Text Similarity Recommendation System

This is a repository for Item RecSys models in Python. You can get the similar Items based on text similarity as follows.

Data Description

Input

This model recommends items that are highly related to each item in Items, which means the source of the recommended items is also Items. If you add some text data related to the corresponding Items to related_to_Items(e.g., Items description, category, etc.), it helps to increase the model accuracy.

Items = [
          'Netflix movie',
          'Netflix party',
          'Netflix top',
          'Netflix ratings',
          'rotten tomatoes ratings',
          'IMDb Top 250 Movie ratings'
          ]
          
related_to_Items = [
          ["movie top", "Netflix"],
          ["party pricing", "Netflix"],
          ["top TV shows',","Netflix"],
          ["ratings"],
          ['tomatoes'],
          ['ratings']
          ]

Output

Netflix movie
1: rotten tomatoes ratings
2: IMDb Top 250 Movie ratings
3: Netflix top

Netflix top
1: IMDb Top 250 Movie ratings
2: Netflix movie
3: Netflix ratings

IMDb Top 250 Movie ratings
1: Netflix ratings
2: Netflix top
3: Netflix movie

Process

video-TextSimila drawio

Tokenization

extract nouns from each sentence

# Example
['Netflix movie', 'Netflix party']

[['Netflix', 'movie'], ['Netflix', 'party']]

Embedding

get embedding vector from each sentence

# Example
[['Netflix', 'movie'], ['Netflix', 'party']]

[[0.94, 0.13], [0.94, 0.741]]

After training tokenization and embedding models, the models are saved automatically. You can either train models with your own corpus or use the pre-trained models.

Calculate cosine similarity

calculate the similarity between item embedding vectors using cosine similarity.

$$ emb_A : \text{embedding vector of item A}\\ emb_B : \text{embedding vector of item B}\\ cos(emb_A,emb_B) = \frac{emb_A\cdot emb_B}{ |emb_A| |emb_B|} $$

Installation

pip install TextSimila

Prerequisites

python version should be greater than 3.7.x

pip install -r requirements.txt

Quick Start

Example notebooks

Refer to sample_code.ipynb if you want to run code in a jupyter environment

Parameter Description

The tables below describe the parameters of the class text_sim_reco

class text_sim_reco(
            Items,
            related_to_Items: list =  None,
            saved: Boolean = False,
            lang = Literal["en","ko"],
            reco_Item_number: int = 3,
            ratio: float = 0.3,

            # tokenize
            pretrain_tok: Boolean = False,
            stopwords: list = None,
            extranouns: list = None,
            verbose: Boolean = False,
            min_noun_frequency: int = 1,
            max_noun_frequency: int = 80,
            max_frequency_for_char: int = 20,
            min_noun_score: float = 0.1,
            extract_compound: Boolean = False,
            model_name_tok: str = None,
            
            # embedding
            pretrain_emb: Boolean = False,
            vector_size: int = 15,
            window: int = 3,
            min_count: int = 1,
            workers: int = 4,
            sg: Literal[1, 0] = 1,
            model_name_emb: str = None)
Parameters Attributes
Items : List[str] (required) A list of text data to recommend
related to Items : List[List] (optional) A list of text data related to Items that helps to recommend
saved: Boolean, default = False (optional) Whether to save the model
lang: Literal[“en”,”ko”], default = “en” The configure model language– ‘ko’: Your Items are in Koran – ‘en’: Your Items are in English
reco_Item_number : int, default = 3 The number of recommendations for each Item
ratio: float, default = 0.2 The minimum percentage that determines whether to create a corpus

Parameters for tokenization with Korean custom dataset Attributes
pretrain_tok: Boolean, default = False Whether to use Pre-trained model
min_noun_score = float, default = 0.1 The minimum noun score. It decides whether to combine single nouns and compounds
min_noun_frequency : int, default = 1 The minimum frequency of words that occur in a corpus. It decides whether to be a noun while training(noun extracting)
extract_compound = boolean, default = False Whether to extract compounds components ‘compounds components’: Information on single nouns that make up compound nouns
verbose: boolean, default = False Whether to print out the current vectorizing
stopwords : List, default = None (Post-preprocessing option) A List of high-frequency of words to be filtered out
extranouns: List, default = None (Post-preprocessing option) A List of nouns to be added
max_noun_frequency: int, default = 80 (Post-preprocessing option) The maximum frequency of words that occur in a corpus. It decides whether to be a noun after training
max_frequency_for_char: int, default = 20 (Post-preprocessing option) max_noun_frequency option for words with length one
model_name_tok: str = None Pre-trained model name

Parameters for embedding Attributes
pretrain_emb: Boolean, default = False Whether to use Pre-trained model
vector_size : int, default = 15 Dimensionality of the word vectors
window: int, default = 3 The maximum distance between the current and predicted word within a sentence
min_count: int, default = 3 The model ignores all words with total frequency lower than this
workers: int, default = 3 The number of worker threads to train
sg: Literal[1, 0], default = 1 Training algorithm: skip-gram if sg=1, otherwise CBOW
model_name_emb: str, default = None Pre-trained model name

Command Prompt

By running exe.py, you can perform all the processes in sample_code.ipynb at once. Note that it saves the model and the predictions in the following format at every run

# Top3_prediction.json
{
  "Item_1": [
    "recommendation_1",
    "recommendation_2",
    "recommendation_3"
  ],

  ...

  "Item_10": [
    "recommendation_1",
    "recommendation_2",
    "recommendation_3"
  ]
}

Precautions

Make sure that the following two files exist in the two folders below before executing exe.py

  1. yaml file in config folder
  2. json file in data folder

1. yaml file

If you want to adjust the hyperparameters, modify existing model.yaml.

You can also create your own yaml file, but you must follow the existing model.yaml form and save it in config folder.

2. json file

If you want to use your custom data, you must process and save it according to the format below.

[
  {
      "Items": "Item_1",
      "related_to_Items": ["related_Items", "Item_1_discription"]
  },
  
  ...

  {
      "Items": "Item_10",
      "related_to_Items": ["Item_10_channel"]
  }

]

Execute the file

To predict with newly-trained model

$ python exe.py [yaml_name] [file_name] --saved [saved]

To predict with Pre-trained model

※ If you want to use English custom dataset

$ python exe.py [yaml_name] [file_name] --pretrain_tok [pretrain_tok] --pretrain_emb [pretrain_emb]

To make it simpler,

$ python exe.py [yaml_name] [file_name] -tok [pretrain_tok] -emb [pretrain_emb]

For example,

Train ver.

# If you want to train the model without saving
$ python exe.py model.yaml sample_eng

# If you want to train the model and then save them
$ python exe.py model.yaml sample_eng --saved True

Pre-trained ver.

# If you want to use Pre-trained model for tokenization and embedding
$ python exe.py model.yaml sample_eng -tok True -emb True

GitHub

View Github