Filtering Configuration¶

The Filtering Configuration module defines the preprocessing strategies applied to the dataset before the splitting phase. Filtering is a fundamental step when the dataset contains redundant or low-quality interactions, or when its size exceeds the available computational resources.

By applying filters, WarpRec ensures that the resulting dataset is both computationally manageable and more representative of the target recommendation task.

General Configuration Format¶

Filtering strategies must be declared under the filtering section of the configuration file. Each strategy is specified by name, followed by its parameters (if required).

filtering:
    strategy_name_1:
        arg_name_1: value_1
    strategy_name_2:
        arg_name_1: value_1
        arg_name_2: value_2
...

Important

Strategies are executed top to bottom in the exact order they are listed.
Some strategies (e.g., MinRating and UserAverage) are incompatible with implicit feedback datasets.
Incorrect strategy names or invalid parameter keys will cause WarpRec to raise an error.

Supported Filtering Strategies¶

WarpRec currently supports the following filtering strategies:

Strategy	Category	Description
`MinRating`	Rating-based	Remove interactions below a rating threshold.
`UserAverage`	Rating-based	Remove interactions below each user's average rating.
`ItemAverage`	Rating-based	Remove interactions below each item's average rating.
`UserMin`	Frequency-based	Remove users with fewer than N interactions.
`UserMax`	Frequency-based	Remove users with more than N interactions.
`ItemMin`	Frequency-based	Remove items with fewer than N interactions.
`ItemMax`	Frequency-based	Remove items with more than N interactions.
`IterativeKCore`	Core-based	Iterative k-core filtering until convergence.
`NRoundsKCore`	Core-based	k-core filtering for a fixed number of rounds.
`UserHeadN`	Temporal	Retain the first N interactions per user.
`UserTailN`	Temporal	Retain the last N interactions per user.
`DropUser`	Entity removal	Remove specific users by ID.
`DropItem`	Entity removal	Remove specific items by ID.

1. MinRating

Removes all interactions where the rating value is strictly below the specified threshold. Not compatible with implicit feedback datasets.

filtering:
    MinRating:
        min_rating: 3.0

2. UserAverage

Removes all interactions for which the rating is below the corresponding user's average rating. Not applicable to implicit feedback scenarios.

filtering:
    UserAverage: {}   # No parameters required

3. ItemAverage

Removes all interactions for which the rating is below the corresponding item's average rating. Not applicable to implicit feedback scenarios.

filtering:
    ItemAverage: {}   # No parameters required

4. UserMin

Removes all interactions involving users with fewer interactions than the given threshold.

filtering:
    UserMin:
        min_interactions: 5

5. UserMax

Removes all interactions involving users with more interactions than the given threshold. This is particularly useful for cold-start user analysis.

filtering:
    UserMax:
        max_interactions: 2

6. ItemMin

Removes all interactions involving items with fewer interactions than the given threshold.

filtering:
    ItemMin:
        min_interactions: 5

7. ItemMax

Removes all interactions involving items with more interactions than the given threshold. Useful for analyzing cold-start item scenarios.

filtering:
    ItemMax:
        max_interactions: 2

8. IterativeKCore

Applies UserMin and ItemMin iteratively until no further interactions can be removed (i.e., until a stable state is reached).

filtering:
    IterativeKCore:
        min_interactions: 5

9. NRoundsKCore

Applies UserMin and ItemMin for a fixed number of iterations. This is a simplified variant of IterativeKCore that does not require full convergence.

filtering:
    NRoundsKCore:
        rounds: 3
        min_interactions: 5

Tip

IterativeKCore ensures dataset stability, but may be computationally expensive. NRoundsKCore is recommended when deterministic runtime is preferred over convergence.

10. UserHeadN

Selects and retains the first N interactions for each user. If timestamps are available, interactions are sorted chronologically before selection. If no timestamps are provided, the original ordering of interactions is preserved.

filtering:
    UserHeadN:
        num_interactions: 30

11. UserTailN

Selects and retains the last N interactions for each user. If timestamps are available, interactions are sorted chronologically before selection. If no timestamps are provided, the original ordering of interactions is preserved.

filtering:
    UserTailN:
        num_interactions: 30

12. DropUser

Filter out all interactions involving specific users identified by their user IDs.

filtering:
    DropUser:
        user_ids_to_filter: [123, 456, 789]

13. DropItem

Filter out all interactions involving specific items identified by their item IDs.

filtering:
    DropItem:
        item_ids_to_filter: [123, 456, 789]

Example Filtering Pipeline¶

The following example demonstrates a pipeline where:

All ratings below 3.0 are removed.
Users with fewer than 10 interactions are filtered out.

filtering:
    MinRating:
        min_rating: 3.0
    UserMin:
        min_interactions: 10