2b:null 2c:["$","title",null,{"children":"Recommender systems | Sandraviz | Observable"}] 31:T5c3,## Introduction Recommender systems (RS) influence, the videos we watch, the books we read, the music we hear, the games we play, the investments we make, the friends we meet, the food we eat, the news we monitor, the products we buy, the software we code, the photos we share, the events we attend, the people we date, the jobs we apply for, in the end they practically and actually impact increasingly **the life we live.** [

](https://www.wired.co.uk/article/algorithm-decision-making) - Which of these situations are familiar to you and which ones seem prompt? - How did we use to take decision in these situations before using algorithms? - In which of the described situations using "algorithms" might lead to a huge improvement? - In which ones on the other hand you sense some uselessness? - And in which situations do you get the impression it might become unethical or dangerous for some humans? - What is your overall takeaway? **Definition** As seen above RS can mean better choice and therefore empower people. In the first and final analysis RS generate advice, which is personalised, contextualised and customised. - Which one is your favourite? Slightly more formally, RS are functions that take information about a user’s preferences as an input, and output a prediction about the rating that a user would give for the items under evaluation.32:T3fb5,The purpose of RS is suggesting relevant items to users. To achieve this task, the algorithms transform data intro relevant recommendations by finding, calculating and ranking the most interesting correlations and co-occurrences for users. It is important to note, that apart from seeking similarities, novelty is equally important. This can be a challenging trade-off. ## Data Independent from the approach, we choose three different data sources are relevant: users, items and the interactions. **Users** The recommendations are rated, ranked and optimised for them. User data is all about profiles, preferences and personalisation. User features, characteristics and choices must be captured, sorted and quantified. **Items** They are the objects or experiences that recommenders recommend. Item data describe qualities, features and attributes that make those items distinctive or desirable. How might items be classified, categorised, labeled, or tagged to enhance reliable recommendation? **Interactions** They happen whenever users meet items. The perceived value or utility of those interactions determine what gets recommended and why. They are either explicit or implicit. Explicit interactions reflect user decision to rate, rank or review items, e.g. Facebook's thumbs-up or Amazon's stars review. On the other hand implicit *"You are a very long row of numbers in a very, very large table. This row describes everything you've looked at, clicked at and purchased on the site. The rest of the table represents the millions of other Amazon shoppers. Your row changes every time you enter the site, and it changes again with every action you take. That information in turn affects what you see on each page you visit and what special offers you receive from the company."* Joseph Konstan and John Riedl ## Methods Popularity is recommendation's simplest success story, it is the transcendental similarity. Moreover it is the easiest and most common way to deal with the "cold start" recommender problem. **Finding similarities** Similarities explore, exploit and explain the data driven dynamics of "people like you" and "people who bought this also bought that recommendations. Similarities can be sliced, diced and digitally defined along radically different dimensions. Quantifying similarities makes ever more personalised recommendations possible. On the other hand surprise matters enormously. Variety of similarity is, indeed, the spice of recommendation life. The paradoxical genius of similarity is the mathematical promise of predictable surprise also known as serendipity. At their core RS are nothing but similarity hunters, with the calculated likelihood of surprise turns them into serendipity hunters. Successful recommender algorithm artfully navigate between continuity and novelty. Two major categories of methods are used: Collaborative filtering and Content based methods, and their hybrids.

These methods live in the utility matrix, the enormous tables discussed above are matrixes that hold the affinities and preferences, the data- defined relationships, between users and items. Each row and column contains essential data about the relevant characteristics and attributes of users and items - alike users are rows and items are columns. Each user/item intersection reveals the degree of user preference for that item. ### Collaborative Filtering

The core idea assumption is that people with similar tastes will rate things similarly, therefore they actually work without any specific knowledge about the item itself. They are based solely on the past interactions recorded between users and items in order to produce new recommendations. The assumption is that people who agreed in the past will agree in the future, and that they will like similar kinds of items as they liked in the past. Then, the main idea that rules collaborative methods is that these past user-item interactions are sufficient to detect similar users and/or similar items and make predictions based on these estimated proximities. The class of collaborative filtering algorithms is divided into two sub-categories that are generally called memory based and model based approaches. #### Memory based technique Memory based approaches directly work with values of recorded interactions, assuming no model, and are essentially based on nearest neighbours search (for example, find the closest users from a user of interest and suggest the most popular items among these neighbours). In searching for a neighbour who shares recognition with him, the things which were already valued by the consumer before play an appropriate role. When a neighbour of user is found, various algorithms are employed to merge neighbours' preferences to produce recommendations. Widespread popularity was gained in real world applications due to the efficacy of these techniques. As no latent model is assumed, these methods have theoretically a low bias but a high variance. #### Model based technique Model based approaches assume an underlying “generative” model that explains the user-item interactions and try to discover it in order to make new predictions. To enhance the efficiency in Collaborative filtering, this method uses the previous rankings for learning a model. Using the techniques like machine learning or data mining methods, the process of building the model can be completed. In fact by making use of pre-computed model, really can easily suggest a collection of items and in fact have been proven to deliver recommendation results which are close to neighbourhood-based methods of recommendation. **Association rule** These algorithms bring out the rules which predict the presence of an object based upon the existence of other objects in any transaction. **Clustering** In various fields, clustering approaches have been used, namely pattern recognition, digital image processing, analytical data