohiro Furusawa (<a href="https://twitter.com/rilmayer_jp">@furufuru</a>), a member of Mercari’s AI Team, has always been involved in improving the discovery experience in his three years at Mercari. <a href="https://ai.mercari.com/en/articles/ai/furufuru_interview_2/">In the first part of the interview</a>, Tomohiro talks about how Mercari’s Recommendation Team has developed since its establishment, and some of the things he wants to work on in the future.
In the second part, he talks about the things he’s focused on while working at Mercari, and Mercari’s unique organizational culture, as someone with experience in a range of positions, including PM, engineer, and scrum master.
<h2>Profile</h2>
Tomohiro Furusawa (<a href="https://twitter.com/rilmayer_jp">@furufuru</a>) | Product Manager 
Tomohiro discovered the joys of data science and software engineering, and before he knew it, he made the leap to the world of IT. After finishing graduate school, used his background in library and information science to develop products for information retrieval and recommendations. He joined Mercari Inc. in 2019 to work on search evaluations as a member of the Data Science Team. He then established the Recommendation Team and is currently handling product improvement as a Product Manager.
<h2>Why experience as both a PM and an engineer makes the development process smoother</h2>
 ──You have held a variety of positions in your three years at Mercari, including PM, engineer, and scrum master. Have you ever felt like you’re able to add value at Mercari precisely because you have experience as both a PM and an engineer?
Yes. One way is that I can approach technical debt in the same way I approach functional improvements. When you build a product, you might create technical debt that doesn’t directly affect customer value but does affect the development experience, and thus the speed of improvement over the long term, and that’s something I now pay attention to. My engineering background also made it easy for me to figure out things like how much time it would take to develop a certain function, and which elements I should focus on. I feel like I’m able to communicate with engineers quickly and smoothly.
Another way is that I can look at things from the perspective of a PM even when I’m working as an engineer. When we launched the Recommendation Team, the internal organization and development environment weren’t as mature as they are today. So, when it came time to launch the new recommendation function, a lot of changes needed to be made in the system, and I was able to communicate with everyone with a good idea of which members we’d need to push forward with the development.
To give a little more detail, at that time, we had to make all our changes while striking a balance between the client code, the back-end microservice code, and the code for multiple peripheral microservices. That required communicating with the engineers on each team, and talking to the designers and the CS team as well. I remember communicating with a bunch of different people throughout the company just to put out one function.
Those sorts of efforts made it possible to move forward with the development of the recommendation functions.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/a67047d0-image_3-1024x683.jpg" alt="" />
<h2>As a Tech PM, the ability to work with data is essential</h2>
 ──What are some of the skills you need to be a Tech PM at Mercari?
The first thing I think a Tech PM at Mercari needs is the ability to work with data. You need to be able to analyze data to drive your decisions—to look at the data in front of you and interpret what’s happening, define “success” and “failure” based on customer behavior as seen in the data, and if something is failing, determine the cause and plan your next move.
When we put a feature in place, we decide in advance how we will measure success before we release it. The easiest way to measure success is to see if the feature increased sales. That said, a single function rarely causes sales to explode, so we often follow our customers’ behavior to see if there has been any change from before a function was deployed. Sometimes the numbers change as expected, and sometimes they stay the same. Sometimes they get worse! Sometimes figures we were monitoring which led to sales improved, but sales themselves actually dropped. A situation like that means the hypothesis was wrong in the first place and needs to be reworked from scratch. This is a job where you’re constantly looking at data, making decisions, and repeating the process, so I think data analysis skills are essential.
