Ranking is a problem in machine learning where the objective is to sort a list of documents for an end user in the most suitable way, so the most relevant documents appear on top. Ranking appears in several domains of data science, starting from recommender systems where an algorithm suggests a set of items for purchase and ending up with NLP search engines where by a given query, the system tries to return the most relevant search results.
The question which arises naturally is how to estimate the quality of a ranking algorithm. As in classical machine learning, there does not exist a single universal metric that would be suitable for any type of task. Why? Simply because every metric has its own application scope which depends on the nature of a given problem and data characteristics.
That is why it is crucial to be aware of all the main metrics to successfully tackle any machine learning problem. This is exactly what we are going to do in this article.
Nevertheless, before going ahead let us understand why certain popular metrics should not be normally used for ranking evaluation. By taking this information into consideration, it will be easier to understand the necessity of the existence of other, more sophisticated metrics.
Note. The article and used formulas are based on the presentation on offline evaluation from Ilya Markov.
There are several types of information retrieval metrics that we are going to discuss in this article:
Imagine a recommender system predicting ratings of movies and showing the most relevant films to users. Rating usually represents a positive real number. At first sight, a regression metric like MSE (RMSE, MAE, etc.) seems a reasonable choice to evaluate the quality of the system on a hold-out dataset.
MSE takes all the predicted films into consideration and measures the average square error between true and predicted labels. However, end users are usually interested only in the top results which appear on the first page of a website. This indicates that they are not really interested in films with lower ratings appearing at the end of the search result which are also equally estimated by standard regression metrics.
A simple example below demonstrates a pair of search results and measures the MSE value in each of them.
Though the second search result has a lower MSE, the user will not be satisfied with such a recommendation. By first looking only at non-relevant items, the user will have to scroll up all the way down to find the first relevant item. That is why from the user experience perspective, the first search result is much better: the user is just happy with the top item and proceeds to it while not caring about others.
The same logic goes with classification metrics (precision, recall) which consider all items as well.
What do all of described metrics have in common? All of them treat all items equally and do not consider any differentiation between high and low-relevant results. That is why they are called unranked.
By having gone through these two similar problematic examples above, the aspect we should focus on while designing a ranking metric seems more clear:
A ranking metric should put more weight on more relevant results while lowering or ignoring the less relevant ones.
Kendall Tau distance
Kendall Tau distance is based on the number of rank inversions.
An invertion is a pair of documents (i, j) such as document i having a greater relevance than document j, appears after on the search result than j.
Kendall Tau distance calculates all the number of inversions in the ranking. The lower the number of inversions, the better the search result is. Though the metric might look logical, it still has a downside which is demonstrated in the example below.
It seems like the second search result is better with only 8 inversions versus 9 in the first one. Similarly to the MSE example above, the user is only interested in the first relevant result. By going through several non-relevant search results in the second case, the user experience will be worse than in the first case.
Precision@k & Recall@k
Instead of usual precision and recall, it is possible to consider only at a certain number of top recommendations k. This way, the metric does not care about low-ranked results. Depending on the chosen value of k, the corresponding metrics are denoted as precision@k (“precision at k”) and recall@k (“recall at k”) respectively. Their formulas are shown below.
Imagine top k results are shown to the user where each result can be relevant or not. precision@k measures the percentage of relevant results among top k results. At the same time, recall@k evaluates the ratio of relevant results among top k to the total number of relevant items in the whole dataset.
To better understand the calculation process of these metrics, let us refer to the example below.
There are 7 documents in the system (named from A to G). Based on its predictions, the algorithm chooses k = 5 documents among them for the user. As we can notice, there are 3 relevant documents (A, C, G) among top k = 5 which results in precision@5 being equal to 3 / 5. At the same time, recall@5 takes into account relevant items in the whole dataset: there are 4 of them (A, C, F and G) making recall@5 = 3 / 4.
recall@k always increases with the growth of k making this metric not really objective in some scenarios. In the edge case where all the items in the system are shown to the user, the value of recall@k equals 100%. precision@k does not have the same monotonic property as recall@k has as it measures the ranking quality in relation to top k results, not in relation to the number of relevant items in the whole system. Objectivity is one of the reasons precision@k is usually a preferred metric over recall@k in practice.