and information discovery. To identify meaningful classes which exists among them, the clustering algorithm aims to divide a collection of data based on the similarity features into a number of sub-clusters. **Decision tree** The method is modeled as per the tree graph technique that is developed by evaluating a collection of training samples for which we define the labels for the class. Then, there are applied to identify examples previously not seen. If training is done on a very good quality of data, then accurate predictions can be expected. **Regression** This technique is employed when a linear relationship is assumed to consistently link two or more variables. It is a powerful process of diversity to evaluate associative relationships among a dependent variable as well as one or many independent variables. Regression uses curve fitting, estimation, and a very systematic hypotheses among relationships between evaluation of variables. **Artificial Neural-network** This is a composition of several linked neurons (nodes) that are organised in systemic ways in layers. Depending on the quantity of influence that one neuron has with another neuron, the connections among neurons will have weights linked with them. In some special problem cases, the use of neural networks has some benefits. For example, an ANN is very much robust against noisy as well as data sets which are erroneous since it contains several neurons, also it allocates weight to each every connection. **Link Analysis** This is a method of creating interconnected object networks in order to investigate trends and patterns. It has created tremendous potential for improving the efficiency in web searching. Link analysis includes algorithms for Page Ranking as well as HITS. A web page is treated by most of the link analysis algorithms as an individual node in the graph of web. As a model for user-item interactions is assumed, this methods has theoretically a higher bias but a lower variance than methods assuming no latent model. #### Advantages - It’s simple. The collaborative filtering algorithm only needs one input: the items other people liked. It doesn’t need any data at all about the items themselves. - It uses the knowledge implied in histories. Each history is the outcome of a string of decisions made by a person. - This is why a collaborative filtering algorithm can give recommendations that are both surprising and accurate. #### Challenges - Updating "user neighbourhoods can be computationally expensive and inefficient. - Collaborative filtering cannot give predictions to items that have never been rated by any user. - The model will help users identify new interests. The ML framework does not know whether the user is really interested in a given item in isolation, but the model may still suggest it because that item is of interest to similar users. To some degree, a matrix factorisation model needs only the feedback matrix to be trained by the method. - For a given (user, item) pair, the model prediction is the dot product of the corresponding embeddings. So, if an object is not seen during training, it cannot be embedded by the system and can not question the model for this item. The cold-start issue is also called this problem. ### Content-Based Filtering

Unlike collaborative methods that only rely on the user-item interactions, content based approaches use additional information about users and/or items. They rely on the assumption that items with similar properties and features will be rated similarly. Determining which features and attributes are most predictive for a given user is the challenge. If we consider the example of a movies RS, this additional information can be, for example, the age, the sex, the job or any other personal information for users as well as the category, the main actors, the duration or other characteristics for the movies (items). Then, the idea of content based methods is to try to build a model, based on the available “features”, that explain the observed user-item interactions. Loosely speaking content-based recommendation means to predict whether a user likes a certain product by using the user's online history. That includes, among others, likes the user gave (e.g. on Facebook), keywords he/she searched for (e.g. on Google), and simply clicks and visits he/she made to certain websites. All in all, it focuses on the user’s own preferences. We can for example think of a simple binary classifier (or regressor) that outputs a click-through rate (or rating) for a certain ad-group for this user. **Keyword matching** A straightforward way of matching users and items is keyword matching. For example, for job recommendations, one can match a job description to job seekers’ resumes. Term frequency-inverse document frequency is often used to put more weights to keywords that are unique for an item or user. **Examples** At Pandora, a team of musicians labeled each music with more than 400 attributes. Then when a user selects a music station, songs that match the station’s attributes will be added to the playlist. This is content-based recommendation. Users or items have profiles describing their characteristics and the system would recommend an item to a user if the two profiles match. For Pandora, manual efforts/costs are needed to create music attributes, but there are many cases without such a need. Stitch Fix’s customers provide their own preferences, Linkedin users provide their own working experiences and skills, merchants on Amazon provide information about their product items, all are freely usable for content-based recommendations. As for model based collaborative methods, a model is assumed. However, this model is more constrained (because representation of users and/or items are given) and, so, the method tends to have the highest bias but the lowest variance. #### Advantages - computationally fast and interpretable. - suffer far less from the cold start problem than collaborative approaches: new users or items can be described by their characteristics (content) and so relevant suggestions can be done for these new entities. Only new users or items with previously unseen features will logically suffer from this drawback, but once the system is old enough, this has few to no chance to happen. #### Challenges - If the data base is constantly and rapidly changing: For example readers of the news tend to read about recent happenings. Recommending the most important articles news and fresh news in place of old news articles is therefore a major challenge. - RS for the presentation of news articles written in only one language, from a single one source are typically developed. Recommendation of multi-lingual articles related to news from different outlets and deploying productive news articles using similarity methods are a difficult job for proper recommendation. ### Hybrid techniques Hybrid Filtering: In order to get better device optimisation, the hybrid filtering technique incorporates various techniques of recommendation to avoid certain disadvantages and problems faced by some pure or clear recommendation systems. The concept behind hybrid techniques is that algorithms are combined to provide recommendations that are more accurate and efficient when compared to that of a single algorithm, as one algorithm can resolve the