Glossary:R-Metric: Difference between revisions

From SuperMemo Help
Jump to navigation Jump to search
mNo edit summary
mNo edit summary
Line 1: Line 1:
;<span id="R-Metric">R-Metric</span> (Recall Metric)
;<span id="R-Metric">R-Metric</span> (Recall Metric)
:[[File:Daily Algorithm SM-17 performance metric.jpg|thumb|SuperMemo 18 performance metric]]absolute measure of performance of two [[Glossary:Spaced_repetition|spaced repetition]] algorithms based on their ability to predict recall before a [[Glossary:Grade|grade]] is scored. In [[What's new in SuperMemo 18?|SuperMemo 18]], '''R-Metric''' is used solely to compare [https://supermemo.guru/wiki/Algorithm_SM-15 Algorithm SM-15] (known from SuperMemo 16) and the new [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18]. It is shown as percentage in '''Statistics''' and '''[[Toolkit menu|Toolkit]] : [[Toolkit menu#Statistics|Statistics]] : [[Analysis]] : [[Analysis#Use|Use]] : Efficiency : [[Analysis#Use_:_Efficiency_:_R-Metric|R-Metric]]'''. '''R-Metric''' is a difference between the performance of the two algorithms: <code>R-Metric=LSRM(Alg-15)-LSRM(Alg-18)</code>, where <code>LSRM</code> is the least squares predicted recall measure for a given algorithm. '''R-Metric''' greater than zero shows superiority of [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18]. '''R-Metric''' less than zero indicates underperformance of the new algorithm. <code>LSRM</code> is a square root of the average of squared absolute differences in recall predictions: <code>abs(Recall-PredictedRecall)</code>, where <code>Recall</code> is 0 for failing [[Glossary:Grade|grades]] and <code>Recall</code> is 1 for passing grades. <code>PredictedRecall</code> is a prediction issued by the algorithm before the [[Glossary:Repetition|repetition]]. In [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18], the prediction is a weighted average of the value taken from the [[Glossary:Recall_matrix|Recall[] matrix]], and [[Glossary:Retrievability|R (retrievability)]] computed from [[Glossary:Stability|S (stability)]] and the used [[Glossary:Interval|interval]]. The weight used is based on prior repetition cases which inform of the significance of the Recall[] matrix prediction (the prediction becomes more meaningful with more prior repetition data)
:[[File:Recall_metric.jpg|thumb|SuperMemo 18 performance metric]]absolute measure of performance of two [[Glossary:Spaced_repetition|spaced repetition]] algorithms based on their ability to predict recall before a [[Glossary:Grade|grade]] is scored. In [[What's new in SuperMemo 18?|SuperMemo 18]], '''R-Metric''' is used solely to compare [https://supermemo.guru/wiki/Algorithm_SM-15 Algorithm SM-15] (known from SuperMemo 16) and the new [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18]. It is shown as percentage in '''Statistics''' and '''[[Toolkit menu|Toolkit]] : [[Toolkit menu#Statistics|Statistics]] : [[Analysis]] : [[Analysis#Use|Use]] : Efficiency : [[Analysis#Use_:_Efficiency_:_R-Metric|R-Metric]]'''. '''R-Metric''' is a difference between the performance of the two algorithms: <code>R-Metric=LSRM(Alg-15)-LSRM(Alg-18)</code>, where <code>LSRM</code> is the least squares predicted recall measure for a given algorithm. '''R-Metric''' greater than zero shows superiority of [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18]. '''R-Metric''' less than zero indicates underperformance of the new algorithm. <code>LSRM</code> is a square root of the average of squared absolute differences in recall predictions: <code>abs(Recall-PredictedRecall)</code>, where <code>Recall</code> is 0 for failing [[Glossary:Grade|grades]] and <code>Recall</code> is 1 for passing grades. <code>PredictedRecall</code> is a prediction issued by the algorithm before the [[Glossary:Repetition|repetition]]. In [https://supermemo.guru/wiki/Algorithm_SM-18 Algorithm SM-18], the prediction is a weighted average of the value taken from the [[Glossary:Recall_matrix|Recall[] matrix]], and [[Glossary:Retrievability|R (retrievability)]] computed from [[Glossary:Stability|S (stability)]] and the used [[Glossary:Interval|interval]]. The weight used is based on prior repetition cases which inform of the significance of the Recall[] matrix prediction (the prediction becomes more meaningful with more prior repetition data)

Revision as of 15:37, 12 April 2019

R-Metric (Recall Metric)
SuperMemo 18 performance metric
absolute measure of performance of two spaced repetition algorithms based on their ability to predict recall before a grade is scored. In SuperMemo 18, R-Metric is used solely to compare Algorithm SM-15 (known from SuperMemo 16) and the new Algorithm SM-18. It is shown as percentage in Statistics and Toolkit : Statistics : Analysis : Use : Efficiency : R-Metric. R-Metric is a difference between the performance of the two algorithms: R-Metric=LSRM(Alg-15)-LSRM(Alg-18), where LSRM is the least squares predicted recall measure for a given algorithm. R-Metric greater than zero shows superiority of Algorithm SM-18. R-Metric less than zero indicates underperformance of the new algorithm. LSRM is a square root of the average of squared absolute differences in recall predictions: abs(Recall-PredictedRecall), where Recall is 0 for failing grades and Recall is 1 for passing grades. PredictedRecall is a prediction issued by the algorithm before the repetition. In Algorithm SM-18, the prediction is a weighted average of the value taken from the Recall[] matrix, and R (retrievability) computed from S (stability) and the used interval. The weight used is based on prior repetition cases which inform of the significance of the Recall[] matrix prediction (the prediction becomes more meaningful with more prior repetition data)