I also think it’s essential for any PM, not just Tech PMs, to be able to keep ROI in mind constantly. Being able to think intuitively and build strategies on how to implement effective measures at a lower cost is an important skill to have.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/35fc0787-image_2-1024x683.jpg" alt="" />
<h2>“Mercari engineers aren’t just people who build whatever a PM comes up with” was a shock to me.</h2>
 ──Tell us about any particular skills you gained or growth you’ve experienced in the Mercari organization.
Maybe the ability to communicate things clearly. Mercari is a diverse organization and there are members from a wide variety of backgrounds. When I first started, there were a lot of times where I thought I was speaking clearly, but the other members and I weren’t understanding each other. When I would tell my team “this is what we’re going to do,” I would sometimes get comments like “I really don’t see the value in working on this feature.” When I looked back on why they weren’t seeing the value in what I was saying, I realized I’d been trying to communicate it too carefully, which had the opposite effect of making it harder to understand. I thought about, first, what was most important to get across and, second, where the value of what we were doing lies, from the standpoint of conveying information to members from different backgrounds in an easy-to-understand manner. Then I worked hard to strip down the amount of information and extract just the most important points. I found that as a PM, focusing on choosing information to keep or discard based on the backgrounds of the other members made it easier for us to work together.
Also, in a pyramid-shaped organization, information flows from the top down, right? Mercari also has a bottom-up flow, where information moves horizontally, and members communicate with each other to get approval from upper levels. The unique communication style Mercari uses to create this bottom-up flow is something else I’ve picked up.
 ──Is there anything else you can tell us about the unique culture at Mercari?
One of the things that struck me when I joined Mercari was that there are so many ideas coming from engineers. It really hit me how Mercari engineers aren’t just people who build whatever things a PM comes up with. Their stance tends to be “we’re not going to build something if we don’t want to or if we don’t see its significance,” so I always try to make sure I can explain the meaning behind what I’m creating.
On the other hand, from an engineers’ perspective, it’s not considered acceptable to just say “OK!” when a PM asks you to create something. The entire team must be convinced, in terms of whether there are any systemic problems, what the cost of implementation will be, and what the feature’s significance is.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/a67047d0-image_3-1024x683.jpg" alt="" />
So, as a PM, I have to communicate the value and context of the feature to the engineers and get them on board, even if the company has already decided it’s a feature that has to be made. Nothing good comes from one-sided communication where I just say, “Build this!”
Since I have experience as both an engineer and a PM, I understand how both sides feel. As a PM, I learned the necessity of having engineers work after accurately communicating background information and meaning to them, and as an engineer, I felt like I needed to be proactive about giving my opinions to PMs.
<h2>The Mercari discovery experience “still has a lot of room for improvement”</h2>
 ──Have you achieved what you set out to do at Mercari when you first came on board? 
I’ve consistently been working on improving the search and discovery experience ever since I joined Mercari. Although I think the customer experience has actually gotten better, it still has a lot of room for improvement. Mercari doesn’t have any boundaries that say you can only go so far as an engineer, so you’ll be encouraged to take on challenges, and in my experience, I was able to actually put the Recommendation Team on track as a PM. The will of individual workers is something the company values, as long as it matches the direction of the company.
Looking back now, I was lucky to have the chance to work as a PM despite joining Mercari as a data scientist, so as long as it fits the direction I want to follow in the future, I’d like to keep challenging myself in that role.
<img src="https://storage.googleapis.com/prd-ai-asset/2022/12/2298cb41-image_4-1024x683.jpg" alt="" />
<ul>
<li>Photo : Shun Nakayama (<a href="https://shunnakayama.myportfolio.com/">Web</a>/<a href="https://twitter.com/shunnakayama_">Twitter</a>)</li>
<li>Text : Ayumi Iga (<a href="https://twitter.com/aymmm_i">Twitter</a>)</li>
</ul>

