LSTM_Last.fm

This repository contains notebooks supporting a study on Sequence Aware Recommender System and the importance of the order within the user hsitory. The dataset is built from Last.fm dataset (data are organized in sequences of length 20, which are the splitted in two equal parts (10,10): first part will be used as input for training or inference, the second part as label for calculating metrics).
These notebooks has been uploaded from a google/colab environment.

Directories

LSTM

This directory contains the notebooks coding LSTM models (a seq2one and a seq2seq models)

• "target_ten_lastFM_1k_LSTM_Embeddings_multilabel_session_dataset_filtered_item.ipynb" for LSTM Seq2Seq model
  • as this model give the better results, we implement in data preparation the possibility of removing items form validation and test set which are not in the train set (see "Filter items from validation and test set which are not in the train set"). this step can be skipped
• "target_one_lastFM_1k_LSTM_Embeddings_multilabel_session_dataset_filtered_item.ipynb" for LSTM Seq2one model

The data building is included in these notebooks. However as it takes time to run, we implement the possibility to output files, so that they can be reload later (without running the data processing). this process is lead by the boolean DATA_SAVED (if True, the data building part should be skipped, and only

Baselines

Contains a notebooks gathering many common recommender used as baseline. Contain as well at the end of the notebook the part "Commutativity stats" performed on the dataset

• "RS_assessement.ipynb" for common Recommender systems used as reference

Dataset

Contains the dataset used (one file for train, one file for test)