It would be nice if the configuration file and a small 'toy' dataset could be added to the file with the binary executables. So that all files necessary are available at once and in one place.

• add icons at the beginning of README: Crate Version, API docs, Crate License
• removed step for installing unnecessary libraries in the beginning of the release CI

Currently our github workflow generates binary artifacts and cargo crate. The workflow takes care of the versioning for the artifacts, but it does not update the versioning for the cargo crate (in the cargo.toml file), so there's a mismatch of version numbers between these -- and, if cargo.toml is not updated with the right version, the pipeline fails to publish the crate. Investigate and propose a reasonable way to sync versioning.

The config library currently used is not the most used/supported by the Rust community. It would be great to replace the current approach to use serde and config (example: https://github.com/paulobarchi/stdbm/blob/main/src/settings.rs).

The evaluate_file.rs presents the prediction latencies.

These percentiles that are shown are currently hardcoded.

Make the percentiles configurable using the config file.

in the config file introduce a new section that evaluate_file.rs should use: [evaluation] latency_percentiles = [0, 25, 50, 75, 90, 99.5]

Add the VSKNN_STAN algorithm.

The VSKNN_STAN algorithm is introduced in the paper: Empirical Analysis of Session-Based Recommendation Algorithms

VSKNN_STAN combines the ideas from STAN and VS-kNN in a single approach and adds a sequence-aware item scoring procedure. Its very likely that the VSKNN_STAN algorithm can leverage the VMISIndex to do its computation.

A python implementation can be found here: https://github.com/rn5l/session-rec/blob/master/algorithms/knn/vstan.py

Add the STAN algorithm to Serenade. Sequence and Time Aware Neighborhood for Session-based Recommendations: STAN https://dl.acm.org/doi/10.1145/3331184.3331322

This method called “Sequence and Time Aware Neighborhood” was presented at SIGIR ’19. STAN is based on SKNN, but it additionally takes into account the following factors for making recommendations: i) the position of an item in the current session, ii) the recency of a past session w.r.t. to the current session, and iii) the position of a recommendable item in a neighboring session. Their results show that STAN significantly improves over SKNN.

Its very likely that the STAN algorithm can leverage the VMISIndex to do its computation.

A python implementation of STAN can be found here: https://github.com/rn5l/session-rec/blob/master/algorithms/knn/stan.py

The evaluation metric (MRR@20) that is used for hyperparameter search is currently hardcoded. Serenade has supports for multiple evaluation metrics such as MRR, F1, HitRate, Ndcg, Popularity. Precision, Recall etc.

The choice for which evaluation metric to be used, should be configurable by the user via the configuration file.

Proposed solution: in the config file introduce a new variable for the section 'hyperparam' used by tpe_hyperparameter_optm.rs: metric_name = "MRR" metric_length = 20

in the config file introduce a new section that evaluate_file.rs should use: [evaluation] metric_name = "MRR" metric_length = 20 training_data_path = "train.txt" test_data_path = "test.txt"

Both the tpe_hyperparameter_optm.rs and evaluate_file.rs should be included to support this feature.

The trait SessionMetric could be extended with a function fn is_larger_better(&self) -> bool
By default all metrics must return true for fn is_larger_better except for popularity.rs which should return false

The tpe_hyperparameter_optm.rs ( line 130 ) should use the 'max' or 'min' score depending on the is_larger_better value.