Discoveries through a career as both PM and engineer at Mercari

Blog

<h2>Overview</h2>
The 1st Workshop on End–to-End Customer Journey Optimization was co-located with the KDD 2022 conference. The theme of the workshop was to discuss ways to improve the user experience across the entire customer journey. 
<h2>Key points of presentation</h2>
<a href="https://arxiv.org/abs/2207.14798">Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions</a>
We summarize our major contributions as follows: 
<ul>
<li>We propose a framework of promotion decision making by targeting customer direct and enduring effect. Our framework consists of direct and enduring response predictions, incentive allocation decision making and business impact evaluation. </li>
<li>We propose a customer direct and enduring effect (CDEE) model, which provides predictions of the direct purchase propensity and the enduring purchase amount of each customer. </li>
<li>We compare the CDEE model against benchmarks on two promotional campaigns in our company and achieve a significantly better performance.</li>
</ul>
<h2>Background</h2>
In the e-commerce marketplace, marketers run various types of promotions as a part of the customer relationship management (CRM) strategy in order to maintain customer relationships and guide consumers towards desired actions. In Mercari, multiple treatment promotions are also a common marketing strategy. To better cater to each customer’s needs, personalized promotion decision making is necessary and has garnered the attention of the research community. Some research optimizes promotion decision making by applying purchase prediction. While the modeling and optimization on the direct purchase probability brings profits to the business, there is still room for improvement. 
First, different customers contribute differently to the revenue even when they complete a purchase. For promotions which aim to maximize revenue, we should consider not only the purchase probability but also the purchase amount in the decision making process.
Second, promotions can induce purchases but these often only turn out to be one-shot deals. On top of that, promotions might cause purchase acceleration or stockpiling. Therefore a post promotion dip in purchases is widely observed in marketing. To create promotions that aim to build long-term customer engagement and boost customer loyalty, the post-promotion period effect should be taken into account.
<h2>Summary of paper</h2>
To target the efficient customers from the perspective of revenue during and post promotion, we propose a framework of promotion decision making by modeling customer direct and enduring response. The direct response corresponds to the customer’s re- sponse during the promotion period, while the enduring response corresponds to the customer’s response during and after the promotional period. Our framework consists of both direct and enduring response predictions, incentive allocation decision making and business impact evaluation. 
We propose a customer direct and enduring effect (CDEE) model which provides predictions of the direct purchase propensity and the enduring purchase amount of each customer. 
We formalize the incentive allocation problem. Our optimization goal is to pick the best incentive types for customers maximizing the total enduring purchase amount while keeping the cost under budget. 
To estimate the effect of decision making, we apply an unbiased evaluation approach on the business metrics with randomized control trial (RCT) data.We compare our method with benchmarks using two promotions in Mercari and achieve significantly better results on both model evaluation and business metrics.
<h2>About MDS Team</h2>
In the Marketing Data Science (MDS) team, we use statistics, machine learning, and mathematical optimization to improve the efficiency of marketing campaigns.

Mercari AI’s research paper —”Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions”—accepted to KDD2022 1st Workshop

<h2>Overview</h2>
The 1st Workshop on End–to-End Customer Journey Optimization was co-located with the KDD 2022 conference. The theme of the workshop was to discuss ways to improve the user experience across the entire customer journey. 
<h2>Key points of presentation</h2>
<ul>
<li>Discount coupons are one of the marketing tools used by Mercari.</li>
<li>The effectiveness of the discount coupons has a strong effect on the business impact of our marketing campaigns. </li>
<li>The timing of a coupon can be a crucial factor in the effectiveness of the coupon. </li>
<li>We used uplift modeling to estimate the effect of timing on the effectiveness of the coupon without the need for a dedicated randomized control trial. </li>
</ul>
<h2>Background</h2>
Mercari uses a number of coupons to motivate users to buy and sell items in the marketplace. The effectiveness of marketing campaigns (and therefore the improvement on the bottomline) is strongly dependent on how effective the coupons used in the campaigns are. While many factors can affect the effectiveness of a coupon, one can hypothesize it to depend on the timing the distribution of the coupon was scheduled. However, verifying such a hypothesis would typically require dedicated randomized control trials which costs time and effort. 
<h2>Summary of paper</h2>
<a href="https://drive.google.com/file/d/1U8ILIrbPmDh4SBNARgOKomOkcXmBklXE/view">Estimating the Effect of Timing on Coupon Effectiveness</a>
In this paper, we performed a natural experiment with uplift modeling to measure the effect of the timing of the distribution of the coupon on its effectiveness. This circumvented the need for a dedicated randomized control trial thereby saving on resources. Using uplift modeling, we found that the quicker we send coupons to users after they create an account, the greater the effectiveness of the coupon.
<h2>About MDS Team</h2>
In the Marketing Data Science (MDS) team, we use statistics, machine learning, and mathematical optimization to improve the efficiency of marketing campaigns.

