How MaxCompute Helped YCLOSET Build Smart O&M Tools

This article is based on the lecture on how to build smart O&M tools based on MaxCompute given by Cheng Yiding, the CTO of YCLOSET.

How has MaxCompute helped YCLOSET, the largest fashion sharing platform in Asia, solve their problems such as slow data extraction and different data standards? Cheng Yiding described the data architecture of YCLOSET and analyzed how MaxCompute helps YCLOSET build smart O&M tools from three aspects: user operations application, product operations application and algorithm recommendation system.

The following content includes highlights from the lecture video.


YCLOSET is the largest fashion sharing platform in Asia. YCLOSET provides monthly women’s clothing rental services. YCLOSET members can select their favorite clothes on the app and continuously change their choices to wear different clothes within a month at a fixed monthly fee. Clothes selected by YCLOSET members will be mailed to them. When a member no longer wants to wear their currently rented clothing from the platform, she can return the clothing at a scheduled time and location. The member can always place new orders to continuously wear different clothes. We have a wide range of customers, from students who are still studying at university to white-collar females who have worked for years. Customers can find the style of clothing that suits them best.

Today, e-commerce has matured significantly, and offline shopping is also a preferred lifestyle for many people. So, why do many people still rent clothing? Because YCLOSET can solve the following customer pain points:

  1. After browsing through her wardrobe, a customer may end up finding nothing she would like to wear. YCLOSET allows our customers to flexibly change what they want to wear and try different styles of clothing.

Why Did We Choose MaxCompute?

The front-end and back-end operations and maintenance of YCLOSET are very complex. On the front-end, we need our customers to know about and understand us quickly to begin customer conversion. YCLOSET has obvious differences from common e-commerce companies in the back-end operations and maintenance. At normal e-commerce companies, after customers place orders on the front-end, the order management system forwards these orders to the warehouse center where item selection, order preparation, and packaging are performed and the logistics process begins before purchased items are delivered to the customers. For common e-commerce companies, the entire business process basically ends at this point. However, for YCLOSET, the entire process is only half done at this point. We also need to accept returns and clean rented clothing and perform quality inspection on them before they are made available to customers again. The big data support is indispensable in this complicated O&M process, and MaxCompute, standing at the heart of the data architecture, helps YCLOSET a lot.

What problems had we encountered before we adopted MaxCompute? The main problems are as follows:

  • Slow data extraction. SQL ran slower and slower with increasing system complexity.

These problems led us to seek out better solutions. At the beginning, we had intended to use Hadoop, but our evaluation found that Hadoop would bring burdensome O&M costs and resource consumption to our data team and O&M team. In addition, Hadoop would not bring immediate positive effects. Fortunately, we later came across MaxCompute. Starting with a trial, we gradually built our data warehouse and data architecture on MaxCompute.

What advantages does MaxCompute have? MaxCompute mainly has the following advantages:

  1. MaxCompute is a cloud-based big data warehouse that does not require complex O&M.

How to Implement MaxCompute?

The following diagram shows the data architecture of YCLOSET. The bottom-layer is the data collection layer, including RDS production databases and log services. Data from data sources are transmitted to the data compute layer through services such as Log Service and DataHub. In the data compute layer, the core component MaxCompute is used together with our own scripts and UDFs for data storage and computing, with generated results sent to the front-end data application layer. The data application layer consists of data analysis and presentation tools such as YConsole, QuickBI, and DataV. We use DataV as the dashboard that displays the global business data and allows us to know who has placed an order nationwide and the stock information at the warehouse centers.

We also need to support additional features such as customer persona, product persona, recognition of popular items, YCLOSET index, and real-time monitoring.

User Operations Application

For Internet companies, one important aspect is user operations, which features two business metrics: conversion and retention. Conversion refers to the act of informing our customers of our value and converting them to paying customers. Retention refers to our ability to provide our existing customers with a better experience on the platform, integrate our service into their routine and eventually retain these customers.

We combine MySQL data and user logs and put them into MaxCompute. On the PAI platform, we perform data analysis ourselves, including data dimension analysis by using data mining algorithms and random forests, and set up many quantitative metrics targeting our user operations. We monitor our operations metrics through emails and daily reports. We also have developed our own user grouping system based on these metrics to facilitate refined user operations. Based on the Alibaba Cloud machine learning platform PAI, we have built a predicative model to predict our customer conversion rate. The customer churn rate compute and alerting system allows us to implement targeted retention marketing policies and activities based on the customer churn index and reduce the membership defection by more than 50%.

Product Operations Application

Each YCLOSET item has structured data. Our professional buyers add labels in 20+ dimensions to our items. We also summarize our user behaviors. The interactions between customers and items are put into MaxCompute to perform correlation analysis and form a series of item metrics. For example, with the recognition of popular items, we can predict what kind of clothing attributes and dimensions are very likely to make items popular. Our buyers will use professional data tools when they purchase clothes, significantly improving the operations efficiency.

Based on comprehensive performance dimensions such as stock depth and rental information, we also calculate YCLOSET indexes of our items, which can be used to rank our items. In addition to rental incomes to our partners, we also provide data tools to maximize the value of the business model.

A core part about items is labels. If labels are made fine-grained enough, we can get a thorough understanding of items and implement some predicative metrics. We can provide guidance on stocking up items based on different labels and the item popularity indexes in different style scenarios.The association between the product allocation and stock is based on the final item vacancy rate. Because the YCLOSET index reflects the item popularity, we can facilitate the rental and selling business according to user behaviors and implement a closed item flow loop.

Algorithm Recommendation System

How do we help our customers to quickly find clothes that look perfect on them for one month? This requires the support from recommendation algorithms.

YCLOSET’s recommendation algorithm is also based on MaxCompute. It collects users’ item-related behaviors into MaxCompute through logs, creates customer personas, and performs model training based on customer personas before the final item recommendations are displayed to customers. The use of the recommendation system has significantly improved our business. The click rate on the clothes selection page has increased by 70% and the average clicks per user have increased by 50. The click rate for a recommended item has increased by 150%, and the per-user click rate has increased by 110%.

To sum up, in the era of big data, MaxCompute provides a very low threshold for startups to utilize the power of big data and make big data play an important role and benefit various companies. This is all I want to share today. Thank you all for your attention.

Original Source