Past Accuracy: Embracing Serendipity and Novelty in Suggestions for Lengthy Time period Person Retention | by Christabelle Pabalan | Jun, 2023


An examination of the elements that contribute to a superb suggestion and long-term consumer retention

Composition by the creator utilizing DALL-E

A Bond of Belief Fashioned at a Espresso Store

You’re sitting in a espresso store, savoring your favourite espresso variation (a cappuccino, in fact) and engrossed in dialog with a pal. Because the dialog flows, the subject shifts to the newest gripping TV collection that you simply each have been hooked on. The shared pleasure creates a bond of belief, to the extent that your pal eagerly turns to you and asks, “What ought to I watch subsequent? Do you have got any suggestions?”

At that second, you change into the curator of their leisure expertise. You are feeling a way of accountability to protect their belief and supply recommendations which are assured to captivate them. Moreover, you’re excited on the alternative to, maybe, introduce them to a barely new style or storyline they hadn’t explored earlier than.

However what elements affect your decision-making course of as you take into account the right suggestions to your pal?

Photograph by Thibault Penin on Unsplash

First, you faucet into your understanding of your pal’s tastes and pursuits. You recall their fondness for intricate plot twists and darkish humor; moreover, they loved crime dramas like “Sherlock” and psychological thrillers like “Black Mirror.” Armed with this data, you navigate your psychological library of TV exhibits.

To play it protected?

You’re tempted to recommend an inventory of exhibits which are virtually similar, with slight variations, to the one you had simply been raving over, which embody each crime and thrill. You additionally take into consideration how others with comparable tastes have loved these exhibits to slim your decisions. In any case, they’re virtually assured to get pleasure from this set; it’s the protected and straightforward alternative. Nevertheless, you take into account that relying solely on their previous favorites could restrict their publicity to new and numerous content material and don’t significantly wish to depend on the protected and straightforward alternative.

You recall a current sci-fi collection that ingeniously blends thriller, journey, and supernatural intrigue. Though it deviates from their typical style, you are feeling assured it would present a refreshing and charming change of narrative.

The Lengthy Tail Drawback, Suggestions Loop & Filter Bubbles

Suggestion techniques intention to copy this course of on a bigger scale. By analyzing huge quantities of knowledge about people’ preferences, behaviors, and previous experiences, these techniques attempt to generate personalised suggestions that embody the complexity of human decision-making.

Nevertheless, historically, suggestion techniques have centered primarily — if not, solely — on taking part in it protected and counting on the suggestions which are assured to fulfill (a minimum of, within the quick time period).

A method they do that is by prioritizing fashionable or mainstream content material. Because of this, this fashionable content material receives extra publicity and interactions (recognition bias), making a suggestions loop that reinforces its prominence. Sadly, this typically leaves lesser-known or area of interest content material struggling to achieve visibility and attain the meant viewers (the lengthy tail downside).

Photograph by Writer

In truth, there was loads of literature in the previous few years that look at “equity” in suggestion techniques. For instance, in “Fairness in Music Recommender Systems: A Stakeholder-Centered Mini Review”, Karlijn Dinnissen and Christine Bauer discover the problem of equity in music recommender techniques; they analyze gender equity and recognition bias from the attitude of a number of stakeholders e.g. the impression of recognition bias on the illustration of artists.

Within the article, “Equity in Query: Do Music Suggestion Algorithms Worth Range?”, Julie Knibbe shares:

As a former product director at a streaming platform, I typically obtain questions like “do streaming companies select to advertise fashionable artists over indies and area of interest music?” Intuitively, folks assume that on these huge platforms “the wealthy get richer.”

Afterward within the article, Knibbe additionally echoes the sentiment of Dinnissen and Bauer:

“Within the context of music suggestion […] equity is commonly outlined by way of publicity or consideration. Streaming companies are additionally a two-sided market, which implies that “neutral and simply therapy” should apply to each streaming companies’ customers and artists.

Each sources spotlight the twin nature of equity in recommender techniques, underscoring the significance of contemplating “neutral and simply therapy” for customers and content material creators.

What does the best final result appear to be?