limitations of other algorithm. In a combined model, using several recommendation techniques will suppress the limitations of a single technique. In any of the following ways, the combination of methods are done: separate implementation of the algorithm and outcome combination, using few content-based filtering methods in collaborative mode, or using few collaborative filtering methods in content-based technique by creating a single recommendation framework that brings all approaches together. **Weighted hybrid** This hybrid combines scores from each component using linear formula. Therefore, components must be able to produce their recommendation score which can be linearly combinable. Also, the components have to be consistent relative accuracy across the product space and to perform uniformly. **Switching hybrid** The issue of this hybrid is selecting one recommender among candidates. This selection is made according to the situation it is experiencing. The criterion for the selection like confidence value or external criteria should exist and the components might have different performances with different situations. **Mixed hybrid** This is a hybrid that is based on the merging and presentation of multiple ranked lists into one. Each component of this hybrid should be able to produce recommendation lists with ranks and the core algorithm of mixed hybrid merges them into a single ranked list. The issue here is how the new rank scores should be produced. **Feature combination hybrid** There exist two very different recommendation components for this hybrid, contributing and actual recommender. The actual recommender works with data modified by the contributing one. The contributing one injects features of one source to the source of the other component. **Feature augmentation hybrid** This is similar to the feature combination hybrids but different in that the contributor generates new features. It is more flexible and adds a smaller dimension than the feature combination method. **Cascade hybrid** This one is a tie-breaker. The secondary recommender is just a tiebreaker and does refinements. **Meta-level hybrid** For this one, contributing and actual recommenders exist but the former one completely replaces the data for the latter one, not just part of it. #### Challenges When the sparsity increases, more number of items are introduced to the scheme, this issue will be more prevalent in the news domain, for instance, in a continuation fashion, latest news when published online and the number of news items increases rapidly. 33:Te6f,## Use Case: Netflix Netflix is synonymous to most people in this day and age as the go-to streaming service for movies and tv shows What most people do not know, however, is that Netflix started out in the late 1990s with a subscription-based model, posting DVDs to people’s homes in the US. The website makes recommendations by comparing the watching and searching habits of similar users (i.e., collaborative filtering) as well as by offering movies that share characteristics with films that a user has rated highly (content-based filtering). ### The Netflix Prize In 2000, Netflix introduced personalised movie recommendations and in 2006, launched Netflix Prize, a machine learning and data mining competition with a $1 million dollar prize money. Back then, Netflix used Cinematch, its proprietary RS which had a root mean squared error (RMSE) of 0.9525 and challenged people to beat this benchmark by 10%. The team who could achieve the target or got close to this target after a year would be awarded the prize money. The winner of the Progress Prize a year later in 2007 used a linear combination of Matrix Factorisation (a.k.a. SVD) and Restricted Boltzmann Machines (RBM), achieving a RMSE of 0.88. Netflix then put those algorithms into production after some adaptations to the source code. What is worth noting is that despite some teams achieving a RMSE of 0.8567 in 2009, the company did not put those algorithms into production due to the engineering effort required to gain the marginal increase in accuracy. This serves an important point in real-life RS — that there is always a positive relationship between model improvements and engineering efforts. #### Privacy issues A number of privacy issues arose around the dataset offered by Netflix for the Netflix Prize competition. Although the data sets were anonymised in order to preserve customer privacy, in 2007 two researchers from the University of Texas were able to identify individual users by matching the data sets with film ratings on the Internet Movie Database. As a result, in December 2009, an anonymous Netflix user sued Netflix in Doe v. Netflix, alleging that Netflix had violated United States fair trade laws and the Video Privacy Protection Act by releasing the datasets. This, as well as concerns from the Federal Trade Commission, led to the cancellation of a second Netflix Prize competition in 2010 ### Streaming A more important reason why Netflix did not incorporate the improved models from the Netflix Prize is because it introduced streaming in 2007. With streaming, the amount of data it has surged dramatically. It has to change the way its RS was generating recommendations and ingesting data. Fast forward to 2020, Netflix has transformed from a mail service posting DVDs in the US to a global streaming service with 182.8 million subscribers. Consequently, its RS transformed from a regression problem predicting ratings to a ranking problem, to a page-generation problem, to a problem maximising user experience (defined as maximising number of hours streamed i.e. personalising everything that can be personalised). ### Netflix as a Business Netflix has a subscription-based model. Simply put, the more users Netflix has, the higher its revenue. Revenue can be seen as a function of three things: - Acquisition rate of new users - Cancellation rates - Rate at which former members rejoin 80% of stream time is achieved through Netflix’s RS, which is a highly impressive number. Moreover, Netflix believes in creating a user experience that will seek to improve retention rate, which in turn translates to savings on customer acquisition (estimated $1B per year as of 2016).34:T712,## Evaluation of RS It is important to understand what makes a good RS. The main questions are - How well do the RS work, do people follow the advices? - Do RS leave people feel more empowered or disappointed and dependent? In the Netflix prize the answer is Root Mean Squared Error (RMSE), but as we know accuracy is not everything in the real world. Users just don’t care how accurately we can predict the rated movies they see, rather they care about the new things that they love to see. The same is true when recommending star wars movie to someone one watched one of them. What matters as well is novelty and diversity. To make it operational one needs to specify, among other things, three parameters: - a) What is the space of options? - b) What counts as a good recommendation? - c) How can RS performance be evaluated? Specifying these parameter choices is highly dependent on the domain of application and the level of abstraction. Typically, the literature implements three LoAs: catalogue-based, decision support, and multi-stakeholder environment. - In e-commerce applications, the space of options may be the items in the catalogue, while a good recommendation may be specified as one which ultimately results in a purchase. To evaluate the system performance, one may compare the RS’s predictions to the actual user behaviour after a recommendation is made. - In the domain of news recommendations, a good recommendation may be defined as a news item that is relevant to the user, and one may use click-through rates as a proxy to evaluate the accuracy of the system’s recommendations. Similar RS are designed to develop a model of individual users and to use it to predict the users’ feedback on the system’s recommendation, which is essentially a prediction problem. 