Recommender systems have become an important personalization technique
on the web and are widely used especially in e-commerce applications.
However, operators of web shops and other platforms are challenged by
the large variety of available algorithms and the multitude of their
possible parameterizations. Since the quality of the recommendations that are
given can have a significant business impact, the selection
of a recommender system should be made based on well-founded evaluation
data. The literature on recommender system evaluation offers a large
variety of evaluation metrics but provides little guidance on how to choose
among them. The paper which is presented in this presentation focuses on the often neglected aspect of clearly defining the goal of an evaluation and how this goal relates to the
selection of an appropriate metric. We discuss several well-known
accuracy metrics and analyze how these reflect different evaluation goals. Furthermore we present some less well-known metrics as well as a variation of the area under the curve measure that are particularly suitable for the evaluation of
recommender systems in e-commerce applications.
Setting Goals and Choosing Metrics for Recommender System Evaluations
1. Setting Goals and Choosing Metrics for Recommender
System Evaluations
Gunnar Schröder, Maik Thiele, Wolfgang Lehner
Gunnar Schröder UCERSTI 2 Workshop
T-Systems Multimedia Solutions at the 5th ACM Conference on
Dresden University of Technology Recommender Systems
Chicago, October 23th, 2011
2. How Do You Evaluate Recommender Systems?
RMSE
Precision
F1-Measure
Recall MAE
ROC Curves
Qualitative Techniques
Quantitative Techniques
User-Centric Evaluation
Mean Average Precision Area under the Curve
Accuracy Metrics Non-Accuracy Metrics
But why do you do it exactly this way?
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
3. Some of the Issues This Paper Tries to Touch
A large variety of metrics have been published
Some metrics are highly correlated [Herlocker 2004]
Little guidance for evaluating recommenders and choosing metrics
Which aspects of the usage scenario and the data influence the choice?
Which metrics are applicable?
What do these metrics express?
What are differences among them?
Which metric represents our use-case best?
How much do the metrics suffer from biases?
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
4. Factors That Influence the Choice of Evaluation Metrics
Objectives for recommender usage
Business goals User interests
Recommender task and interaction
Prediction Classification Ranking Similarity Presentation
Preference data
Explicit Implicit Unary Binary Numerical
Choice of metrics
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
5. Major Classes of Evaluation Metrics
Prediction Accuracy Metrics
Ranking Accuracy Metrics
Classification Accuracy Metrics
Non-Accuracy Metrics
5.0 4.8 4.7 4.3 3.8 3.2 2.4 2.1 1.6 1.2
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
6. Why Precision, Recall and F1-Measure May Fool You
Ideal recommender (example a – f) vs. Worst-case recommender (ex. g – l )
Four recommendations (R1 – R4) e.g. Precision@4
Ten items with a varying ratio of relevant items (1 – 9 relevant items)
Precision, recall and F1-measure are very sensitive to the ratio of relevant items Figure 3
They fail to distinguish between an ideal recommender and a worst-case recommender if
the ratio of relevant items is varied
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
7. What is the Ideal Length for a Top-k Recommendation List?
A typical ranking produced by a recommender on a set of ten item with four items being
relevant
The length of the top-k recommendation list is varied in examples a (k=1) to j (k=10)
Figure 1
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
8. What is the Ideal Length for a Top-k Recommendation List?
A typical ranking produced by a recommender on a set of ten item with four items being
relevant
The length of the top-k recommendation list is varied in examples a (k=1) to j (k=10)
2.
1.
2.
2.
3.
part of
Figure 1
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
9. What is the Ideal Length for a Top-k Recommendation List?
A typical ranking produced by a recommender on a set of ten item with four items being
relevant
The length of the top-k recommendation list is varied in examples a (k=1) to j (k=10)
2. 3.
1. 2. 1.
3. 1. 2.
3. 3.
3.
part of
Markedness = Precision + InvPrecision – 1 Figure 1
Informedness = Recall + InvRecall – 1
Matthew’s Correlation =
[Powers 2007]
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
10. From Simple Classification Measures to Partial Ranking Measures
Moving a single relevant item among the recommenders ranking (examples a - j)
Idea: Consider both classification and ranking for the top-k recommendations Figure 2
Area under the Curve => Limited Area under the Curve
Boolean Kendall’s Tau => Limited Boolean Kendall’s Tau
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
11. A Further More Complex Example to Study at Home
Figure 4
Conclusions:
For classification use markedness, informedness and Matthew’s correlation instead
of precision, recall and F1 measure
Limited area under the curve and limited boolean Kendall’s tau are useful metrics for
top-k recommender evaluations
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
12. Conclusion and Contributions
Important aspects that influence the metric choice
Objectives for recommender usage
Recommender task and interaction
Aspects of preference data
Some problems of Precision, Recall and F1-Measure
The advantages of markedness, informedness and Matthew’s correlation
Two new metrics that measure the ranking of a limited top-k list
Limited area under the curve, limited boolean Kendall’s tau
Guidelines for choosing a metric (See paper)
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder
13. Thank You Very Much!
Do not hesitate to contact me, if you have any
questions, comments or answers!
Slides are available via e-mail or slideshare
Setting Goals and Choosing Metrics for Recommender System Evaluation - Gunnar Schröder