Mercari AI’s research paper—”Estimating the Effect of Timing on Coupon Effectiveness”— accepted to KDD2022 1st Workshop

<h2>Overview</h2>
Earlier this year, the CRE ML Team's research paper on violation detection systems in chat messages was accepted to the <a href="https://sites.google.com/view/ecnlp">ECNLP</a> (European Community for 
Neuro-Linguistic Programming), a workshop of the ACL (The Association for Computational Linguistics).
<a href="https://aclanthology.org/2022.ecnlp-1.8/">Textual Content Moderation in C2C Marketplace - ACL Anthology</a>
The <a href="https://aclanthology.org/">ACL</a> is one of the industry’s most prestigious language processing conferences. ACL 2022, the 60th workshop of its kind, was held in Dublin, Ireland.
<h2>Key points of presentation</h2>
<ul>
<li>Implemented a machine learning system to detect violations of the terms of service written in chat messages</li>
<li>Improved violation detection accuracy while significantly reducing the number of manual checks by customer support</li>
<li>Created a system that is easy to maintain and continuously improves detection accuracy</li>
</ul>
<h2>Background</h2>
After an item is purchased on the Mercari marketplace, the seller and buyer can freely communicate with each other via a chat system. To facilitate smooth transactions, harassing behavior such as making defamatory or threatening comments, is prohibited in Mercari’s terms of service. However, it is difficult for customer support to visually check all such communications, and therefore Mercari uses a system that automatically screens for violations.
In the past, the system used heuristics such as keyword searches to detect violations, but the detection accuracy was limited. However, because trouble can easily follow communications between sellers and buyers, and there is a significant risk of overlooking a violation, it is necessary to ensure the reproducibility of violation detection (also known as recall). An additional issue was that the number of false positives remained high, even if a conversation had passed the heuristic screening, and therefore customer support representatives had to check huge volumes of conversations manually.
In this study, researchers implemented a system that used a machine learning model specialized for the aforementioned screening process. 
The study reported that implementing the system reduced the number of false positives while ensuring reproducibility, resulting in a significant reduction in the number of conversations that needed to be checked manually.
<h2>Summary of paper</h2>
After a buyer purchases an item on the Mercari marketplace, the seller and buyer are able to freely send chat messages to each other. However, there are cases where the content of these messages makes sense in the context of the parties’ conversation that started on the separate public chat message board used prior to the item purchase. We therefore devised a machine learning model that makes inferences based on the context of the previous conversation, rather than estimating whether or not the text in isolation is in violation of Mercari’s terms of service.
The model itself is designed to be extremely lightweight, and can make inferences at a cost that is reasonable in relation to the vast number of chat messages exchanged on Mercari. In addition, the architecture allows each type of violation to be added, modified, or deleted with ease since the content of violations stipulated in the terms of service may be modified in the future to adapt to changing social norms. The moderation staff then checks the conversations that the machine learning system has screened. The decisions made at this time are added to the training data as true labels. In this way, a feedback loop is created within the violation detection system, which is expected to improve detection accuracy continuously.
In addition, Mercari confirms chat messages from the viewpoint of preventing fraud and trouble, in compliance with laws and regulations, and with the consent of each user.
<h2>About CRE ML Team</h2>
The CRE ML team operates several other machine learning services and provides technical support for the safety and security of our users.