Naturally, there exists an inherent imbalance within the distribution of content material. A part of what makes the human expertise wealthy lies inside its community intricacy; some content material resonates with a broader viewers, whereas others forge connections inside area of interest teams, creating a way of richness and personalization. The target is to not artificially promote much less fashionable content material for the sake of it, striving for a uniform distribution. Slightly, our intention is to floor area of interest content material to people who genuinely relate and may admire the content material creator’s work, thereby minimizing missed alternatives for significant connections.

What does the business say about this?

In 2020, the analysis group at Spotify launched an article titled, “Algorithmic Effects on the Diversity of Consumption on Spotify.” Of their analysis, they examined the connection between listening variety and consumer outcomes.

Photograph by Fath on Unsplash

They aimed to reply the questions: “How does variety relate to necessary consumer outcomes? Are customers who hear diversely roughly happy than those that hear narrowly?”

The researchers found that “customers with numerous listening are between 10–20 share factors much less more likely to churn than these with much less numerous listening […] listening variety is related to consumer conversion and retention.

Moreover, based on Julie Knibbe:

“TikTok’s suggestion algorithm was lately talked about among the many high 10 […] by MIT know-how evaluate. What’s modern of their method isn’t the algorithm itself — it’s the metrics they’re optimizing for, weighing in additional on variety than different elements.”

Subsequently, there’s a connection between the attribute of discoverability inside a platform and consumer retention. In different phrases, when suggestions change into predictable, customers may search different platforms that supply a larger sense of “freshness” in content material, permitting them to flee the confines of filter bubbles.

So how can suggestion techniques emulate the thoughtfulness and instinct that you simply employed in curating the right suggestion to your pal?

Properly, within the article, “Range, Serendipity, Novelty, and Protection: A Survey and Empirical Evaluation of Past-Accuracy Goals in Recommender Techniques”, authors Marius Kaminskas and Derek Bridge spotlight:

“Analysis into recommender techniques has historically centered on accuracy […] nonetheless, it has been acknowledged that different suggestion qualities — similar to whether or not the record of suggestions is numerous and whether or not it accommodates novel objects — could have a major impression on the general high quality of a recommender system. Consequently […] the main focus of recommender techniques analysis has shifted to incorporate a wider vary of ‘past accuracy’ aims”


Sifting via the literature in an try to know what ‘variety’ is in recommender techniques was brutal, as every article offered its personal distinctive definition. Range may be measured each on the particular person degree or on the international degree. We’ll go over 3 ways to conceptualize variety, within the context of giving present suggestions to a pal.

Prediction Range

Prediction variety refers back to the measure of how diversified the suggestions are inside a given set. If you recommend a set of exhibits to your pal, prediction variety assesses the extent to which the suggestions differ from each other by way of genres, themes, or different related elements.

A better prediction variety signifies a wider vary of choices throughout the advisable set, providing your pal a extra numerous and probably enriching viewing expertise.

A method that is measured is through the use of intra-list-diversity (ILD), which is the typical pairwise dissimilarity among the many advisable objects. Given the advisable merchandise record, the ILD is outlined as follows:

Person Range

Person variety, within the context of offering present suggestions to a pal, examines the typical variety of all of the suggestions you have got ever given to that particular pal. It considers the breadth and number of content material instructed to them over time, capturing the vary of genres, themes, or different related elements coated.

You too can assess consumer variety by analyzing the typical dissimilarity between the merchandise embeddings inside every set of suggestions per pal.

World Range

Then again, international variety seems past a selected pal and assesses the typical variety of all of the suggestions you have got given to any pal.

Typically, that is known as congestion — a mirrored image of advice uniformity or the crowding of suggestions.

A few metrics that you should use to research international variety embody the Gini index and entropy.

The Gini index, tailored from the sphere of earnings inequality measurement, can be utilized to evaluate the equity and steadiness of advice distributions in suggestion techniques. A decrease Gini index signifies a extra equitable distribution, the place suggestions are unfold evenly, selling larger variety and publicity to a wider vary of content material. Then again, a better Gini index suggests a focus of suggestions on just a few fashionable objects, probably limiting the visibility of area of interest content material and lowering variety within the suggestions.

Entropy is a measure of the quantity of knowledge contained within the suggestion course of. It quantifies the extent of uncertainty or randomness within the distribution of suggestions. Just like the Gini index, optimum entropy is attained when the advice distribution is uniform, which means that every merchandise has an equal chance of being advisable. This means a balanced and numerous set of suggestions. Increased entropy suggests a extra diversified and unpredictable suggestion system, whereas decrease entropy signifies a extra concentrated and predictable set of suggestions.