AP@k (Average Precision) & MAP@k (Mean Average Precision)
The problem with vanilla precision@k is that it does not take into account the order of relevant items appearing among retrieved documents. For example, if there are 10 retrieved documents with 2 of them being relevant, precision@10 will always be the same despite the location of these 2 documents among 10. For instance, if the relevant items are located in positions (1, 2) or (9, 10), the metric does differentiate both of these cases resulting in precision@10 being equal to 0.2.
However, in real life, the system should give a higher weight to relevant documents ranked on the top rather than on the bottom. This issue is solved by another metric called average precision (AP). As a normal precision, AP takes values between 0 and 1.
AP@k calculates the average value of precision@i for all values of i from 1 to k for those of which the i-th document is relevant.
In the figure above, we can see the same 7 documents. The response to the query Q₁ resulted in k = 5 retrieved documents where 3 relevant documents are positioned at indexes (1, 3, 4). For each of these positions i, precision@i is calculated:
- precision@1 = 1 / 1
- precision@3 = 2 / 3
- precision@4 = 3 / 4
All other mismatched indexes i are ignored. The final value of AP@5 is computed as an average over the precisions above:
- AP@5 = (precision@1 + precision@3 + precision@4) / 3 = 0.81
For comparison, let us look at the response to another query Q₂ which also contains 3 relevant documents among top k. Nevertheless, this time, 2 irrelevant documents are located higher in the top (at positions (1, 3)) than in the previous case which results in lower AP@5 being equal to 0.53.
Sometimes there is a need to evaluate the quality of the algorithm not on a single query but on multiple queries. For that purpose, the mean average precision (MAP) is utilised. Is is simply takes the mean of AP among several queries Q:
The example below shows how MAP is calculated for 3 different queries:
RR (Reciprocal Rank) & MRR (Mean Reciprocal Rank)
Sometimes users are interested only in the first relevant result. Reciprocal rank is a metric which returns a number between 0 and 1 indicating how far from the top the first relevant result is located: if the document is located at position k, then the value of RR is 1 / k.
Similarly to AP and MAP, mean reciprocal rank (MRR) measures the average RR among several queries.
The example below shows how RR and MRR are computed for 3 queries:
Though ranked metrics consider ranking positions of items thus being a preferable choice over the unranked ones, they still have a significant downside: the information about user behaviour is not taken into account.
User-oriented approaches make certain assumptions about user behaviour and based on it, produce metrics that suit ranking problems better.
DCG (Discounted Cumulative Gain) & nDCG (Normalized Discounted Cumulative Gain)
The DCG metric usage is based on the following assumption:
Highly relevant documents are more useful when appearing earlier in a search engine result list (have higher ranks) — Wikipedia
This assumption naturally represents how users evaluate higher search results, compared to those presented lower.
In DCG, each document is assigned a gain which indicates how relevant a particular document is. Given a true relevance Rᵢ (real value) for every item, there exist several ways to define a gain. One of the most popular is:
Basically, the exponent puts a strong emphasis on relevant items. For example, if a rating of a movie is assigned an integer between 0 and 5, then each film with a corresponding rating will approximatively have double importance, compared to a film with the rating reduced by 1:
Apart from it, based on its ranking position, each item receives a discount value: the higher the ranking position of an item, the higher the corresponding discount is. Discount acts as a penalty by proportionally reducing the item’s gain. In practice, the discount is usually chosen as a logarithmic function of a ranking index:
Finally, DCG@k is defined as the sum of a gain over a discount for all first k retrieved items:
Replacing gainᵢ and discountᵢ with the formulas above, the expression takes the following form:
To make DCG metric more interpretable, it is usually normalised by the maximum possible value of DCGₘₐₓ in the case of perfect ranking when all items are correctly sorted by their relevance. The resulting metric is called nDCG and takes values between 0 and 1.
In the figure below, an example of DCG and nDCG calculation for 5 documents is shown.