35:Ted9,## Hands-on Let's discuss some of the typical use cases and how they are presented in media. ### Choose an example Below you find recent articles explaining different RS applications. #### [It’s 2021, why are dating app algorithms still so bad?](https://www.wired.co.uk/article/dating-app-algorithms) But when it comes to dating, these algorithms are doing a terrible job. A 2016 study of match rates across all sexual orientations on Tinder run by Queen Mary University of London revealed jaw-droppingly low rates of reciprocal interest: the match rate was 10.5 per cent for women and just 0.6 per cent for men. #### [Facebook partially documents its content recommendation system](https://techcrunch.com/2020/08/31/facebook-partially-documents-its-content-recommendation-system/?guccounter=1&guce_referrer=aHR0cHM6Ly9kdWNrZHVja2dvLmNvbS8&guce_referrer_sig=AQAAAGEP3s9sXltqIvbuHtw3k44JFmDLoFHBjFqCRguFnw96YtrVLG3PLF3vHc3d-Q-t6ESz5THLt2_GZm3AGYf3f9woQfzy0NBx-rns6zHJqQ3-aPPdXmB2DaBEU07-3CjagDgMhXWtMENjT9G77-81ElFRi9nLSvg_uPmsI-2TK4Va) But when it comes to dating, these algorithms are doing a terrible job. A 2016 study of match rates across all sexual orientations on Tinder run by Queen Mary University of London revealed jaw-droppingly low rates of reciprocal interest: the match rate was 10.5 per cent for women and just 0.6 per cent for men. #### [How Spotify’s Algorithm Manages To Find Your Inner Groove](https://analyticsindiamag.com/how-spotifys-algorithm-manages-to-find-your-inner-groove/) Have you ever wondered how Spotify manages to recommend you that perfect song, playlist or even that ‘daily mix’? The answer is simple — data. Spotify’s algorithm is always finding new ways to understand the kind of music one listens to — from the songs that are always on repeat to the favourite genre that one can’t let go. #### [TikTok Has a Pro-Anorexia Problem](https://www.wired.com/story/opinion-tiktok-has-a-pro-anorexia-problem/) “PRO-ANA” COMMUNITIES—WEBSITES, BLOGS, forums, and social media spaces dedicated to promoting the worsening of eating disorders like anorexia—have been a fixture of the web more or less since its inception. So it’s no surprise that, as Buzzfeed reported last month, some TikTok users have found disturbing pro-ana content on their For You page, a personalized section of the platform that displays videos users are likely to enjoy. #### [The Tinder algorithm, explained](https://www.vox.com/2019/2/7/18210998/tinder-algorithm-swiping-tips-dating-app-science) If there’s one thing I know about love, it’s that people who don’t find it have shorter life spans on average. Which means learning how the Tinder algorithm works is a matter of life and death, extrapolating slightly. #### [The Real Story Behind Netflix’s Top 10 Lists](https://www.vulture.com/article/netflix-top-ten-most-popular-explained.html) Streaming TELEVISION guaranteed to change house entertainment by releasing audiences from the tyranny of old-fashioned tv schedules. You might enjoy whatever you fancied, at whatever time you wanted. Amazing! But now it appears the BBC is attempting to put the genie back in the bottle, by means of complex speculative software application that will make it much easier for the iPlayer to “recommend” brand-new material. Forget free choice – and invite back to the bosom the Beeb Hive Mind. ### Discussion Now think first a moment by yourself about the following questions and then discuss them with your group. - What is the situation/problem? - What knowledge do you gain about the RS? - Which critic is mentioned and if none what are your reflections? - Is the "problem" in society (e.g. biases) or created by the technology itself (e.g. filter bubble)? - Any ideas how to address it? 36:T49d4,## Ethical challenges The main questions are: - Who is responsible for good or manipulative advice? - Should RS be liable for bad outcome? - Should people be punished for ignoring good advice? Research into the ethical issues posed by RS is still in its infancy. The debate is also fragmented across different scientific communities, as it tends to focus on specific aspects and applications of these systems in a variety of contexts. The current fragmentation of the debate may be due to two main factors: - the relative newness of the technology, which took off with the spread of internet-based services and the introduction of collaborative filtering techniques in the 1990s - the proprietary and privacy issues involved in the design and deployment of this class of algorithms. The details of RS currently in operation are treated as highly guarded industrial secrets. This makes it difficult for independent researchers to access information about their internal operations, and hence provide based assessment. In the same vein, due to privacy concerns, providers of recommendation systems may be reluctant to share information that could compromise their users’ personal data . ### How to map ethical challenges In order to identify what is ethically at stake in the design and deployment of a RS, let us start with a formal taxonomy. This is how we propose to design it. The question about which moral principles may be correct is deeply contentious and debated in philosophy, we do not have to take a side because all we need is a distinction about which there is a general consensus: there are at least two classes of variables that are morally relevant, *actions and consequences*. The value of some consequences is often measured in terms of the utility they contain. So, it is reasonable to assume that any aspect of a RS that could impact negatively the utility of any of its stakeholders, or risk imposing such negative impacts, constitutes a feature that is ethically relevant. Thinking in terms of actions and consequences, we can identify two ways in which a RS can have ethical impacts. First, its operations can - impact(negatively) and/or - violate their rights. Second, these two kinds of ethical impact may be leading to a decrease in utility for the user or they may expose the relevant parties to future risks. In the case of RS, for example, the risks may involve exposing users to undue privacy violations by external actors, or the exposure to potentially irrelevant or damaging content. Given the previous analysis, we may now categorise the ethical issues caused by recommender systems along two dimensions - whether a RS negatively impacts the utility of some of its stakeholders or, instead, constitutes a rights violation, which is not necessarily measured in terms of utility - whether the negative impact constitutes an immediate harm or it exposes the relevant party to future risk of harm or rights violation.

