Skip to content

Complexly represent contents, build recommender systems, evaluate them. All in one place!

License

Notifications You must be signed in to change notification settings

swapUniba/ClayRS

Repository files navigation

ClayRS logo

ClayRS

Build Status   Docs   codecov   Python versions

ClayRS is a python framework for (mainly) content-based recommender systems which allows you to perform several operations, starting from a raw representation of users and items to building and evaluating a recommender system. It also supports graph-based recommendation with feature selection algorithms and graph manipulation methods.

The framework has three main modules, which you can also use individually:

ClayRS

Given a raw source, the Content Analyzer:

  • Creates and serializes contents,
  • Using the chosen configuration

The RecSys module allows to:

  • Instantiate a recommender system
    • Using items and users serialized by the Content Analyzer
  • Make score prediction or recommend items for the active user(s)

The EvalModel has the task of evaluating a recommender system, using several state-of-the-art metrics

Code examples for all three modules will follow in the Usage section

Installation

ClayRS requires Python 3.7 or later, while package dependencies are in requirements.txt and are all installable via pip, as ClayRS itself.

To install it execute the following command:

pip install clayrs

Usage

Content Analyzer

The first thing to do is to import the Content Analyzer module

  • We will access its methods and classes via dot notation
import clayrs.content_analyzer as ca

Then, let's point to the source containing raw information to process

raw_source = ca.JSONFile('items_info.json')

We can now start building the configuration for the items

  • Note that same operations that can be specified for items, could be also specified for users, via the ca.UserAnalyzerConfig class
# Configuration of item representation
movies_ca_config = ca.ItemAnalyzerConfig(
    source=raw_source,
    id='movielens_id',  # id which uniquely identifies each item
    output_directory='movies_codified/'  # where items complexly represented will be stored
)

Let's represent the plot field of each content with a TfIdf representation

  • Since the preprocessing parameter has been specified, then each field is first preprocessed with the specified operations
movies_ca_config.add_single_config(
    'plot',
    ca.FieldConfig(ca.SkLearnTfIdf(),
                   preprocessing=ca.NLTK(stopwords_removal=True,
                                         lemmatization=True),
                   id='tfidf')  # Custom id
)

To finalize the Content Analyzer part, let's instantiate the ContentAnalyzer class by passing the built configuration and by calling its fit() method

  • The items will be created with the specified representations and serialized
ca.ContentAnalyzer(movies_ca_config).fit()

RecSys

Similarly above, we must first import the RecSys module

import clayrs.recsys as rs

Then we load the rating frame from a TSV file

  • In this case in our file the first three columns are user_id, item_id, score in this order
    • If your file has a different structure you must specify how to map the column via parameters, check documentation for more
ratings = ca.Ratings(ca.CSVFile('ratings.tsv', separator='\t'))

Let's split with the KFold technique the loaded rating frame into train set and test set

  • since n_splits=2, train_list will contain two train_sets and test_list will contain two test_sets
train_list, test_list = rs.KFoldPartitioning(n_splits=2).split_all(ratings)

In order to recommend items to users, we must choose an algorithm to use

  • In this case we are using the CentroidVector algorithm which will work by using the first representation specified for the plot field
  • You can freely choose which representation to use among all representation codified for the fields in the Content Analyzer phase
centroid_vec = rs.CentroidVector(
    {'plot': 'tfidf'},
  
    similarity=rs.CosineSimilarity()
)

Let's now compute the top-10 ranking for each user of the train set

  • By default the candidate items are those in the test set of the user, but you can change this behaviour with the methodology parameter

Since we used the kfold technique, we iterate over the train sets and test sets

result_list = []

for train_set, test_set in zip(train_list, test_list):
  
  cbrs = rs.ContentBasedRS(centroid_vec, train_set, 'movies_codified/')
  rank = cbrs.fit_rank(test_set, n_recs=10)

  result_list.append(rank)

EvalModel

Similarly to the Content Analyzer and RecSys module, we must first import the evaluation module

import clayrs.evaluation as eva

The Evaluation module needs the following parameters:

  • A list of computed rank/predictions (in case multiple splits must be evaluated)
  • A list of truths (in case multiple splits must be evaluated)
  • List of metrics to compute

Obviously the list of computed rank/predictions and list of truths must have the same length, and the rank/prediction in position will be compared with the truth at position

em = eva.EvalModel(
    pred_list=result_list,
    truth_list=test_list,
    metric_list=[
        eva.NDCG(),
        eva.Precision(),
        eva.RecallAtK(k=5)
    ]
)

Then simply call the fit() method of the instantiated object

  • It will return two pandas DataFrame: the first one contains the metrics aggregated for the system, while the second contains the metrics computed for each user (where possible)
sys_result, users_result =  em.fit()

Note that the EvalModel is able to compute evaluation of recommendations generated by other tools/frameworks, check documentation for more