The Tokopedia Way: Getting Relevant Products Through a Deep Recommender System
Do you enjoy Tokopedia’s personalized experience every time you visit our site? How often do you click, or even buy products from the “Terlaris Untukmu” section? Countless applications use recommendation engines, just like Tokopedia, which aim to improve our customer experience and also to increase our revenue. The coolest thing about recommendations is that they’re awesome both for business and users. But why? Why is a recommender system beneficial for us?
For 11 years, technology has always been the heart of every Tokopedia’s innovation to keep developing for the better. Along with the development, Tokopedia is now transforming into “Data & Artificial Intelligence (AI) first company”. It means, how can Tokopedia adopt AI technology to achieve customer excellence. One of the Tokopedia’s features and innovation that has utilized AI technology is the fast recommendation for more than 350 million products which are corresponding to every Tokopedia’s user interest.
Tokopedia Technology Workshop
As a form of contribution in developing digital talent and supporting tech enthusiasts in Indonesia with the latest technology in Tokopedia, Tokopedia Academy with Tokopedia Data Scientist held a Tokopedia Technology Workshop with the topic: “The Tokopedia Way: Getting Relevant Products Through a Deep Recommender System.” With this topic, Tokopedia invited the participants to get to know more about AI, especially a Recommender System that could help millions of buyers and sellers fulfilling their needs through Tokopedia.
During the workshop, the participants not only learned how the recommendation system could help Tokopedia improve user personalized experience, but also tried to build recommender systems from scratch, starting from the data exploratory, features engineering, modeling processes, and final implementation by providing recommended products to the users.
Introduction to Recommender System
In the first session, the introduction explained a recommender system, various approaches, and also the benefits in general.
A recommender system is defined to help people make choices when there is no base knowledge or experience in choosing. In e-commerce, a recommender system is used to predict which items users might like. To complete this task, there are two approaches, namely collaborative filtering, and content-based, even though modern recommender systems combine these two approaches. The first one is collaborative filtering, which provides predictions based on the similarity to other users. The next one is content-based, which focuses on the attributes or descriptive characteristics of items to generate product recommendations.
For Tokopedia, a Recommender System can help achieve one of the values that Tokopedia has, which is “Focus on Consumer”. The technology built must have an impact on Tokopedia’s users.
For buyers, the existence of a recommender system helps to show the items that are being needed, thus shortening the time to get the appropriate items. Personalized experience increases customers’ satisfaction which can be translated into the increased loyalty, increasing consumption, and more profit. Also, from the seller’s point of view, a Recommender System will help marketing the items to the correct target market in a short time. This can increase sellers’ satisfaction by the increase of selling items’ exposure level, increasing sales, and of course loyalty.
Exploratory Data & Feature Engineering
In this session, participants recognize each of the tables in the sample data. Also, participants were asked to answer questions from stakeholders to find insights such as how much income and also the average rating for each category of items. Not only answering these questions, but participants must also be able to make business decisions that could be done based on the insights that have been found.
In the Feature Engineering session, participants were preparing the data required for the model. Not only practice, but they also got to know about Feature Engineering techniques that were commonly used such as one-hot encoding.
The input required for the model is interaction data from the user-item to get a review value (Figure 3), a feature that describes a user, and a feature that describes a product.
Modeling & Implementation
The next session is building a recommender system model. To help the model learn which products to recommend to our user, we will build a model to predict the rating. Rating is used, because by using rating not only get the kind of product the user interacts with, but also the feedback of the products. The inputs of this model are user and product data, and the output of the model is a rating.
At the end of the modeling session, there is the last most important step, which is the model implementation process to get the recommended product to the users. To produce this product, the steps taken are to cache all products, get the user rating prediction, sort the products prediction by the rating prediction, and the last is get top-k products for each user.
After modeling was done, the next step is to monitor and maintain the performance of the model. One of the most important things is to retrain the model, to keep the model performing well on real data. Furthermore, there are several things we can do to make it better, such as extract more features, improve the model, add diversity, and add more data.
As Tokopedia DNA, “Make it Happen, Make it Better”, we still need to go further than just all these steps and redo the things to give the best implementation. Although we have come a long way to reproduce a robust recommender system, as long as we support our customer needs and achieve customer excellence, it will not be a problem because we are all doing this with the spirit of our DNA value.
Special thanks to the Data Science team at Tokopedia for helping me with this post.