- Which examples come to your mind, where recommendations may harm their users? ### Ethical content In the case of e-commerce applications one can suggests that there are five ethically problematic areas: - the practices of user profiling, - data publishing, - algorithm design, - user interface design, - online experimentation or A/B testing, i.e. the practice of exposing selected groups of users to modifications of the algorithm, with the aim of gathering feedback on the effectiveness of each version from the user responses. Moreover, users’ choice of parameters can reveal sensitive information about the users themselves. For example, adding a filter to exclude some kind of content gives away the information that the user may find this content distressing, irrelevant, or in other ways unacceptable. **User-centred solutions** But above all, the main problem is that, although user-centred solutions may foster the transparency of recommender systems, they also shift the responsibility and accountability for the protection of rights and utility to the users. These points highlight how user-centred solutions in general are challenged by their demanding nature, as they may constitute a mere shift in responsibility when the users are only nominally empowered but actually unable to manage all the procedures needed to protect their interests. This may therefore be an unfair shift, since it places undue burdens on the users, and is in any case problematic because the effectiveness of these solutions varies with the level of awareness and expertise of the users themselves, which may lead to users experiencing different levels of protection depending on their ability to control the technology. **Ethical filter** Implementing an “ethical filter” for a recommender system, would also be controversial in some applications, for example if it were used by a government to limit citizens’ ability to access some politically sensitive contents. As for the eudaimonic approach, this goes in the direction of designing a recommender system that is an optimal decision support, yet it seems practically unfeasible, and at least much more research would be needed. Figuring out what is a “good human life” is something that millennia of reflection have not yet solved. #### Privacy - What issues come into your mind? - Are you concerned when you use any personalised RS (Youtube, Spotify, Facebook etc.)? User privacy is one of the primary challenges for RS. This may be seen as inevitable, given that a majority of the most commercially successful RS are based on hybrid or collaborative filtering techniques, and work by constructing models of their users in order to generate personalised recommendations. Privacy risks occur in different stages. - They can arise when data are collected or shared without the user’s explicit consent. - Once data sets are stored, there is the further risk that they may be leaked to external agents, or become subject to deanonymization attempts (example: Netflix price) At both stages, privacy breaches expose users to risks, which may result in loss of utility (for example, if individual users are targeted by malicious agents as a result), or in rights violations (for example, if users’ private information is utilised in ways that threaten their individual autonomy. - Independently of how securely data are collected and stored, privacy concerns also arise at the stage of inferences that the system can draw from the data. Users may not be aware of the nature of these inferences, and they may object to this use of their personal data if they were better informed. - Privacy risks do not only concern data collection because, for example, an external agent observing the recommendation that the system generates for a given user may be able to infer some sensitive information about the user. #### Autonomy RS can encroach on individual users’ autonomy, by providing recommendations that nudge users in a particular direction, by attempting to “addict” them to some types of contents, or by limiting the range of options to which they are exposed. These interventions can range from being benign (enabling individual agency and supporting better decision making by filtering out irrelevant options), to potentially questionable (persuasion, nudging), to possibly malign (being manipulative and coercive). Algorithmic classification used to construct user models on the basis of aggregate user data can reproduce social categories. Our experience of personal identity is mediated by the categories to which we are assigned. Algorithmic profiling, performed by RS, can disrupt this individual experience of personal identity, for at least two main reasons. First, the RS’s model of each user is continuously reconfigured on the basis of the feedback provided by other users’ interactions with the system. In this sense, the system should not be conceptualised as tracking a pre-established user identity and tailoring its recommendations to it, but rather as contributing to the construction of the user identity dynamically. Second, the labelling that the system uses to categorise users may not correspond to recognisable attributes or social categories with which the user would self-identify (for example, because machine-generated categories may not correspond to any known social representation), so even if users could access the content of the model, they would not be able to interpret it and connect it with their lived experiences in a meaningful way. These features of RS create an environment where personalisation comes at the cost of removing the user from the social categories that help mediate their experiences of identity. In this context, an interesting take on the issue of personal autonomy in relation to RS comes from the “captology” of RS. From an anthropological perspective: As recommenders spread across online cultural infrastructures and become practically inescapable, thinking with traps offers an alternative to common ethical framings that oppose tropes of freedom and coercion. RS appear to function as “sticky traps” insofar as they are trying to “glue” their users to some specific solutions. This is reflected in “captivation metrics” (i.e. that measure user retention), which are commonly used by popular RS. A prominent example is YouTube’s recommendation algorithm, which received much attention recently for its tendency to promote biased content and “fake news”, in a bid to keep users engaged with its platform. Regarding RS as traps requires engaging with the minds of the users: traps can only be effective if their creators understand and work with the target’s world view and motivations, so the autonomous agency of the target is not negated, but effectively exploited. Given this captological approach, and given the effectiveness and ubiquity of the traps of RS, the question to ask is not how users can escape from them, but rather how users can make the traps work for them. #### Opacity In theory, explaining how personalised recommendations are generated for individual users could help to mitigate the risk of encroaching on their autonomy, giving them access to the reasons why the system “thinks” that some options are relevant to them. It would also help increase the transparency of the algorithmic decisions concerning how to class and model users, thus helping to guard against