pub struct NextAppModelArtifact {Show 16 fields
pub schema_version: String,
pub model_id: String,
pub dataset_id: String,
pub trained_at_ms: i64,
pub config: NextAppModelConfig,
pub app_vocab: Vec<String>,
pub global_popularity: Vec<AppScore>,
pub naive_bayes: NaiveBayesModel,
pub markov: MarkovModel,
pub feature_lift: FeatureLiftModel,
pub user_frequency: BTreeMap<String, Vec<AppScore>>,
pub user_recency: BTreeMap<String, String>,
pub markov_context: BTreeMap<String, Vec<AppScore>>,
pub ensemble_combiner: EnsembleCombiner,
pub ensemble_logistic: LogisticRerankerModel,
pub training_summary: TrainingSummary,
}Fields§
§schema_version: String§model_id: String§dataset_id: String§trained_at_ms: i64§config: NextAppModelConfig§app_vocab: Vec<String>§global_popularity: Vec<AppScore>§naive_bayes: NaiveBayesModel§markov: MarkovModel§feature_lift: FeatureLiftModelLog-lift feature ensemble. The JSON field name remains xgboost for
backward compatibility with existing artifacts, but the implementation
is a lightweight deterministic feature-lift model, not an XGBoost
gradient-boosted tree ensemble.
user_frequency: BTreeMap<String, Vec<AppScore>>Per-user app-usage frequency ranking (MFU), keyed by user id. Unlike the per-user Markov table this is unconditional on the current app: it ranks the apps a user opens most often overall. Defaults to empty for backward compatibility with older artifacts.
user_recency: BTreeMap<String, String>Hard recency pointer: the single most recent next-app observed per
"{user}\t{current}" key in training order (last write wins). Unlike the
per-user Markov distribution this is a one-app “what did this user open
last time from here” signal, a strong top-1 predictor on repetitive
personal usage. Defaults to empty for backward compatibility.
markov_context: BTreeMap<String, Vec<AppScore>>Context-aware Markov transitions keyed by temporal features. Keys are
"{current}\t{hour}" or "{current}\t{weekday}", values are ranked
next-app distributions. This captures time-of-day and day-of-week
patterns that the global Markov table averages over. Defaults to empty
for backward compatibility with older artifacts.
ensemble_combiner: EnsembleCombinerLearned reciprocal-rank-fusion combiner for the ensemble algorithm. The weights are fit offline on a held-out validation slice (never on the test split) and locked into the artifact, so runtime inference is pure and deterministic with no extra dependencies. Defaults to empty, in which case the ensemble falls back to the legacy fixed-weight fusion so older committed artifacts keep working.
ensemble_logistic: LogisticRerankerModelLearned logistic candidate reranker over the same component features as
the ensemble. Empty means “use ensemble_combiner / legacy fusion”.
training_summary: TrainingSummaryTrait Implementations§
Source§impl Clone for NextAppModelArtifact
impl Clone for NextAppModelArtifact
Source§fn clone(&self) -> NextAppModelArtifact
fn clone(&self) -> NextAppModelArtifact
1.0.0 · Source§fn clone_from(&mut self, source: &Self)
fn clone_from(&mut self, source: &Self)
source. Read more