Sklearn Recommender

This library implements some common recommender functions based on the Estimator and Transformer interfaces from sklearn.

Getting Started

Install it through PyPi through:

pip install sklearn-recommender

Install the library on your local distribution through:

pip install .

Tutorial

All functions of the library are build around Transformer and Estimator, allowing them to be used through a Pipeline as well as with GridSearchCV. In general the system assumes the input to be

import sklearn_recommender as skr

Transformer

User-Item

Uses a list of user-item interactions to create a user-item matrix that can also be used as input to the similarity transformer. This also supports binary interactions by setting the binarize flag in the constructor.

tf = skr.transformer.UserItemTransformer(user_col='user_id', item_col='item_id', value_col='ranking', agg_fct='mean')
user_item = tf.transform(df_reviews)

Similarity

Creates a similarity matrix based on the given input data. It assumes that each row in a matrix is single item and computes the similarity between them according to the features listed in the given cols. The resulting dataframe will have the index_col as index (or preserves the original index if index_col is not given). Note that cols can be a list of names or a tuple defining the position of the first and last column to use.

tf = skr.transformer.SimilarityTransformer(cols=(2, -1), index_col='item_id', normalize=True)
sim_mat = tf.transform(items)

GloVe

Based on the Global Vector for Word Embeddings, this implements a transform that create n-dimensional word embeddings based on a list of input texts. The library comes with functions to download pre-trained models from the project website (note of caution: these models can take 3+GB of additional disk space). There are current two pre-trained models integrated: 'wikipedia' (which only has 300-dimensional embeddings) and 'twitter' (coming with 25, 50, 100 and 200 dimensional embeddings). There are also multiple ways to create embeddings for the given text (as it spans more than one word):

• word - generates a list of simple word embeddings (only recommended for single words)
• sent - creates a document embedding by adding up all vectors and normalizing them
• sent-matrix - creates a matrix with max_feat rows that contains the embeddings for the first max_feat words (if less words it is filled with random vectors according to distribution of vector space)
• centroid - Takes all word embeddings in the given text and computes the max_feat centroids for the clusters of the vectors

# optionally download the requried models
skr.glove.download('twitter')
tf = skr.glove.GloVeTransformer('twitter', 25, 'sent', tokenizer=skr.nlp.tokenize_clean)

Recommender

Similarity

Recommendations are made based on the similarity of item. That requires the id of an item to be given and returns the n most similar candidates.

rec = skr.recommender.SimilarityRecommender(5)
rec.fit(sim_mat)
# predict the 5 most similar items to the given items 5, 6 and 7 respectively
rec.predict([5, 6, 7])

Cross-Similarity

Collaborative-Filtering based approach. This uses the most similar items on one dimensions (e.g. most similar users to the given user) to predict the most relevant items along a different dimension (e.g. the items the most similar users interacted the most with).

rec = skr.recommender.CrossSimilarityRecommender(5)
rec.fit((user_item, sim_mat, ))
rec.predict([10, 12])

Helper Functions

Apart from the sklearn extensions, there are also various helper functions that help to prepare data for the training or reasoning process.

Train-Test Split:

Train-Test split for user-item interactions.

df = ...
# create a 30% size test set
train_df, test_df = skr.train_test_split(df, split_size=0.3)

NLP Functions:

In combination with text embeddings, there are some functions to tokenize input words using functions from nltk.

tokens = skr.nlp.tokenize_clean('Just a simple sample text.')

Design Philosophy

The library is build on top of the sklearn interfaces to allow easy chaining of pipelines and expects pandas dataframes as inputs. The general goal is to allow the quick and easy exploration of data relevant to recommender systems as well as the quick building of a baseline recommender.

Future Work

• Implement sufficient test coverage
• Add type tests to code (+ conversions to required datatypes between numpy and pandas)
• Implement additional guarantees into the train_test_split (e.g. coverage of item ids)

License

The code is published under MIT License