Mercari AI’s research paper—”Textual Content Moderation in C2C Marketplace”—accepted to ACL2022 ECNLP Workshop

Hi! I&#8217;m <a href="https://twitter.com/MLBear2">ML_Bear</a>, an ML Engineer on Mercari&#8217;s Recommendation Team. 
I&#8217;m in charge of improving the recommendation components displayed on the Mercari home page and elsewhere. Today I&#8217;m going to show you examples of how I improved the recommendation logic by using Mercari&#8217;s massive amount of log data and the metadata assigned to items sold on the Mercari app by users. (More details on the metadata below.)
<h2>Item Metadata</h2>
To improve the item search experience for users, Mercari assigns various kinds of metadata to the items on the app. For example, metadata for apparel includes color and the fabric&#8217;s texture, while for home appliances it contains model numbers. The different kinds of data that indicate an item&#8217;s traits is what we call metadata. 
I&#8217;d like to take a look right now at the metadata we assign to books and manga (what I&#8217;ll refer to as “title data” below). 
In the Mercari app, metadata is assigned to books and manga to show what series each item belongs to. For example, &quot;Kingdom Vol. 1,&quot; &quot;Kingdom Vol. 2,&quot; &quot;Kingdom Vol. 25-50,&quot; and &quot;Complete set of Kingdom&quot; get the same &quot;Kingdom&quot; title data. 
At Mercari, we also record all logs of items viewed by users. 
So, I tested a way I thought we could improve book and manga recommendations to match user interests by reconciling those logs and title data while using the Item2Vec recommendation technique.
<h2>Using Item2Vec</h2>
Word2Vec1 is one well-known technique for natural language processing.It is a method that proposes words can be processed in a variety of ways by vectorizing them. Although it&#8217;s a fairly old technology, it is highly versatile, so it&#8217;s used in many kinds of research and in machine learning competitions. 
Item2Vec2 is a way to apply that technique to item recommendations. 
Implementation and training are relatively simple, and it&#8217;s a very easy way to produce accuracy. That&#8217;s why other companies have also shared use cases. It&#8217;s difficult to use if there aren&#8217;t many user logs, but personally, I think it&#8217;s highly suited to a service like Mercari that has a massive number of users and an enormous number of logs collected on a daily basis. 
I created a recommendation logic with the following method.
<ol>
<li>
Derive the similarity and relationship of each item. (With Python, process monthly by using data from N past months.) 
a. Extract users who viewed book or manga items over the past N months. 
b. Collect users&#8217; item browsing histories over the past N months. 
c. Extract arrays of titles viewed by session. 
d. Regard title arrays as sentences and each title as a word, then derive title vectors with Item2Vec. 
e. Use title vectors to derive relationships between titles (cos similarity). 
f. Store data on relationship between titles in BigQuery.
</li>
<li>
Generate recommendations for users. (With BigQuery, process hourly by using data from N past hours.) 
a. Extract the history of items browsed by the user over the past N hours that are also tied to title data. 
b. Derive &quot;titles similar to item titles the user has viewed.&quot; 
c. Store the data derived in the preceding step as &quot;list of titles to recommend to users&quot; in Bigtable. 
d. If there is a title list to recommend in Bigtable when a user visits the home page, search for those titles and display them in the recommended item list.
</li>
</ol>
<img src="https://storage.googleapis.com/prd-engineering-asset/2022/09/992a7635-image2.png" alt="" /><center> 
