Naver launches program to support SME tech capabilities

[Courtesy of Naver Corp.]

Naver Corp. is launching a program to help small and medium-sized enterprises (SMEs) improve their technological capabilities, the company announced Wednesday.

The “Store Growth” program aims to help SMEs seeking growth strategies after starting their business by providing consulting services and growth points to apply various commerce solutions to their Naver Smart Stores.

Naver plans to pilot the Store Growth program for four weeks, targeting up to 100 Smart Store sellers in the home, kids, leisure and pet categories who have never used the various technological tools and commerce solutions on the commerce platform. According to Naver, sellers who use marketing and technology solutions more actively see a more-than twofold jump in their cumulative annual transaction volume.

“This program aims to lower the entry barriers to technological solutions so that SME business owners can use them according to their business characteristics,” the company said.

Naver and a research team led by Professor Park Chan-young at the Korea Advanced Institute of Science and Technology (KAIST) have also jointly developed technology to significantly improve the performance of large language model (LLM) product recommendation via “collaborative filtering,” KAIST said on Wednesday.

LLM is the foundational technology for generative artificial intelligence (AI), and product recommendation services using LLM have gained attention on e-commerce platforms. But these services have not been successful because they require the “fine-tuning” of data to train LLM, which involves considerable time and cost for learning and inference.

Instead of training LLM, the research team used “collaborative filtering,” which uses information from other users who have consumed similar products in the neural network and allows for the lightweight use of neural networks.

With collaborative filtering, it increased learning and the inference speed by 253 percent and 171 percent respectively compared to existing research, while also achieving an average 12 percent improvement in product recommendation performance.

The team specifically achieved a 42 percent improvement in multi-domain product recommendations, where models trained on shopping malls of different product categories make recommendations on the current shopping mall. It also saw an average 20 percent improvement in product recommendations for users with limited purchase history.

“This is a new method that extracts user-product interaction information from traditional collaborative filtering models. This could be used to develop conversational recommendation systems or personalized product information generation,” Park said.