bias. **Explanations for RS** Designing and evaluating explanations for recommender systems can take different forms, depending on the specific applications. As reported, several studies have pursued a user-centred approach to evaluation metrics, including metrics to evaluate explanations of recommendations. What counts as a good explanation depends on several criteria: the purpose of the recommendation for the user; whether the explanation accurately matches the mechanism by which the recommendation is generated; whether it improves the system’s transparency and scrutability; and whether it helps the user to make decisions more efficiently (e.g. more quickly), and more effectively, e.g. in terms of increased satisfaction. These criteria are satisfied by factual explanations.5 However, factual explanations are notoriously difficult to achieve. As noted recommendations generated by collaborative filtering techniques can, on a simple level, be conceptualised as analogous to “word of mouth” recommendations among users. However, offline word of mouth recommendations can work on the basis of trust and shared personal experience, whereas in the case of recommender systems users do not have access to the identity of the other users, nor do they have access to the models that the system uses in order to generate the recommendations. As we mentioned, this is an issue in so far as it diminishes the user’s autonomy. It may be difficult to provide good factual explanations in practice also for computational reasons (the required computation to generate a good explanation may be too complex), and because they may have distorting effects on the accuracy of the recommendations. For example, explaining to a user that a certain item is recommended because it is the most popular with other users may increase the item’s desirability, thus generating a self-reinforcing pattern where the item will be recommended more often because it is popular. This, in turn, reinforces its popularity, ending in a winner-takes-all scenario that, depending on the intended domain of application, can have negative effects on the variety of options, plurality of choices, and the emergence of competition. Arguably, this may be one of the reasons why Amazon does not automatically privilege products with less than perfect scoring but that have been rated by a large number of reviewers. #### Fairness Fairness in algorithmic decision making is a wide-ranging issue, made more complicated by the existence of multiple notions of fairness, which are not all mutually compatible. In the context of RS, several articles identified in this review address the issue of recommendations that may reproduce social biases. They may be synthesised around two approaches. On the one hand, several possible sources for unfairness in collaborative filtering have been identified, they note that these methods assume that the missing ratings (i.e., the ones that the system needs to infer from the statistical data to predict a user’s preferences) are randomly distributed. However, this assumption of randomness introduces a potential source of bias in the system’s predictions, because it is well documented that users’ underlying preferences often differ from the sampled ratings, since the latter are affected by social factors, which may be biased, also identify the two primary sources of bias in RS with two problematic patterns of data collection, namely observation bias, which results from feedback loops generated by the system’s recommendations to specific groups of users, and population imbalance, where the data available to the system reflect existing social patterns expressing bias towards some groups. They propose a probabilistic programming approach to mitigate the system’s bias against protected social groups. On the other hand, it is suggested to consider fairness in recommendation systems as a multi-sided concept. Based on this approach, he focuses on three notions of fair recommendations, taking the perspective of either the user/consumer (C-fairness); or the provider (P- fairness); or a combination of the two (CP-Fairness). This taxonomy enables the developer of a recommendation system to identify how the competing interests of different parties are affected by the system’s recommendations, and hence design system architectures that can mediate effectively between these interests. In both approaches, the issue of fairness is tied up with choosing the right LoA for a specific application of a recommender system. Given that the concept of fairness is strongly tied to the social context within which the system gathers its data and makes recommendations, extending the same approach to any application of recommender systems may not be viable. #### Polarisation A much-discussed effect of some recommender systems is their transformative impact on society. In particular, news recommender systems and social media filters, by nature of their design, run the risk of insulating users from exposure to different viewpoints, creating self-reinforcing biases and “filter bubbles” that are damaging to the normal functioning of public debate, group deliberation, and democratic institutions more generally. A closely related issue is protecting these systems from manipulation by (sometimes even small but) especially active groups of users, whose interactions with the system can generate intense positive feedback, driving up the system’s rate of recommendations for specific items. News recommendation systems, streaming platforms, and social networks can become an arena for targeted political propaganda, as demonstrated by the recent Cambridge Analytical scandal in 2018, and the documented external interference in US political elections in recent years. The literature on the topic proposes a range of approaches to increase the diversity of recommendations. A point noted by several authors is that news recommendation systems, in particular, must reach a trade-off between the expected relevance to the user and diversity when generating personalised recommendations based on pre-specified user preferences or behavioural data argue that the design of algorithmic tools to combat informational segregation should be more sensitive to the democratic norms that are implicitly built into these tools. In general, the approaches to the issue of polarisation and social manipulability appear to be split between bottom-up and top-down strategies, prioritising either the preferences of users (and their autonomy in deciding how to configure the personalised recommendations) or the social preference for a balanced public arena. Once again, some authors take a decidedly user-centred perspective. For example propose the use of different “recommendation personae”, or “pre-configured and anthropomorphised types of recommendation algorithms” expressing different user preferences with respect to novelty, diversity, relevance, and other attributes of a recommendation algorithm. In the same vein stresses the importance of promoting serendipity even at the cost of sacrificing aspects of the user experience, such as diminished relevance of the recommendations. Filter bubbles (detailed mainly by Eli Praiser) limit user exposure to conflicting, contracting and/or challenging viewpoints. This can cause intellectual isolationism and promotes malign political and social polarisation.