Fig1. Concept diagram for separating titles by sessions in item browsing logs
<img src="https://storage.googleapis.com/prd-engineering-asset/2022/09/3d14bef2-img_1-473x1024.png" alt="" /> 
Fig2. The new recommendation section&quot;Recommended Books &amp; Manga for You&quot; 
</center>
I won&#8217;t discuss details about the system when it generates recommendation data or displays it in the app, but below is an overview (everything is processed on GCP)
<ul>
<li>Process most data concerning item viewing logs with BigQuery (BigQuery makes it easy to process hundreds of millions of lines of log data.)</li>
<li>Use Gensim on Colab to process Item2Vec calculations.</li>
<li>For the data pipeline, use Cloud Compower (Airflow) to execute hourly and monthly data processing.</li>
<li>Use Bigtable for the data store in the production environment.</li>
</ul>
<h2>Item2Vec Results</h2>
I&#8217;d like to show you some of the results from calculating with Item2Vec. 
Below are the results from the most_similar function in the model I calculated with <a href="https://radimrehurek.com/gensim/">Gensim</a>. (The number indicates the degree of similarity.)
Titles similar to the manga &quot;Dragon Zakura&quot;
<pre><code>1[(&#039;Dragon Zakura 2 17&#039;, 0.8193),
2 (&#039;Angel Bank: Dragon Zakura Gaiden&#039;, 0.6674),
3 (&#039;Let&#039;s Go to University of Tokyo: Dragon Zakura Official Guidebook&#039;, 0.628),
4 (&#039;7 Skills from Dragon Zakura for Parents to Bring Out the Ability for Kids to Pass the University of Tokyo Entrance Exam: Longshan High School “Parent Skills” Special...&#039;, 0.5366),
5 (&#039;Dragon Zakura Official Supplementary Reader, Age 16 Textbook 2, Where Do Study and Work Connect?&#039;, 0.5265),
6 (Gin no Anchor: Naitei Ukeoi Manga&#039;, 0.5145),
7 (&#039;The Great War of Archimedes&#039;, 0.5118),
8 (&#039;Bucho Shima Kosaku: Bilingual Edition&#039;, 0.5098),
9 (&#039;Investor Z Complete&#039;, 0.5034),
10 (&#039; Dragon Zakura Official Supplementary Reader, Age 16 Textbook, Why Study? What to Study?&#039;, 0.4913),
11...</code></pre>
Dragon Zakura is a Japanese manga series written and illustrated by Norifusa Mita. 
It tells the story of high school students studying to enter the University of Tokyo. 
Supplementary readers for entrance exams are mixed in with Norifusa Mita&#8217;s works. I found this unexpected result very interesting.
Titles similar to the manga &quot;Shoujiki Fudousan&quot;
<pre><code>1[(&#039;24 Pieces of Divine Knowledge for Beating Real Estate Companies&#039;, 0.5547),
2 (&#039;Stand Out at All Stages, From Novice to Veteran: The Most Powerful Real Estate Investing Textbook for Investors 1...&#039;, 0.493),
3 (&#039;Kurosagi&#039;, 0.4731),
4 (&#039;The Easiest Real Estate Investing Textbook in the World, 1st-Year Students: Perfect for Beginning Again&#039;, 0.4692),
5 (&#039;Rumored Secret Drugs in Industry: Big Real Estate Techniques&#039;, 0.4675),
6 (&#039;Souzoku Tantei&#039;, 0.4628),
7 (&#039;The Future of Real Estate: Prepare for the Big Housing Turnover&#039;, 0.4405),
8 (&#039;Worst Properties of the Year&#039;, 0.4351),
9 (&#039;Real Estate Investing Don&#039;ts&#039;, 0.4313),
10 (&#039;Monkey Peak&#039;, 0.4291),
11...</code></pre>
This is the story of a real estate agent who can&#8217;t tell a lie. The list includes &quot;Kurosagi,&quot; which Takeshi Natsuhara (the original writer of &quot;Shoujiki Fudousan&quot;) worked on. There is indeed a strong connection because the original writer is the same. It&#8217;s also interesting that the list includes regular real estate books among the manga. 
