Preference learning

Step 3 of the pipeline

Preference learning infers, for each user and based on their data, a model of the user's preferences. This corresponds to the idea of "algorithmic representatives", which will participate in the digital democracy to remedy users' lack of activity and reactivity.

PreferenceLearning ¶

Bases: ABC

call ¶

__call__(
    judgments: Judgments,
    users: DataFrame,
    entities: DataFrame,
    initialization: Optional[
        dict[int, ScoringModel]
    ] = None,
    new_judgments: Optional[Judgments] = None,
) -> dict[int, ScoringModel]

Learns a scoring model, given user judgments of entities

Parameters:

Name	Type	Description	Default
`judgments`	`Judgments`	May contain different forms of judgments, but most likely will contain "comparisons" and/or "assessments"	required
`entities`	`DataFrame`	entity_id: int, index May contain others, such as vector representation	required
`initialization`	`Optional[dict[int, ScoringModel]]`	Starting models, added to facilitate optimization It is not supposed to affect the output of the training	`None`
`new_judgments`	`Optional[Judgments]`	New judgments This allows to prioritize coordinate descent, starting with newly evaluated entities	`None`

Returns:

Name	Type	Description
`user_models`	`dict[int, ScoringModel]`	user_models[user] is the learned scoring model for user

user_learn ¶

user_learn(
    user_judgments: dict[str, DataFrame],
    entities: DataFrame,
    initialization: Optional[ScoringModel] = None,
    new_judgments: Optional[dict[str, DataFrame]] = None,
) -> ScoringModel

Learns a scoring model, given user judgments of entities

Parameters:

Name	Type	Description	Default
`user_judgments`	`dict[str, DataFrame]`	May contain different forms of judgments, but most likely will contain "comparisons" and/or "assessments"	required
`entities`	`DataFrame`	entity_id: int, index May contain others, such as vector representation	required
`initialization`	`Optional[ScoringModel]`	Starting model, added to facilitate optimization It is not supposed to affect the output of the training	`None`
`new_judgments`	`Optional[dict[str, DataFrame]]`	New judgments This allows to prioritize coordinate descent, starting with newly evaluated entities	`None`

Returns:

Name	Type	Description
`model`	`ScoringModel`

UniformGBT ¶

UniformGBT(
    prior_std_dev: float = 7.0,
    convergence_error: float = 1e-05,
    cumulant_generating_function_error: float = 1e-05,
    high_likelihood_range_threshold: float = 1.0,
)

Bases: GeneralizedBradleyTerry

Parameters (TODO)

LBFGSUniformGBT ¶

LBFGSUniformGBT(
    prior_std_dev: float = 7,
    convergence_error: float = 1e-05,
    cumulant_generating_function_error: float = 1e-05,
    max_iter: int = 100,
    high_likelihood_range_threshold: float = 1.0,
)

Bases: LBFGSGeneralizedBradleyTerry

Parameters (TODO)

cumulant_generating_function ¶

cumulant_generating_function(score_diff: Tensor) -> Tensor

For.

Parameters:

Name	Type	Description	Default
`score_diff`	`Tensor`	Score difference	required

Returns:

Name	Type	Description
`out`	`float`

Preference learning