Skip to content

Latest commit

 

History

History
55 lines (42 loc) · 3.06 KB

hiije_architecture.md

File metadata and controls

55 lines (42 loc) · 3.06 KB
Raw

HIIJE

Hiije is a recommendation system currently built for grocery recommendations. The system itself is an implementation of the work of Karypis and Deshpande. The description there is very technical amd is recommended for people who actually intend to implement, extend or adapt the algorithm in some way. For

Hiije Architecture

Hiije is organised into 2 bits:

  1. Setup
  2. Recommender
  3. hiije_api

Setup

The setup basically comes up with the Item similarity matrix.

This should be done in the following ways:

  1. Ready-to-drink database file

    A file that already contains the item-similarity matrix values in some format. This file would come from a separate source where the similarity values were computed and stored in file.

  2. Text-based historical record file

    In this case the input is a file containing plain text in csv format. The system will take this file and use it to generate an item similarity matrix. A few things which will also need to be generated along the way are:

    • Unique item list. A list of each unique item appearing in the input file
    • Unique IDs for each unique item. This may be the name as given in the unique item list or a numerical UID or some other form of appropriate unique identifier
    • The item similarity matrix in some form such as a csv file

The setup also * creates default config parameters * creates database tables (it requires the database to be pre existing) * populates the database from the Unique item List file and * runs unit/integration tests

Recommender

Having access to the item similarity matrix, the recommender provides a class whose objects will:

  • Accept a list of unique items. This means that 2 hotdogs and a soda will be taken to mean [hotdog, soda].
  • Confirm that each item in the list exists in the unique items database. Raise an error if this is not the case.
  • Generate a recommendation using the input list and similarity matrix.
  • Persist the response in a DB
  • Return the reponse as a dictionary with the elements:
    • Reponse_ID- A unique ID for this response
    • Request- The items in the original request
    • Recommendation_num- The number of recommended items
    • Recommendation(s)- The recommended item(s)
    • Recommendation_score(s)- The weight given to each recommendation. This is a number between 0 and 1 where 0 means not recommended at all and 1 means highly recommended.
    • Request_time- Timestamp of the time when the request was received.
    • Request_duration- The amount of time it took to compute and return the recommendations To do this, a recommender object needs access to the item similarity matrix (file)

hiije_api

The api provides access to the services of the recommender. Using the api, it is possible to:

  • Make a request and receive a recommendation. This uses a post request with a list of either unique item names or UID's. The recommendation is returned as a json object.
  • See the last 10 recommendations.
  • See a particular recommendation given its recommendation ID