Chapter 14

Ethics & Fairness

Accuracy is not enough. A recommender that maximises click-through can also concentrate attention on a few items, trap users in filter bubbles, under-serve minorities, or push toward extreme content. This chapter asks how to *measure* these harms — the Gini coefficient for concentration, KL calibration for genre balance — and how to *fix* them — MMR and popularity-aware re-ranking — with case studies from Spotify and Facebook on how real systems face the trade-offs.

Reading: ~52 min 0 Source: Polimi Recommender Systems 2024/25 — Ethics of Recommendation (L. Boratto, guest lecture) · Carbonell & Goldstein, The Use of MMR for Reordering Documents (SIGIR 1998) · Steck, Calibrated Recommendations (RecSys 2018) · Anderson, Maystre, Anderson, Mehrotra & Lalmas, Algorithmic Effects on the Diversity of Consumption on Spotify (WWW 2020) · Bakshy, Messing & Adamic, Exposure to Ideologically Diverse News on Facebook (Science 2015)

key

The big idea

Recommendation is not value-neutral. Optimising solely for relevance concentrates attention, creates feedback loops, and can disadvantage minority users and providers. Fairness-aware recommendation means measuring the unintended harms — concentration via the Gini coefficient, bias via consumption-vs-recommendation distributions — and correcting them, either post-hoc through re-ranking (MMR, calibration, popularity penalties) or at training time through multi-objective optimisation. The goal is not maximal diversity, but to make the trade-off with relevance explicit.

01 · What could go wrong?

Fairness issues in personalised algorithms

Five ways a recommender can cause harm — even when it’s accurate.

The lecture identifies five classes of unfairness:

Filter bubbles — users are gradually locked into a narrow set of items, never seeing content outside their confirmed interests.
Extreme content — recommending content maximally aligned with a user’s interests can push them toward ever more extreme versions of those interests.
Under-representation — users or groups with sparse data receive less effective recommendations.
Narrowness — recommendations capture only a user’s predominant interest, missing the breadth of their actual behaviour.
Provider unfairness — vendors, creators, or long-tail items are systematically disadvantaged.

Two low-level algorithmic biases illustrate the problem. Maximum inner product (MF, two-tower) tends to recommend “extreme” items — if you like action, you should like a lot of action. Nearest neighbour (item-CF) tends to recommend items very close to what you’ve already consumed. Both narrow the breadth of what users see.

02 · Quantifying the problem

Measuring diversity and concentration

If you can’t measure it, you can’t fix it. Two tools: distribution comparison and the Gini coefficient.

The simplest diagnostic compares the distribution of consumptions (which items users actually interact with) against the distribution of recommendations (which items the model serves). Are the most-consumed items also the most-recommended? Are recommendations more dominated by popular items than consumptions are? The Gini coefficient quantifies inequality in a distribution:

Gini coefficient

G = \frac{\sum_{i=1}^{n}\sum_{j=1}^{n}\lvert x_i - x_j\rvert}{2\,n^2\,\bar x},\qquad \bar x = \frac{1}{n}\sum_i x_i.

$G\approx 0$ is uniform (every item gets a similar share); $G\approx 1$ is concentrated (a few items dominate). The key comparison is $G_{\text{recs}}$ vs $G_{\text{int}}$ : if $G_{\text{recs}} > G_{\text{int}}$ the recommender is concentrating attention (making popular items more popular); if $G_{\text{recs}} < G_{\text{int}}$ it is diversifying. For Goodreads, interactions sit at $G=0.72$ and recommendations at $G=0.77$ — slight concentration.

key

The feedback loop is real

Fleder & Hosanagar (2009): with simulated users, recommendations grow more concentrated over time. Nguyen et al. (2014): with real users, both recommendations and interactions grow less diverse over time. Recommend popular items → they get more popular → the model recommends them even more.

03 · Post-hoc corrections

Re-ranking strategies for diversification

The simplest fix: take a relevance-ranked list and re-order it. No retraining needed.

