evaluate_topics_temporally

Inputs

Outputs

from spl import SPLClient

result = SPLClient.embedded().call(
    "splime://@splime.project/case4-hn/evaluate_topics_temporally@2",
    kwargs={'cutoff_utc': 'replace-with-value', 'max_features': 0, 'min_df': 0, 'min_documents': 0, 'n_topics': 0, 'random_state': 0, 'snapshot': {}, 'stop_words': [], 'train_fraction': 0.0},
    trust=True,
)

Review before running. Public bundles contain executable code. Verify the publisher and exact content hash; a virtual environment isolates dependencies, not the operating system.

Evidence

Evidence-backed concepts

• Evaluate topics temporally
• No supervised accuracy
• Fit vocabulary
• Topics only earlier titles
• Inspect later title coverage
• Future popularity forecast claimed
• Train fraction
• N topics

• v231 Oct 20262 downloads
• v131 Oct 20262 downloads

• train_topic_modelFit a simple TF-IDF and nonnegative matrix factorization topic model on supplied training titles only. Returns portable JSON vocabulary, IDF weights, learned components, automatic top-term names and a content identity; no rule-generated labels.
• hn_topic_digestFetch real HN stories, normalize them, apply the required immutable embedded trained model, count observed attention and render notebook-safe HTML. Training is a separate operation. Published implementation executes locally in embedded mode.
• tag_titlesAssign HN titles to topics with the immutable trained model embedded as a required native graph dependency. Automatic names use learned top terms; no caller model file or retraining. Returns hn-topics-v1 predictions.
• predict_topicsApply a supplied fitted HN topic model without fitting vocabulary or components. Returns topic assignments, automatic keyword tags, unknown-topic diagnostics and unchanged observed popularity values.
• analyze_titlesApply explicit topic rules and compare document shares in adjacent equal-duration windows. Counts describe the supplied snapshot, not all Hacker News activity.