# recommerder

**Repository Path**: NoaRicky/recommerder

## Basic Information

- **Project Name**: recommerder
- **Description**: recommder algorithm by tf2
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2019-06-20
- **Last Updated**: 2021-04-30

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Sequential Recommendation System for Software Crowdsourcing with pytorch

## Project Structure

- fm_rmd: project folder
    - data: modules for handling different datasets
    - models: modules for collecting different models: deep learning model or others
    - encoder: modules for encoding different features (onehot, multi-label, scalar, sequence)
- docs: contains markdown files
- tests: test samples for validating the accuracy of models
- notebooks: folders for storing jupyter notebooks
- experiments: folders for storing sh cmd files for experiment

## Content of List

## TODO List

- Training problems
  - prme model only contains 150 step records
  - transFM model loss is nan
- Global Metric Store class modification
  - preset the store tensor and set the value while interating the evaluation dataset
- Model Block: Generative Residual Block
  - The Block contains generative process in VAE and residual block in ResNet.
  - The fm block is similar with attention block. Therefore, we can try different attention machnism in experment

## Roadmap

- training base models in stackoverflow dataset with 3090
- training base models in stackoverflow dataset with 3090
- Category data encoding
  - for sequential category data, because of its sparsity, the encoder will use sparse tensor to transforme to embedding space
  - the category columns in dataframe will contain two dimensions, where first dimension is the matrix row, the other is the category id
- Item content feature appending
  - for prediction item-user pairs whose item is not viewed in training dataset, we need to add content feature for each item (not pre_item)
  - Also need to concat the content feature when negative sampling