Community Blog PPWang's Exploration of the Search Business in B2B E-Commerce

PPWang's Exploration of the Search Business in B2B E-Commerce

This article reviews the "Best Practice Sharing of the Search Business in B2B E-Commerce" held during the Apsara Conference 2020.

Catch the replay of the Apsara Conference 2020 at this link!

By ELK Geek with special guest Cong Xixing, CEO of PPWang

Introduction to PPWang:

PPWang is an e-commerce platform on mobile devices for clothing wholesale based on traditional wholesale markets. It is the first e-commerce platform that combines livestreaming and B2B clothing. From 2015 to today, PPWang has become a livestreaming spokesperson in the wholesale clothing industry.

Mission: Promote industry development with technology and make wholesale simple

Vision: Become an excellent online wholesale platform in China

Values: Integrity, passion, introspection, teamwork, and innovation


The Tortuous Process of Search R&D

By introducing engineers on search algorithms and setting up a special team for search projects, PPWang utilized an open-source search framework and performed redevelopment. However, PPWang failed to meet the requirements of search optimization. The problems that PPWang encountered are listed below:

  1. The Effect Is Not Ideal: PPWang wants to improve the search experience by writing simple algorithms, but the effect is not obvious.
  2. The Lack of Personnel: It is very difficult to find professional and appropriate algorithm personnel.
  3. The Lack of Data Timeliness: It is difficult to balance the relation between high-quality commodities and newly released commodities.
  4. Malicious Ranking Competition Among Merchants: Some merchants find ranking loopholes and get top positions on the search results through piling up search keywords, which makes the user experience bad.

The Turn of Search R&D Process – Alibaba Cloud's OpenSearch Solution

1. Sorting Optimization:

Users are allowed to use expressions to adjust the sorting results in real-time without relying on development engineers.

  • The first-phase is rough sorting, which selects relevant documents from the specified documents collection.
  • The second-phase is fine-grained sorting, which further filters the results of rough sorting and supports any complex expressions and syntax.
  • Correlation Scoring Strategy: Score the recalled rank_size documents (1 million at present) according to the definition of the rough sorting expression. Score and sort N results (at the hundred level) with the highest rough sorting scores, based on the fine-grained sorting expression. Then, return the corresponding results to the user according to Start and Hit settings. If the number of obtained results exceeds the number of documents after fine-grained sorting, the obtained results will be those documents after rough sorting.


2. Distinct - The Function for Aggregation and Breaking Apart

It can balance the display opportunities of "high-quality commodities" and "newly released commodities".

  • Distinct helps present diverse results and improve user experience. For example, users may obtain a large number of documents in a single query. However, if several documents of the same provider all have relatively high scores, they will rank on the top of the list. This explains the reason why almost all results displayed on the first page belong to the same provider, which is neither conducive to the result display nor user experience. In this regard, distinct can be used to extract documents from each provider. Thus, all providers are entitled to present their documents.

3. The Semantic Understanding of Query:

It is very convenient to choose and use preset search features.


The Bright Future of the Search R&D Process

There are more advanced features of OpenSearch that will be available to PPWang in the future.

  • Popularity Model: The Popularity Model calculates the values of static quality and popularity of each commodity. Those values will be constantly trained and calculated to produce popularity scores. Moreover, the model creates a more fine-grained sorting model to accurately meet search demands.
  • Category Prediction Model: It is used for predicting the category of results that users tend to search for based on search words. Combined with sorting expressions, this model can present more suitable results for users.
  • A/B Testing: A/B testing can help test businesses before carrying out and allocating certain proportions of traffic. By doing so, users can avoid the negative impact of online businesses.
0 0 0
Share on

Alibaba Clouder

2,600 posts | 753 followers

You may also like