By the way, I only just now learned about the manga &quot;Souzoku Tantei&quot; from writing this tech blog, so I took a break from writing, bought the digital edition, and gave it a read. It&#8217;s pretty good. It certainly feels similar to &quot;Shoujiki Fudousan.&quot;
Titles similar to the novel &quot;Clouds Above the Hill&quot; by Ryotaro Shiba
<pre><code class="language-1[(&#039;Tobu">2 (&#039;Yo ni Sumu Hibi Complete&#039;, 0.7067),
3 (&#039;Natsukusa no Fu Vol. 1 &amp; 2&#039;, 0.6681),
4 (&#039;Owl&#039;s Castle&#039;, 0.6624),
5 (&#039;When the Last Sword Is Drawn Vol. 1 &amp; 2&#039;, 0.6594),
6 (&#039;Kashin Vol. 1-3&#039;, 0.653),
7 (&#039;Sekigahara Vol. 1 &amp; 2&#039;, 0.6471),
8 (&#039;Oda Nobunaga&#039;, 0.647),
9 (&#039;Komyo ga Tsuji&#039;, 0.6468),
10 (&#039;The Last Shogun: The Life of Tokugawa Yoshinobu&#039;, 0.6324),
11...</code></pre>
There are lots of works by Ryotaro Shiba.
Titles similar to the novel &quot;Downtown Rocket&quot; by Jun Ikeido
<pre><code>1[(&#039;Downtown Rocket: Ghost Yata Glass&#039;, 0.9246),
2 (&#039;Downtown Rocket: Yata Glass Ghost&#039;, 0.9115),
3 (&#039;Stock Price Crash&#039;, 0.8829),
4 (&#039;Bank Manager Special Assignment, New Edition&#039;, 0.8819),
5 (&#039;Hanasaki Mai Speaks Out&#039;, 0.8758),
6 (&#039;Bitter Enemy&#039;, 0.8756),
7 (&#039;Scandal, New Edition&#039;, 0.8702),
8 (&#039;Bones of Steel&#039;, 0.8686),
9 (&#039;Financial Detective&#039;, 0.8614),
10 (&#039;Ginko Shiokinin&#039;, 0.8589),
11...</code></pre>
This list also has many works by the author, in this case Jun Ikeido. Mercari users who search for novels might view many different works by the same writer.
<h2>Recommendation Logic A/B Testing</h2>
After completing Item2Vec, I ran A/B testing (also cleverly named: Book2Vec). 
At Mercari, it&#8217;s a big part of our culture to quantitatively assess improvements with A/B testing. Team member yaginuuun released materials about our attempts to standardize A/B testing, so check those out. (<a href="https://speakerdeck.com/shyaginuma/btesutobiao-zhun-hua-hefalsequ-rizu-mi">Slides</a> * This slide is in Japanese only) 
In conclusion, it produced good results (though I can&#8217;t share the exact numbers) and the recommendation logic I made was adopted. I&#8217;m diligently developing it so pages in addition to the home page can use it too. 
The joy I received from having the logic I created and implemented run on a service used by tens of millions of people is the best part about working as a Mercari engineer.
<h2>We’re Hiring!</h2>
In addition to metadata for books and manga, we intend to improve the recommendation logic we use for the massive amount of other data at Mercari. 
As regular users of the Mercari app may be aware, the Mercari recommendation logic still has lots we can do with it and plenty of room for improvement. 
The fact is, because of the massive number of visitors our site gets, a complex logic poses challenges because we can&#8217;t use it in the actual environment. On the other hand, however, if we have just one that works, then it will have extensive economic value. I want to keep on working out ways for users to find the items that are right for them. 
If you&#8217;re interested in this work, I&#8217;d be happy to hear from you, as would my Mercari colleagues, so feel free to contact us. See ya!
<a href="https://apply.workable.com/mercari/j/3333DC5597/">Software Engineer, Machine Learning &amp; Search &#8211; Mercari</a>
<h2>References</h2>
[1] Mikolov T, Sutskever I, Chen K, Corrado GS, Dean J. Distributed representations of words and phrases and their compositionality. In Proceedings of NIPS 2013 (pp. 3111-3119). 
[2] Oren Barkan and Noam Koenigstein. 2016. Item2vec: Neural item embedding for collaborative filtering. arXiv preprint arXiv:1603.04259.