Maximal Marginal Relevance (MMR; Carbonell & Goldstein, 1998) builds a diverse list greedily:

MMR

\text{MMR} = \arg\max_{i\in R\setminus S}\Big[\lambda\cdot\text{rel}(i) - (1-\lambda)\cdot\max_{j\in S}\text{sim}(i,j)\Big].

Start with the most relevant item; for each subsequent pick, balance high relevance $\text{rel}(i)$ against low similarity $\text{sim}(i,j)$ to the already-chosen set $S$ . The knob $\lambda\in[0,1]$ controls the trade-off: $\lambda=1$ is pure relevance, $\lambda=0$ is pure diversity. For cross-user diversity (are different items recommended to different people?), a popularity-based re-ranking weights relevance by novelty (inverse popularity), $\text{score}(i) = \text{rel}(i)\cdot \tfrac{1}{\text{popularity}(i)}$ — spreading recommendations toward less popular items, but only where relevance is high enough.

key

MMR predates recommendation

These strategies are domain-independent — they work for any ranking task. MMR was developed for text summarisation. The slide is explicit: “nothing here is specific to recommendation.”

04 · Evidence from the field

Case studies: Spotify and Facebook

Two large-scale studies on how real users respond to diversity, and how algorithms shape exposure.

Spotify — Anderson et al. (2020): not all users want diversity equally.

Specialists have consumption oriented in one direction (cosine similarity among their items near 1). They are more sensitive to personal relevance and benefit more from recommendations. Inactive users tend to be specialists.
Generalists have spread-out consumption (cosine near 0). They are more likely to churn and need diverse recommendations to stay engaged. Both groups engage more with personalised recs than with popularity alone — but the boost is larger for specialists.

Facebook — Bakshy et al. (2015): filter bubbles in political news.

Algorithmic ranking exposes users to less ideologically diverse news than their social group’s makeup would suggest.
Users interact with cross-cutting content at an even lower rate than their rate of exposure.
Individual choice plays the largest role — people self-select into homogeneous feeds. The algorithm amplifies this, but does not create it from scratch.

05 · Matching history

Calibration

Recommendations should reflect the same genre/attribute distribution as the user’s own history.

Calibration (Steck, 2018): if a user consumed 80% romance and 20% comedy, their recommendations should not be 100% romance. The idea is to match the attribute distribution (genres, actors) between the consumption and recommendation sets, measured by KL divergence:

Calibration KL

\text{calibration} = D_{\text{KL}}\big(p_{\text{hist}}(a)\;\|\;p_{\text{rec}}(a)\big).

Here $p_{\text{hist}}(a)$ is the proportion of attribute $a$ in the user’s history and $p_{\text{rec}}(a)$ its proportion in the recommended list; the goal is to minimise the divergence. In practice calibration works just like MMR — iteratively add items that balance relevance against calibration, controlled by $\lambda$ — but it operates on attribute distributions rather than item-level similarity. $\lambda=0$ is pure relevance; $\lambda=1$ matches the historical distribution closely.

06 · Beyond accuracy alone

Desirable features of a recommender system

A taxonomy of properties that go beyond “does the user click?”

Novelty

Balance discovery of new items against recommending items the user already knows (and would click anyway).

Coverage / balance

Recommended items should span a broad range of categories, matching the user’s historical distribution (calibration).

Consumer fairness

Minority groups should not receive less effective recommendations — the quality gap between groups should be small.

Provider fairness

Minority providers (creators, vendors) should not be systematically under-recommended.

Serendipity

Items should feel unexpected yet relevant — relevance + novelty + unexpectedness. Hard to define, harder to measure.

Mutual compatibility

In some domains (outfits, playlist sequencing) the goal is not individual relevance but compatibility among the recommended items.

