Skip to content

Latest commit

 

History

History
122 lines (65 loc) · 5.37 KB

report.md

File metadata and controls

122 lines (65 loc) · 5.37 KB
Raw

Recipe Generation - Based on Eat-PIM

Team members

  • Sofya Aksenyuk, 150284
  • Uladzimir Ivashka, 150281

Main idea

The goal is to introduce a model that would be capable of generating new recipes based on a user's input ingredients.

Based on Eat-PIM and FlavorGraph generated embeddings, the input user's ingredients would result in a FlowGraph, presenting the recipe proposed by the model.

Project source code

Source code to the project can be found here: Source code

Chain of work

  • Parse provided ingredients
  • Match infringements with FoodOn and Wikidata entities
  • Choose most relevant ingredients (that make sense together out of list of ingredients)
  • Generate a set of the most applicable actions for these ingredients
  • Create a FlowGraph from these ingredients and actions

EDA

Source code to EDA can be found here: EDA

Algorithm

Step 1 - Parse provided ingredients

Source code to Data Parser can be found here: Data Parser

Main methods used in data parsing:

  • Parsing Matches (parse_matches): Read files that contain mappings between ingredients and other entities (like FoodOn and Wikidata), then reformat and save this data for later use

  • Parsing Flow Graphs (parse_flow_graphs): Process recipe flow graph data, which helps in understanding the sequence and hierarchy in recipes, and save the processed data

  • Parsing Entities Embeddings (parse_entities_embeddings): Handle embeddings related to ingredients and other entities by reading, normalizing, and computing cosine similarity matrices

  • Parsing Actions Embeddings (parse_actions_embeddings): Process action-related embeddings

  • Parsing Co-occurrence Data (parse_cooc): Analyze ingredient co-occurrence data to explore common pairings and potential novel combinations in recipes

Example of User Input

Step 2 - Match infringements with FoodOn and Wikidata entities

Source code to Ingredient Selection can be found here: Ingredient Selection

Main methods used in ingredient selection:

  • Identifying Existing Ingredients (ingredients_existing_in_cooc): Filter the input list of ingredients to retain only those present in the co-occurrence dataset. If an ingredient is not directly found, the class attempts to match its short form from the loaded mappings.

  • Calculating the Threshold (determine_threshold): Determine a threshold value for ingredient selection based on their similarity scores in the co-occurrence matrix. The percentile for this threshold is manually adjustable.

  • Selecting Ingredients:

    • Initial Selection Process (select): Select the most common ingredient based on the co-occurrence data. Add ingredients to meet a minimum count requirement, ensuring a basic selection diversity.
    • Expanding the Selection: Continue adding ingredients based on their similarity scores with already selected ingredients. This process continues until the average similarity score falls below the determined threshold.

Example of Selected Ingredients

Step 3 - Generate a set of the most applicable actions for these ingredients and create a FlowGraph

Cluster selected ingredients

Source code to Clustering can be found here: Clustering

  • Apply Elbow method to determine the number of clusters:

    • For Eat-PIM embeddings:

    Conclusions: No particular Number of clusters (k) can be determined

    • Then, for mix FlavorGraph and EaT-PIM embeddings:

    Conclusions: clearly defined k = 4 Number of clusters

Clusters k=4 visualization for selected ingredients

Determine actions (One cluster - one action):

  • Calculate cosine similarity for ingredient-action embeddings:

Conclusions: Actions do not make sense: elbow macaroni - burn; ground pork - support, etc. Therefore, cosine similarity cannot be used for these embeddings, due to the specification of EaT-PIM embeddings

  • Workaround:

    • By the use of Tf-Idf, process all recipes used in embeddings to find the most similar recipes to our set of ingredients:

    Source code to Tf-Idf Embeddings can be found here: Tf-Idf Embeddings

    Source code to Final Recipe Acquisition can be found here: Final Recipe Acquisition

    Conclusions: The most similar recipe was determined (RECIPE_OUTPUT_323553)

    • Take actions from the final recipe and apply them to our ingredients:

    Conclusions: This way we could determine more relevant actions to the set of ingredients: elbow macaroni - cover; ground pork - cook, etc.

Resources