#### Trust abuse The most important ingredient for a RS is the trust that humans place on them. Users confident that RS respect their best interests and are not afraid to take a chance, they make themselves vulnerable. Can this trust be abused, is one of the key ethical question. If the interest of the user and the company providing the RS are not 100% overlapping is this a case of abuse? Technically manipulating RS to favor one brand over another is not hard and the greater the influence of RS on people the greater the temptation to manipulate them. - Do you have situations in your mind where this might happen? - How would you address the conflict of interests? - What would speak against abusing RS from a business point of view?37:T7a2,## Appendix ### Web links - [Introduction to recommender systems](https://towardsdatascience.com/introduction-to-recommender-systems-6c66cf15ada) - [RecSyS Conference](https://recsys.acm.org/) - [The Netflix Recommender System: Algorithms, Business Value and Innovation](https://dl.acm.org/doi/pdf/10.1145/2843948) ### Books

[Recommendation Engines](https://mitpress.mit.edu/books/recommendation-engines) Michael Schrage 2020 ISBN: 978-0262539074

[Machine Learning, revised](https://mitpress.mit.edu/books/machine-learning-revised-and-updated-edition) Ethem Alpaydin 2021 ISBN-13: 978-0262542524

[Hands-On Recommendation Systems with Python](https://github.com/PacktPublishing/Hands-On-Recommendation-Systems-with-Python) Rounak Banik 2018 ISBN: 978-1788993753

[Data Science from Scratch](https://github.com/joelgrus/data-science-from-scratch) Joel Grus 2019 (2nd edition) ISBN-13: 978-1491901427 ### Research papers [Streaming Platform and Strategic Recommendation Bias](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3338744). Marc Bourreau. Telecom ParisTech. Germain Gaudin. University of Freiburg - College of Economics and Behavioral Sciences. 2018 [Recommender Systems and their ethical challenges](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3378581). Silvia Milano, Mariarosaria Taddeo, Luciano Floridi University of Oxford - Oxford Internet Institute 2019 38:T558,39:[{"id":"a6458dd7630be19bd8777a1f0a4cac5d29817049e3f6cab2246551fa959c8205f80eccb3ef698cd1a4e3c347598d59910e6ca4e99acda5ef086fc4e4226655f1","href":"https://static.observableusercontent.com/files/a6458dd7630be19bd8777a1f0a4cac5d29817049e3f6cab2246551fa959c8205f80eccb3ef698cd1a4e3c347598d59910e6ca4e99acda5ef086fc4e4226655f1","name":"ml.jpeg","type":"image/jpeg","lastModified":1633288814594.06,"size":68602},{"id":"b1f01dc9fc8528538e248303a4fcd66ae8ebf9d989282ebd7caaa85513b4106e342cd80ced0c1bf73973710ef6994c88e29ed85af1f1946eb416285af9a11dcd","href":"https://static.observableusercontent.com/files/b1f01dc9fc8528538e248303a4fcd66ae8ebf9d989282ebd7caaa85513b4106e342cd80ced0c1bf73973710ef6994c88e29ed85af1f1946eb416285af9a11dcd","name":"re.jpeg","type":"image/jpeg","lastModified":1633288852146.626,"size":34283},{"id":"540b9df87768d6eccb2ba42bb7b1000a08fe5827119e35bb50fc1399612b3ff7cd5dc5db143bd1a5e40702a57666dceb542490a160e048ef1000e060a584cd2d","href":"https://static.observableusercontent.com/files/540b9df87768d6eccb2ba42bb7b1000a08fe5827119e35bb50fc1399612b3ff7cd5dc5db143bd1a5e40702a57666dceb542490a160e048ef1000e060a584cd2d","name":"wired.jpg","type":"image/jpeg","lastModified":1633292190626.309,"size":289324},{"id":"80ae6cf8ccb082c09784430b22545e429c1b31cf547e492da71a91b33d68448c1dbaf0c3a1693eab7bc44a148facfc95e78acd017053b69b80d9d3bc9cbcc38b","href":"https://static.observableusercontent.com/files/80ae6cf8ccb082c09784430b22545e429c1b31cf547e492da71a91b33d68448c1dbaf0c3a1693eab7bc44a148facfc95e78acd017053b69b80d9d3bc9cbcc38b","name":"types.png","type":"image/png","lastModified":1633295362607.162,"size":36477},{"id":"bf8a795499b3d1da0b11319552780bf6e6145c5db78a295bacba521fd8f7940d86dc0290e2274c02bb9fb2f87bfac5ec6211c79fa64f0a3f01f86d2ead8380bb","href":"https://static.observableusercontent.com/files/bf8a795499b3d1da0b11319552780bf6e6145c5db78a295bacba521fd8f7940d86dc0290e2274c02bb9fb2f87bfac5ec6211c79fa64f0a3f01f86d2ead8380bb","name":"data science .jpeg","type":"image/jpeg","lastModified":1633884228051.568,"size":716254},{"id":"d006a232ad523c7631e389e68ce8511038953c426b548491be5491168a80b1ddf6b7f7c8cdd997b263f6d8967d95ccecc65fdf20a03ccfc5584de3fbe42eb024","href":"https://static.observableusercontent.com/files/d006a232ad523c7631e389e68ce8511038953c426b548491be5491168a80b1ddf6b7f7c8cdd997b263f6d8967d95ccecc65fdf20a03ccfc5584de3fbe42eb024","name":"recommender-1.jpg","type":"image/jpeg","lastModified":1664799986713.935,"size":24102},{"id":"5f0f521ff0bf3c938b20ee3b76a675b3203089ee5c52a27b80afda3efeae9f7daa8dac9dd4602589a9713d725c9b8581d77318d82d9085df29728cd375376dfd","href":"https://static.observableusercontent.com/files/5f0f521ff0bf3c938b20ee3b76a675b3203089ee5c52a27b80afda3efeae9f7daa8dac9dd4602589a9713d725c9b8581d77318d82d9085df29728cd375376dfd","name":"PYTHON.png","type":"image/png","lastModified":1633293811725.782,"size":58981},{"id":"1d6cdecdf91d3114a12f92ceb7caf09e54e12d5c24d8b1bc8b2c34589b5ecf65f7d5eadf8cc9e4e9ddfbf910353227ace33ee36b8f4e0f884d4e1696413a3e91","href":"https://static.observableusercontent.com/files/1d6cdecdf91d3114a12f92ceb7caf09e54e12d5c24d8b1bc8b2c34589b5ecf65f7d5eadf8cc9e4e9ddfbf910353227ace33ee36b8f4e0f884d4e1696413a3e91","name":"rs2.png","type":"image/png","lastModified":1633335115855.495,"size":51004},{"id":"b1fef4a7bbcb50ad1a3e63a15630ca304002e386ba08261f21a371ae184d2a0e0dac5c0cfa467c9ad45f8b0badc20722edf3f7c1