Palette-AI-models (AI & Data)

Description

Palette is a date course recommendation platform for couples.
This repository covers AI technologies used in Palette.

Member

Moonsoo Park	Taeho Kim	Seulbeen Je
@m0onsoo	@Taeho0818	@powerjsv

Data

• All our data was collected from Naver Place, a map and restaurant information platform.
• To this end, we developed a web crawler and collected about 350,000 reviews to be used as interactions.
• The data consists of restaurant, cafe, and bar business names, categories, locations, reviews, dishes, and dish images.

Region	#User	#Restaurant	#Interaction	Density(%)
Gwangjin	11,935	2,108	239,170	0.95
Hongdae	4,158	538	47,179	2.1
Jamsil	4,724	525	59,633	2.4
Total	32,750	3,171	345,982

Recommendation

• LightGCN was used to provide recommendation results to users.
• The graph structure was adopted because it can capture overall tastes well in user records.
• Our main function, couple recommendation, inferred the recommendation results by summing each user's embeddings.
• Due to the lack of a quantitative evaluation method for both users, we conducted a survey to more than 100 random people in their 20s and received positive responses from more than 80%.

Dataset	Yelp2018	Amazon-Book	Ours
NDCG	0.0525	0.0318	0.0718

Best Dish Extraction

• Extract reliable representative dishes from reviews
• To this end, we use KeyBERT, a keyword extraction algorithm.
• Noting that reviews reflect users' subjective feelings, we added a Sentiment Penalty to the KeyBERT score to increase the reliability of the results.

Review Analysis

• We used a fine-tuned T5 to summarize a large amount of restaurant reviews.
• Provides a priority sorting function for reviews of users with similar tastes as mine by calculating the similarity between user embeddings.
• This helps users make decisions.

Deployment

• Implement model serving using FastAPI
• Model Serving Repo

Demo

You can see more details in the video below.

Presentation

Awards

🎉 1st palce in Creatvie Design Competition (Capstone Design Project)!!

References

• Naver Place
• LightGCN
• KeyBERT
• Fine-tuned T5