On-line recommender systems have diferent challenges to overcome to provide content to users. One of these is the potential of isolating users from a diverse set of items by recommending very narrow content. In this paper we propose an item-centric view of a recommender system, looking at the exposure of items to groups of consumers, and how diverse those groups are, to identify if items are recommended to narrower groups of consumers. This is opposite to current practice where diversity of content is typically analysed. Preliminary results on the MovieLens 20M dataset show that recommender systems expose items to narrower groups of consumers, and these groups are less diverse.
Recommender systems have become ubiquitous in the interfaces to product catalogues provided by on-line retailers. From the user’s perspective, recommender algorithms are used to filter a large set of possible selections into a much smaller set of items that the user is likely to be interested in. On the other hand, from the business point of view, as important as users getting engaging recommendations is the utilisation of products in the catalogue.
Sales increase or redistribution across the whole catalogue of items might not be the only business goal to be addressed by a recommender system. In some sense, recommender systems are marketing tools that identify customers and target these customers with personalised items. Questions arise: are we exposing items to users that showed an interest before? Are we promoting items to reach new groups of customers? How diverse are these groups? From market development perspective, recommender system should help us in achieving all of these business goals. To measure and control for this, we need a picture of how items are exposed to diferent groups of people, and if the exposure is diverse.
In this paper we tackle the problem of item exposure to understand who consumes items and if potential consumers are reached by recommendations. We measure diversity of the people getting recommendations for an item, using approaches coming from ecology, such as species diversity of a habitat. This is diferent to the content diversity of recommendations that has been typically considered in the context of diversity in recommender systems. The main goal of this paper is to find the answer to the following question: do
Recommender systems have to deal with the long tail of items that are rarely recommended. This includes niche items that are rarely liked, but also items that have not penetrated the market. To identify and promote these items, we argue it is not enough to ask how many users have rated each item in the past, but also which users have rated the items, which define its item exposure.
An item’s user profile, Ui , contains the set of users who rated the item in the past. A diversity measure over these users gives insight into the extent to which item has been exposed to a wide range of diferent user types. Similarly, the set of users to whom the item is recommended, Ri , can be analysed to reveal the extent to which recommendations extend the exposure of an item. If an item is recommended to diverse consumers, it is possible that the item can reach a wider potential market.
As it is commonplace for marketeers to model their customerbase through customer segmentation, we find it useful to measure the diversity in terms of the spread across diferent consumer segments. Given a partition Pc of U into k consumer segments, U = C1 ∪ C2 ∪ ... ∪ Ck , where Cj is the jth consumer segment, we define consumer diversity of a set of consumers S, as functions of (p1, ..., pk ), where pj = |S ∩Cj | is the proportion of the set S that |S | belong to consumer segment Cj .
A similar problem is considered in ecology, where a habitat can
be quantified in terms of species diversity [
j
j=1
!1/(1−q)
and 1D = limq→1 q D = exp(H (p)). In biodiversity these are called
true diversities or efective number of species [
0.40 0 1
UB 5.11 2.84 2.34
IB 3.73 2.99 1.95
MF 6.43 3.53 2.84 of diferent items to the consumer segments or to compare the exposure of a single item under diferent conditions. 3
ANALYSIS OF CONSUMER DIVERSITY
We investigate consumer diversity on the MovieLens 20M dataset
[
We wonder if recommender systems might sufer not only from narrowing content served to users, but also items being exposed to narrow audiences. To illustrate that, we take a movie (The Matrix) for which we show distribution of consumers over segments— Figure 1. It can be seen that one segment (no. 4) is over-represented almost 4 times in recommendations. We measured its true diversities: richness, Shannon and Simpson indices. Richness decreased from 15 to 13 which means 2 segments are not reached, Shannon and Simpson indices also dropped, respectively, from to 9.80 to 6.14, and 8.85 to 4.43, which means that recommendations are generally less diverse in terms of consumers to which this movie has reached. True diversities are also easy to interpret—they tell the efective number of species, the number of equally abundant species that produce same diversity. In our case, recommendations are 1.5 times less diverse on Shannon index, and 2 times on Simpson index.
Table 1 contains values of considered true diversity indices, averaged over all items. Richness shows that on average items are consumed by users of 13 out of 15 segments, but only recommended to 3-6 segments. If concentration is taken into account, Shannon and Simpson indices drop, indicating items being 2-3 times less diverse. Paired t-test show significance of the diferences ( p < 0.001).
As item’s popularity can afect collaborative filtering methods, we wonder if lower diversity is due to low item popularity. To examine this, we split the items into the head of most popular items (80% of interactions), and the rest in the tail—histogram of Shannon diversity in Figure 2. On dataset, head items have higher diversity, while tail tends to obtain lower values. Recommendations do not follow these—both groups of items have distribution of diversity skewed towards 0. This suggests that even popular items, receiving more interactions, are isolated from wide and diverse consumers. 4
Diversity is commonly studied in the context of items that are
recommended to users, which might help mitigating the problem
of users being exposed to narrower spectrum of item types. A
number of frameworks have been proposed to measure and increase
diversity, such as Intra-List Diversity [
Sales diversity [
In information retrieval, a concept of profile diversity [
In this paper we identified and explored the problem of consumer diversity, which measures how diverse each item is in terms of consumer segments. Our analysis shows that popular recommendation techniques expose items to much narrower and less diverse consumers. Although the overall quality of recommendations might be good, items are hidden from certain groups of people who expressed an interest in them in the past.