EpiK Protocol to make intelligent product recommendations know you better
This article is from Dr. Hongyuan Yu, a domain expert in EpiK Protocol, who is currently working at the National Laboratory of pattern recognition. His main research areas are machine learning, multimodal representation learning and computer vision. Dr. Yu has won the championship of RGBD in VOT (Visual Object Tracking challenge), the runner-up of ROBCUP JAPAN OPEN in Robot World Cup, and the best paper award in ICDAR. He has published many papers in international conferences such as NeruIPS, ECCV, and top journals such as IJCV, PR, TCSVT. Dr. Yu has also carried out academic research and collaboration with Baidu, Microsoft Research Asia and Alibaba Ant Group.
What is intelligent product recommendation?
At present, most of the e-commerce platforms’ product recommendations are based on user input, recommending the popular goods or promotional goods. They also recommend the latest products to boost the sales of the products. But often these do not meet our requirements, leaving a poor shopping experience. This is all because such a recommendation system does not know us well enough.
A recommendation system that better understands you
Recommendation systems need to be trained with a large amount of real user data, while the authenticity and diversity of the data, as well as data construction, can significantly affect the systems. Most e-commerce platforms are trained by their own user shopping data, but can these data really be a true response to user preferences? In many cases, certain products are purchased due to impulse consumption, so the authenticity of these data remains questionable. In order to recommend products that know you and fit you better, Multi-Modal domain expert Hongyuan Yu has collected plenty of real multimodal data through Epik Protocol’s Knowledge Mainland. How does this multimodal data drive the development of intelligent recommendations? Let’s see what Dr. Yu’s team has been working on.
Data Construction
According to Maslow’s Hierarchy of Needs, physiological needs are the most basic and important needs on which all other needs are based, and consumption at this level dominates major e-commerce platforms and supermarkets. Therefore, Dr. Yu’s team decided to build a multimodal dataset from such commodities and designed a detailed data format for these goods. Through the manual labeling on the Knowledge Mainland, real and diverse data were obtained. This is the biggest difference from the regular e-commerce platform dataset.
Data Validation
The accuracy of data is the top priority of AI technology, just imagine how unqualified data can facilitate qualified AI technology. Therefore, in order to ensure accuracy, after labeling the data on the Knowledge Mainland, Dr. Yu’s team will also maintain the accuracy of the data by eliminating or revising the wrong data.
Model Building
Algorithms and models are the two important factors of artificial intelligence technology. To implement AI technology, only high-quality labeled data is not enough, but it also needs efficient algorithms to bring the data into full play. Dr. Yu’s team has built a product intelligence recommendation model based on the world’s most advanced transformer technology and large model pre-training technology. The model input includes both image and graph modal data, and the training objectives are image classification, image segmentation, graph classification, graph prediction, etc. Making full use of data from both modalities, the technology enables separate coding of inputs and selects the most relevant products automatically. When users search for the products they want to buy, the system displays the corresponding products and promotes the products that are more suitable for them, thus ensuring an easier and more enjoyable shopping experience.
At present, Dr. Yu’s team is actively cooperating with many AI teams in China. We believe that in the near future, the intelligent recommendation system incubated through Epik Protocol will excel in the field.
