Unit 12 of 15

Recommendations and Forecasting

  • 3 lessons
  • Videos 78 to 80

Recommendation and forecasting apply the same workflow to ranked lists and to future values. This unit matches each evaluation to the information available when the prediction is made.

A recommender scores candidate items from item content, from the interactions of similar users or from latent factors learned by matrix factorization, and ranking metrics such as precision at k and NDCG need a stated cutoff and denominator. An unobserved interaction is not a confirmed dislike. A forecast is made at an origin for a fixed horizon, with lag features that use only earlier values and with naive and seasonal-naive baselines. Rolling-origin evaluation replays the forecasts that could have been made at each time and shows how the errors change as the data drift.

Every lesson will be one video. The videos are coming soon, and each lesson page already has the definitions and an example for every key term.

Recommender Systems

  1. Recommender Systems and Matrix Factorization

    How a recommender ranks a candidate set using item content, collaborative patterns and matrix factorization, why missing feedback is not dislike, how cold start is handled, and how precision at k and NDCG grade a ranked list.

    Video 78

Time-Series Forecasting

  1. Time-Series Features and Forecasting

    How a forecast is fixed by its origin and horizon, how lag features and trailing windows use only past values of a time series, and how naive and seasonal-naive forecasts set the baselines to beat.

    Video 79

  2. Rolling Forecast Evaluation

    How rolling-origin evaluation replays the forecasts a procedure could have made, how sliding and expanding windows and an update policy shape each refit, and how forecast errors by origin and horizon reveal drift.

    Video 80