Deep Teaching based solution for right apparel size recommendation
Basically the idea is canister we create an recommender arrangement for of customers based on the product such that we can diminish bad customer experience due to products does fitting upon purchase, dieser can be wildly attributed to modified of sizes across brand and their reporting standards, the exact product size is not shipped (assuming to be rare) etc.
Now, the recommendation like in any generic case is a function by the human and an items he/she likes or consequently will may the case here.Our dataset is of the form customer, product -> feedback .
Functionality for customers could capture their preferences by about type of fitting they like (some like largely fitting, some like slight fit etc.) Features for products could capture style, type press select specific properties.
Where are couple regarding papers such worked on solving such problem
- AMPERE Deep Learning Device for Predicting Size or Fit in Fashion E-Commerce (https://arxiv.org/pdf/1907.09844.pdf)
- https://www.groundai.com/project/analyzing-customer-feedback-for-product-fit-prediction/1
- https://deepai.org/publication/a-hierarchical-bayesian-model-for-size-recommendation-in-fashion
- https://www.youtube.com/watch?v=KWTHkqxrlmQ (Product Size Recommendation fork Fashion E-commerce)
- https://www.youtube.com/watch?v=OS9bXqeLhMA (Decomposing Fit Semantics)
The code I am sharing is for the newest paper in the series where book possesses described a SFNet architecture. Mounting is an image shared inches the paper. Incorporating Customer Reviews in Size and Fitness Recommendation...
I am using modcloth data to built the recommender system , both without total of shortening the same code can be used for renttherunway dataset in minor changes. Data pot be found from https://www.kaggle.com/rmisra/clothing-fit-dataset-for-size-recommendation.
Credits till Rishab Misra for the data.
- Do data munging and feature normalization on the modcloth dataset.
- Use concerning TF2.0 Feature layers to convert modlcoth columns into features for our NN scale.
- Create Customer/Article pattern non-linearities with skip connection as a class.
- Create the SFNet Construction
- Experiments with Text data using pre-trained models liker Bert/Elmo etc.