Modelling Skip-Behaviors in Sequence-Aware Recommender Systems

Sequential Recommender Systems (SRSs) operate on sequences of items as inputs. If the sequences contain noisy items, SRS will produce a sub-optimal performance as noise may dilute the signal captured by the model. The project is aimed to improve the performance of SRS in two ways: 1) to model relationships between non-adjacent items (skip-behaviors) to account for heterogeneous sequences 2) to introduce noise-filtering techniques to mitigate the impact of noise on sequences. In particular, two models are proposed: Sequential Recommender with Association Rules (SRAR) uses assocition rules to filter out noisy data; Sequential Recommender with Attention (SRAttn) integrates an attention mechanism to detect and deactivate the embeddigns of irrelevant items.

Requirements

• Python 3
• PyTorch v1.0+

Training

The default hyperparameters are optimal for the MovieLens dataset. In order to get the reported performance on Gowalla, the following hyperparameters should be changed:

python train.py SRAR --dataset=gowalla --d=100 --fc_dim=50 --l2=1e-6 --min_support=9e-5 --min_threshold=0.0366

python train.py SRAttn --dataset=gowalla --d=100 --fc_dim=50 --l2=1e-6 --num_heads=5

Last.fm hyperparameters:

python train.py SRAR --dataset=lastfm --learning_rate=0.002 --d=200 --fc_dim=50 --l2=5e-6 --num_heads=25

Important: for the Last.fm dataset, a restricting condition if p not in rated should be removed in evaluation.py:

predictions = [p for p in predictions if p not in rated]

Since the dataset contains many repetitions, the model is allowed to recommend items that are present in the training set.

Datasets

• The models were trained on three publicly available datasets: MovieLens, Last.fm, and Gowala.
• Each entry in a dataset is represented in the form of a triplet (user−item−rating)
• Each rating is substituted with 1 to adjust the data to the implicit feedback scenario.
• The data is split into train.txt and test.txt files as follows: 80% of the user’s sequence is dedicated to the training set, and 20% to the test set. The 80/20 split is subsequently performed on the training set to generate the training and validation sets that are used for parameter tuning.

Acknowledgements

The project is heavily built on top of Spotlight, Caser, and CosRec. Many thanks for their contribution to the community.