Implementation of the STAMP model; Add SBR datasets description to examples/README

[bokang-ugent]

The descriptions are added to examples/matching/README.md. Ranking metrics obtained by the NARM model is also added.

[bokang-ugent]

This is an implementation of the STAMP (Short-Term Attention/Memory Priority Model for Session-based Recommendation, Liu et. al., KDD 2018) model. The mode is also one of the earlier proposed neural model in the line of SBR research, often used as a baseline.

Changes

• STAMP model is introduced in a new file models/matching/stamp.py. A new line that imports the model is added in models/maching/__init__.py. The implementation follows directly from the paper, with a few features refer to the original TensorFlow implementation as they are not fully clear from the paper (e.g., the normalization of attention weights is not described in the paper).
• Example script examples/matching/run_sbr.py is updated to allow STAMP model to be evaluated on the benchmark datasets.
• README results are also updated to examples/matching/README.md.

Results

	Recall@20	MRR@20
YOOCHOOSE1/64	0.6675	0.2859
YOOCHOOSE1/4	0.7079	0.3074
DIGINETICA	0.5578	0.2303

• The results are similar to the scores reported in the original paper, with DIGINETICA slightly off. This could be: a. the preprocessing of DIGINETICA in this implementation deviates from the existing implementations (notice the NARM is also a bit worse than the results reported original paper on DIGINETICA ); b. the architecture reported in the paper is not exactly the one that is used in the paper, see issues reported the TensorFlow implementation.

Future work

• Tune Hyperparameter on a validation set. This could potentially improve the quantitative metrics.
• Closely follow the TensorFlow implementation could also potentially improve the scores.
• Consult multiple preprocessing pipelines from different SBR paper, and implement the processing logic that agreed by the majority.