Epik Protocol aims to make product recommendations more personalized

What are product recommendations?
Currently, most e-commerce platforms recommend popular items, discounted products, or newly released items based on user input to drive sales. However, these recommendations often fail to meet our requirements, resulting in a poor shopping experience. This is because such recommendation systems need to understand you truly.
A recommendation system that understands you better.

Recommendation systems require a significant amount of actual user data for training. The data’s authenticity, diversity, and content greatly influence the effectiveness of intelligent recommendation systems. E-commerce platforms typically rely on user shopping data from their platforms for training, but is this data a true reflection of users’ preferences?

Data construction
Maslow’s hierarchy of needs theory states that physiological needs are the most basic and essential, upon which all other needs are built. Consumption at this level dominates the traffic on major e-commerce platforms and supermarkets. This is also the key difference between this dataset and regular e-commerce platform datasets.

Data verification
The accuracy of data is of utmost importance in artificial intelligence technology. Just imagine how unqualified data could catalyze competent AI technology. Therefore, to ensure data accuracy, after annotating the data through the Knowledge Continent, Dr. Yu Hongyuan’s team also conducts data quality control by eliminating or correcting erroneous data.

Model construction
Algorithms and models are the two pillars of artificial intelligence technology. High-quality annotated data alone is insufficient to achieve the practical application required. Efficient algorithms are also needed to leverage the power of the data fully. Dr. Yu Hongyuan’s team has developed an intelligent product recommendation model based on state-of-the-art Transformer technology and large-scale pre-training techniques. This model takes input in the form of both images and graphs, and its training goals include tasks such as image classification, image segmentation, graph classification, and graph prediction. The model effectively utilizes data from both modalities by joint training on these tasks. This technology enables the model to encode the input separately and automatically select the most relevant products. When users search for desired items, the system displays the corresponding products and recommends items that are more suitable for them, enhancing their shopping experience with ease and enjoyment.