RBP (Rank-Biased Precision)
In the RBP workflow, the user does not have the intention to examine every possible item. Instead, he or she sequentially progresses from one document to another with probability p and with inverse probability 1 — p terminates the search procedure at the current document. Each termination decision is taken independently and does not depend on the depth of the search. According to the conducted research, such user behaviour has been observed in many experiments. Based on the information from Rank-Biased Precision for Measurement of Retrieval Effectiveness, the workflow can be illustrated in the diagram below.
Parameter p is called persistence.
In this paradigm, the user looks always looks at the 1-st document, then looks at the 2-nd document with probability p, looks at the 3-rd document with probability p² and so on. Ultimately, the probability of looking at document i becomes equal to:
The user examines document i in only when document i has just already been looked at and the search procedure is immediately terminated with probability 1 — p.
After that, it is possible to estimate the expected number of examined documents. Since 0 ≤ p ≤ 1, the series below is convergent and the expression can be transformed into the following format:
Similarly, given each document’s relevance Rᵢ, let us find the expected document relevance. Higher values of expected relevance indicate that the user will be more satisfied with the document he or she decides to examine.
Finally, RPB is computed as the ratio of expected document relevance (utility) to the expected number of checked documents:
RPB formulation makes sure that it takes values between 0 and 1. Normally, relevance scores are of binary type (1 if a document is relevant, 0 otherwise) but can take real values between 0 and 1 as well.
The appropriate value of p should be chosen, based on how persistent users are in the system. Small values of p (less than 0.5) place more emphasis on top-ranked documents in the ranking. With bigger values of p, the weight on first positions is reduced and is distributed across lower positions. Sometimes it might be difficult to find out a good value of persistence p, so it is better to run several experiments and choose p which works the best.
ERR (Expected Reciprocal Rank)
As the name suggests, this metric measures the average reciprocal rank across many queries.
This model is similar to RPB but with a little difference: if the current item is relevant (Rᵢ) for the user, then the search procedure ends. Otherwise, if the item is not relevant (1 — Rᵢ), then with probability p the user decides whether he or she wants to continue the search process. If so, the search proceeds to the next item. Otherwise, the users ends the search procedure.
According to the presentation on offline evaluation from Ilya Markov, let us find the formula for ERR calculation.
First of all, let us calculate the probability that the user looks at document i. Basically, it means that all i — 1 previous documents were not relevant and at each iteration, the user proceeded with probability p to the next item:
If a user stops at document i, it means that this document has already been looked and with probability Rᵢ, the user has decided to terminate the search procedure. The probability corresponding to this event is actually the same as the reciprocal rank equals 1 / i.
From now, by simply using the formula for the expected value, it is possible to estimate the expected reciprocal rank:
Parameter p is usually chosen close to 1.
As in the case of RBP, the values of Rᵢ can either be binary or real in the range from 0 to 1. An example of ERR calculation is demonstrated in the figure below for a set of 6 documents.
On the left, all the retrieved documents are sorted in the descending order of their relevance resulting in the best possible ERR. Contrary to the situation on the right, the documents are presented in the ascending order of their relevance leading to the worst possible ERR.
ERR formula assumes that all relevance scores are in the range from 0 to 1. In case when initial relevance scores are given from out of that range, they need to be normalised. One of the most popular ways to do it is to exponentially normalise them:
We have discussed all the main metrics used for quality evaluation in information retrieval. User-oriented metrics are used more often because they reflect real user behaviour. Additionally, nDCG, BPR and ERR metrics have an advantage over other metrics we have looked at so far: they work with multiple relevance levels making them more versatile, in comparison to metrics like AP, MAP or MRR which are designed only for binary levels of relevance.
Unfortunately, all of the described metrics are either discontinuous or flat making the gradient at problematic points equal to 0 or even not defined. As a consequence, it is difficult for most ranking algorithms to optimise these metrics directly. However, a lot of research has been elaborated in this area and many advanced heuristics have appeared under the hood of the most popular ranking algorithms to solve this issue.
All images unless otherwise noted are by the author.