Serendipity is slippery

The slides admit it “is not clear what serendipity is” — a blend of relevance, novelty, and unexpectedness. Measuring it usually compares recommended items to the user’s history (low cosine = serendipitous), but items a user didn’t expect to be relevant can negatively affect satisfaction (Kotkov et al., 2018). The optimal trade-off among serendipity, diversity, and novelty remains an open question.

07 · Exam intel

What the exam tests

Name the five fairness issues; explain the Gini coefficient for concentration; write the MMR re-ranking formula with $\lambda$ ; distinguish calibration (KL on attribute distributions) from MMR (item-level similarity); describe the Spotify specialist/generalist finding and the Facebook filter-bubble result; and name at least four beyond-accuracy properties.

Worked question — diagnose and correct

A recommender outputs $G_{\text{recs}}=0.85$ while consumption data has $G_{\text{int}}=0.72$ . (a) Is it concentrating or diversifying? (b) Name one re-ranking strategy to reduce concentration. (c) If you apply calibration (Steck, 2018) instead of MMR, what extra requirement does it impose?

(a) Concentrating: $G_{\text{recs}} > G_{\text{int}}$ means recommendations are more skewed toward a few items than actual consumption.
(b) MMR with $\lambda<1$ penalises similarity to already-selected items, raising intra-list diversity. A popularity penalty $\text{score}=\text{rel}/\text{pop}$ pushes toward less popular items.
(c) Calibration requires the attribute distribution of recommendations to match the user’s history (minimising KL divergence) — a stronger constraint than MMR, since it works on category proportions, not just pairwise item similarity.

Traps: confusing serendipity (unexpected + relevant + novel) with diversity (variety in a list) or novelty (unpopularity); assuming $G_{\text{recs}} < G_{\text{int}}$ is always good (over-diversifying hurts relevance); and forgetting that calibration and MMR act at different levels (attribute vs item).

08 · Exam · past papers

Past-paper questions

Past paper Practice Exam 5 · BPR popularity bias and re-ranking (5 pts)

Q. How BPR exhibits a strong popularity bias and why (2); how MMR re-ranking increases diversity (1); other re-ranking strategies based on item popularity or similarity (2).

Model answer. BPR popularity bias. BPR samples negatives uniformly. A popular item is a positive for very many users and is only rarely drawn as the sampled negative for users who like it, so the optimiser keeps pushing its score up across the population — systematically ranking already-popular items high and starving the long tail. MMR. Build the list greedily, each step maximising $\lambda\,\text{rel}(i)-(1-\lambda)\max_{j\in S}\text{sim}(i,j)$ ; penalising similarity to chosen items forces variety into the top-N. Other re-ranking. Popularity-based — divide each score by item popularity raised to a power (à la RP³β) to demote head items. Similarity-based — cap or penalise intra-list similarity (a determinantal point process, or calibration to match the user’s genre mix). All re-order an accurate candidate list to improve diversity/novelty at a small relevance cost.

09 · Self-check

Three questions before you move on

A recommender's Gini coefficient for recommendations is higher than for actual user interactions. What does this indicate?

MMR re-ranking with λ=1 and calibration (Steck, 2018) with λ=0 both produce:

In the Spotify study (Anderson et al., 2020), how do 'specialist' users differ from 'generalists'?

10 · Recap

One-screen summary

Chapter 14 — load-bearing ideas

Measure before fixing: compare consumption vs recommendation distributions. The Gini coefficient (0 = uniform, 1 = concentrated) detects concentration ( $G_{\text{recs}} > G_{\text{int}}$ ) vs diversification.
Post-hoc correction: MMR builds diverse lists greedily ( $\lambda$ trades relevance vs similarity); calibration matches attribute distributions (KL divergence); a popularity penalty $\text{score}=\text{rel}/\text{pop}$ spreads attention.
Real-world evidence: filter bubbles are real (Facebook: the algorithm amplifies self-selection); diversity preference is personal (Spotify: specialists vs generalists); serendipity is desirable but slippery.
Beyond accuracy: novelty, coverage/calibration, consumer and provider fairness, serendipity, mutual compatibility.