Attempt to improve item recommendation accuracy using Item2vec

The MERCARI GEARS Lecture Series is a video series for engineers that features the future direction and technical challenges of each Mercari Group company, including Mercari Inc.
In this edition of MERCARI GEARS, Mercari's Edge AI Team highlights the Mercari app’s real-time item scanning feature, available on iOS and Web.
<center> 
<iframe loading="lazy" width="560" height="315" src="https://www.youtube.com/embed/cKBzwB5StLY" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> 
Edge AI - Realtime Item Scan for iOS and Web [AI@Mercari] 
</center>
Mercari has a feature called &quot;AI Listing&quot; that suggests category, brand, and even selling price for an item when you upload a photo of it.
To use the &quot;AI Listing&quot; feature, you used to have to take and upload a photo of an item. But now, the Edge AI Team has made it possible to to use the feature as soon as your device’s camera is activated and the item is in focus.
In order to create a quicker and more comfortable display experience, the team leveraged recent advances in smartphone technology to create an ML model with low computational cost. As a result, suggestions through this feature can now be made directly on the smartphone without having to go back and forth to a computer server.
We are working on a variety of things, including improving processing time and device compatibility to adapt quantization to the learning model and make inferences, developing tools to monitor performance after implementation, which are explained in detail in the video. Please take a look!

Edge AI  – Realtime Item-Scan for iOS and Web

“<a href="https://medium.com/mercari-engineering" title="Making Mercari">Making Mercari</a>&quot; is an engineering blog for Mercari's US product. AI-related posts on &quot;Making Mercari&quot; will also be featured on this website.
<hr />
Here is a post by Abhishek Munagekar, Software Engineer for the Machine Learning Platform at Mercari US. In this blog post, he describes how his team optimized the cost of computing resources used for machine learning training. More specifically, there were some challenges with downscaling of the Kubernetes cluster used for ML model training. Read the blog post to find out how PodDisruptionBudget, dedicated node pools and the Gatekeeper Assign CRD were used to overcome these challenges.
<a href="https://medium.com/mercari-engineering/how-we-optimized-our-machine-learning-training-infrastructure-costs-c2a64f7451bb" title="**How we optimized our Machine Learning Training Infrastructure Costs | by Abhishek Munagekar | Making Mercari | Apr, 2022 | Medium**">How we optimized our Machine Learning Training Infrastructure Costs | by Abhishek Munagekar | Making Mercari | Apr, 2022 | Medium</a>

How we optimized our Machine Learning Training Infrastructure Costs @ Mercari US

Blogs by mercari AI team, delivering information about AI technology and organization

The Past and Future of Mercari’s Discovery Experience, from the PM who Launched the Recommendation Team

Discoveries through a career as both PM and engineer at Mercari

Mercari AI’s research paper —”Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions”—accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Estimating the Effect of Timing on Coupon Effectiveness”— accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Textual Content Moderation in C2C Marketplace”—accepted to ACL2022 ECNLP Workshop

The Past and Future of Mercari’s Discovery Experience, from the PM who Launched the Recommendation Team

Discoveries through a career as both PM and engineer at Mercari

Mercari AI’s research paper —”Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions”—accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Estimating the Effect of Timing on Coupon Effectiveness”— accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Textual Content Moderation in C2C Marketplace”—accepted to ACL2022 ECNLP Workshop

The Past and Future of Mercari’s Discovery Experience, from the PM who Launched the Recommendation Team

Discoveries through a career as both PM and engineer at Mercari

Mercari AI’s research paper —”Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions”—accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Estimating the Effect of Timing on Coupon Effectiveness”— accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Textual Content Moderation in C2C Marketplace”—accepted to ACL2022 ECNLP Workshop

Latest Posts

The Past and Future of Mercari’s Discovery Experience, from the PM who Launched the Recommendation Team

Discoveries through a career as both PM and engineer at Mercari

Mercari AI’s research paper —”Personalized Promotion Decision Making Based on Direct and Enduring Effect Predictions”—accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Estimating the Effect of Timing on Coupon Effectiveness”— accepted to KDD2022 1st Workshop

Mercari AI’s research paper—”Textual Content Moderation in C2C Marketplace”—accepted to ACL2022 ECNLP Workshop

Attempt to improve item recommendation accuracy using Item2vec

Edge AI – Realtime Item-Scan for iOS and Web

How we optimized our Machine Learning Training Infrastructure Costs @ Mercari US