Protection is outlined because the portion/proportion of attainable suggestions the algorithm can produce. In different phrases, how properly the suggestions cowl the catalog of accessible objects.

For instance, let’s take into account a music streaming platform with an enormous library of songs spanning numerous genres, artists, and many years. The protection of the advice algorithm would point out how successfully it might probably cowl everything of this music catalog when suggesting songs to customers.

Drawback: This metric treats an merchandise advisable as soon as as the identical as an merchandise that was advisable hundreds of occasions


Novelty is a metric used to gauge the extent of newness or originality in advisable objects. It encompasses two points: user-dependent and user-independent novelty. Person-dependent novelty measures how totally different or unfamiliar the suggestions are to the consumer, indicating the presence of contemporary and unexplored content material. Nevertheless, it has change into more and more frequent to discuss with the novelty of an merchandise in a user-independent manner.

To estimate novelty, one frequent method is to contemplate an merchandise’s recognition, measured as Merchandise Rarity. This method inversely relates an merchandise’s novelty to its recognition, recognizing that much less fashionable objects are sometimes perceived as extra novel as a consequence of their deviation from mainstream or widely-known decisions. By integrating this attitude, novelty metrics present insights into the extent of innovation and variety current within the advisable objects, contributing to a extra enriching and exploratory suggestion expertise.

Unexpectedness (Shock)

Shock in suggestion techniques measures the extent of unexpectedness within the advisable objects based mostly on a consumer’s historic interactions. One solution to quantify shock is by calculating the cosine similarity between the advisable objects and the consumer’s previous interactions. A better similarity signifies much less shock, whereas a decrease similarity signifies larger shock within the suggestions.


Discoverability in suggestion techniques may be understood because the consumer’s skill to simply come throughout and discover the suggestions instructed by the mannequin. It’s akin to how seen and accessible the suggestions are throughout the consumer interface or platform.

It’s quantified utilizing a lowering rank low cost operate, which assigns increased significance to suggestions on the high ranks of the advice record and progressively decreases their weight because the rank place goes down.


Serendipity in suggestion techniques encompasses two key points: unexpectedness and relevance.

Serendipity refers back to the prevalence of nice surprises or the invention of attention-grabbing and surprising suggestions. To quantify serendipity, it’s calculated on a per-user and per-item foundation utilizing the method:

By multiplying unexpectedness and relevance, the serendipity metric combines the weather of nice shock and suitability. It quantifies the diploma to which a suggestion is each surprising and related, offering a measure of serendipitous experiences within the suggestion course of.

Total serendipity averaged throughout customers and advisable objects may be computed as:

Because the business evolves, there’s a rising emphasis on refining suggestion algorithms to ship suggestions that embody everything of consumer preferences, together with richer personalization, serendipity, and novelty. Furthermore, suggestion techniques that optimize the steadiness between these dimensions have additionally been related to improved consumer retention metrics and consumer expertise. Finally, the purpose is to create suggestion techniques that not solely cater to customers’ recognized preferences but additionally shock and delight them with contemporary, numerous, and personally related suggestions, fostering long-term engagement and satisfaction.

  1. Diversity, Serendipity, Novelty, and Coverage: A Survey and Empirical Analysis of Beyond-Accuracy Objectives in Recommender Systems
  2. Post Processing Recommender Systems for Diversity
  3. Diversity in recommender systems — A survey
  4. Avoiding congestion in recommender systems
  5. The Definition of Novelty in Recommendation System
  6. Novelty and Diversity in Recommender Systems: an Information Retrieval approach for evaluation and improvement
  7. Quantifying Availability and Discovery in Recommender Systems via Stochastic Reachability
  8. A new system-wide diversity measure for recommendations with efficient algorithms
  9. Automatic Evaluation of Recommendation Systems: Coverage, Novelty and Diversity
  10. Serendipity: Accuracy’s Unpopular Best Friend in Recommenders

From Python to Julia: Function Engineering and ML | by Wang Shenghao | Jun, 2023

Ludwig — A “Friendlier” Deep Studying Framework | by John Adeojo | Jun, 2023