[Paper Review] Recipe representation learning with networks

Title: Recipe representation learning with networks

Authors: Yijun Tian, Chuxu Zhang 

---

Summary

• Recipe representation learning with networks to involve both the textual feature and the structural, relational feature into recipe representations
• Present RecipeNet a large-scale corpus of recipe data
• Propose rn2vec a novel heterogeneous recipe network embedding model
• Combine the objective function of node classification and link prediction

 

Framework

1. Construction of RecipeNet 
2. Encoding instructions through textual CNN and ingredients 
3. Inner ingredients transformer 
4. GNN with hierarchical attention 
5. Combine objective function 
6. Node classification 

• Why Graph-based model? 
  • To capture structural information of RecipeNet
  • Node-level attention module to distinguish the subtle difference of neighbor nodes under a specific relation

 

RecipeNet

: textual feature and the structural, relational feature into recipe representations

• Two types of nodes
  • ingredient
  • recipes (with instruction) 
• Three types of edges
  • recipe-recipe
  • recipe-ingredient
  • ingredient-ingredient

 

 

• 데이터 수집(Recipe1M)
  • Recipe, ingredient, nutritional information
• Transform recipe dataset into RecipeNet
  • Recipe nodes, ingredient nodes, edges
• Network Construction 
  • Recipe node a pre-trained skip-instruction embeddings
  • Ingredient node a nutrient of each ingredient into a vector 
  • Recipe-ingredient edge a connect each recipe and its ingredient, edge weight(weight of each ingredient)
  • Recipe-Recipe edge a similar recipe mining from FoodKG, edge weight(similarity score)
  • Ingredient-ingredient edge a food pairing graph(FlaborGraph)

 

RN2VEC 

• Encoding Instruction and Ingredients
  • Instruction à encoded by textual CNN, Ingredient à use nutrient vectors
  • Type-specific transformation matrix to project the input features of nodes with different types into the same embedding space
• Inner-ingredients Transformer
  • Interaction between ingredients needs to be incorporated
• GNN with Hierarchical Attention
  • Interaction among nodes and learn node embeddings
  • Node-level attention to calculate each relation-specific embedding
  • Relation-level attentions to combining relation-specific embeddings
• Objective Function
  • Combined objective function of node classification and link prediction

 

Model Validation 

using recipe category 

 

---

Paper URL (https://dl.acm.org/doi/abs/10.1145/3459637.3482468)

Recipe Representation Learning with Networks | Proceedings of the 30th ACM International Conference on Information & Knowledge M