b235eceb0df004629577996b","href":"https://static.observableusercontent.com/files/b1fef4a7bbcb50ad1a3e63a15630ca304002e386ba08261f21a371ae184d2a0e0dac5c0cfa467c9ad45f8b0badc20722edf3f7c1b235eceb0df004629577996b","name":"rs3.png","type":"image/png","lastModified":1633335115167.337,"size":56528},{"id":"a9824bcf3f064ca15829d5485adadf03ef47c4d14edfd502d2c90a37fd9cae89b10649303cd1fbf84d534afdef24ef1f11af162725fadad9a5e0d13fd189fbd6","href":"https://static.observableusercontent.com/files/a9824bcf3f064ca15829d5485adadf03ef47c4d14edfd502d2c90a37fd9cae89b10649303cd1fbf84d534afdef24ef1f11af162725fadad9a5e0d13fd189fbd6","name":"tab.jpg","type":"image/jpeg","lastModified":1633421168482.759,"size":104335},{"id":"22973d043ccd95f283572a5d70f87320e1599f3cca727b0e8f02689b2a4e8177f7711b16993e641a5bb11af65f2927c16416b5754e0024067b45a256e5013e66","href":"https://static.observableusercontent.com/files/22973d043ccd95f283572a5d70f87320e1599f3cca727b0e8f02689b2a4e8177f7711b16993e641a5bb11af65f2927c16416b5754e0024067b45a256e5013e66","name":"tab2.jpg","type":"image/jpeg","lastModified":1633421773022.811,"size":169821},{"id":"1028cc5be762a425d9f301efab736e48891386f2c634c9125aa09b50a259000c1958f7b26a642efeb929f76af301944fd01e6a369a59c085a6fa08c33d684836","href":"https://static.observableusercontent.com/files/1028cc5be762a425d9f301efab736e48891386f2c634c9125aa09b50a259000c1958f7b26a642efeb929f76af301944fd01e6a369a59c085a6fa08c33d684836","name":"recommender 3.jpg","type":"image/jpeg","lastModified":1663764192698.229,"size":272720}] 2f:{"id":"f3fe1705621dea2b","slug":"recommender-systems","title":"Recommender systems","thumbnail":"f79b5171b6b4e58ed56ef849773bf017e8ffd89ed7c9d2c2bec81cfd41afdf6d","access_level":"public","create_time":"$D2021-09-01T13:38:46.359Z","update_time":"$D2023-11-28T12:24:10.249Z","view_time":null,"first_public_time":"$D2021-09-01T14:02:46.720Z","fork":{"id":"ebe57cfb3b5ab8f2","slug":"decision-support-systems","title":"Decision support systems","owner":{"id":"6661932e6ac85aef","login":"sandraviz","name":"Sandraviz"}},"stars":31,"starred":false,"owner":{"type":"team","id":"6661932e6ac85aef","login":"sandraviz","name":"Sandraviz","avatar_url":"https://avatars-next.observableusercontent.com/e6a2480f89a40dcfbe75f19d5081ae2296848971c4b934ed0499fd3b7f1973eb","tier":"starter_2024"},"version":1874,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"3d1beb3a90fbfe98","login":"sandraviz","name":"Sandraviz","avatar_url":"https://avatars-next.observableusercontent.com/9b19f1a074469dcaafa9bd24c7b08f4bc969889b551fe70c209497f1d03c6514","tier":null}],"members":[{"type":"individual","id":"3d1beb3a90fbfe98","login":"sandraviz","name":"Sandraviz","avatar_url":"https://avatars-next.observableusercontent.com/9b19f1a074469dcaafa9bd24c7b08f4bc969889b551fe70c209497f1d03c6514","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"subscribed":null,"forks":14,"collection":{"id":"eab5d593efbf8501","slug":"ai","title":"AI ","owner":{"id":"6661932e6ac85aef","login":"sandraviz","name":"Sandraviz"}},"type":"notebook","body":{"title":"Recommender systems","cells":[{"id":0,"value":"# Recommender systems\n\nIn 2021 and 2022 I taught remotely at the **University Arcada (Finland)** responsible and ethical uses of machine learning practice, which is a fundamental precondition to trustworthy development of AI. Recommendation systems, which simultaneously shape and predict the future in nearly all parts of human life, was the main use case during the course. You can find all the material and links in this [Github Repository](https://github.com/sandravizz/Decision-support-systems-verification-Python).\n\n**Bluntly, technology is taking over advice** - Michael Schrage\n\n","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1247,"value":"

","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":181,"value":"Index = toc(\"h2, h3, h4, h5\")","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined"},{"id":1194,"value":"$31","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1422,"value":"$32","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1242,"value":"$33","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1561,"value":"$34","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1682,"value":"$35","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1660,"value":"$36","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1499,"value":"## Examples in python\n\n

\n\n\nYou can find a collection of notebook examples under this [link](https://github.com/sandravizz/Decision-support-systems-verification/tree/main/Recommender%20System)","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1837,"value":"## Assignment\n\nIn this assignment I ask you to create a simple recommender as explained in today’s lecture based on the Amazon Top 50 books dataset (find attached). \n\n**First Part**\n\nEDA, visualise the most important variables and comment on your observations. \n\n**Second Part**\n\nRecommender, follow the 4 steps and make your own decisions/rules \n\n- Metric choice\n- Prerequisites\n- Calculate score\n- List in decrease order ","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":1454,"value":"$37","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined"},{"id":179,"value":"import {toc} from \"@nebrius/indented-toc\"","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined"},{"id":1184,"value":"$38","pinned":false,"hidden":false,"mode":"html","author":"$undefined","summary":"$undefined"}],"files":"$W39","stdlib":"1","titling":"auto"}} 30:"$2f:body"