21 Recommendation Systems Interview Questions and Answers

• A Recommendation System is a subclass of information filtering system that seeks to predict the rating or preference a user would give to an item.
• Recommender systems usually make use of either or both collaborative filtering and content-based filtering, as well as other systems such as knowledge-based systems.

• The recommendations of content-based and collaborative systems are primarily based on historical data.
• The recommendations of knowledge-based systems are based on the direct specifications by users of what they want.

• An important distinguishing characteristic of knowledge-based systems is the high level of customization to the specific domain. This customization is achieved through the use of a knowledge-base that encodes relevant domain knowledge in the form of either constraints or similarity metrics.

• In different domains, such as temporal data, location-based data, and social data, the context of the recommendation plays a critical role. The notion of contextual recommender systems was developed to address the additional side information that arises in these domains.
• This notion is used with different modifications for various types of data.
• Some domain-specific applications are context-based recommender systems, time-sensitive recommender systems, location-based recommender systems, and social recommender systems.

The various components of component-based systems are as follows:

1. Preprocessing and feature extraction: Content-based systems are used in a wide variety of domains, such as Web pages, product descriptions, news, music features, and so on. In most cases, features are extracted from these various sources to convert them into a keyword-based vector-space representation. This is the first step of any content-based recommendation system, and it is highly domain-specific. However, the proper extraction of the most informative features is essential for the effective functioning of any content-based recommender system.

2. Content-based learning of user profiles: A content-based model is specific to a given user. Therefore, a user-specific model is constructed to predict user interests in items, based on their history of either buying or rating items. To achieve this goal, user feedback is leveraged, which may be manifested in the form of previously specified ratings (explicit feedback) or user activity (implicit feedback). Such feedbacks are used in conjunction with the attributes of the items to construct the training data. A learning model is constructed on this training data. The resulting model is referred to as the user profile because it conceptually relates user interests to item attributes.

3. Filtering and recommendation: In this step, the learned model from the previous steps is used to make recommendations on items for specific users. This step needs to be very efficient because the predictions need to be performed in real-time.

• Model-based Collaborative approaches only rely on user-item interactions information and assume a latent model to explain these interactions.
• For example, matrix factorization algorithms consist of decomposing the huge and sparse user-item interaction matrix into a product of two smaller and dense matrices: a user-factor matrix (containing users representations) that multiplies a factor-item matrix (containing items representations).

• Collaborative methods for recommender systems are methods that are based solely on the past interactions recorded between users and items to produce new recommendations.
• Unlike collaborative methods, content-based approaches use additional information about users and/or items.

• A real-world example of a company using collaborative filtering is Last.fm. They would create a 'station' of recommended songs by observing what bands and individual tracks the user has listened to on a regular basis and comparing those against the listening behavior of other users. This approach leverages the behavior of users.
• Pandora uses the content-based approach of recommender systems. It uses the properties of a song or artist to see a 'station' that plays music with similar properties. User feedback is used to refine the station's results, deemphasizing certain attributes when a user 'dislikes' a particular song and emphasizing other attributes when a user 'likes' a song.

When choosing between the implementation of a user-based and an item-based neighbourhood recommender system, the following criteria should be considered: Accuracy

• The accuracy of neighbourhood recommender methods depends mostly on the ratio between the number of users and items in the system. The similarity between two users in user-based methods is normally obtained by comparing the ratings made by these users on the same items.
• An item-based method usually computes the similarity between two items by comparing ratings made by the same user on these items. A small number of high-confidence neighbours is by far preferable to a large number of neighbours for which the similarity weights are not trustable. In cases where the number of users is much greater than the number of items, such as large commercial systems like Amazon, item-based methods can therefore produce more accurate recommendations.

Efficiency

• The memory and computational efficiency of recommender systems also depends on the ratio between the number of users and items.
• When the number of users exceeds the number of items, as is most often the case, item-based recommendation approaches require much less memory and time to compute the similarity weights than user-based ones, making them more scalable.
• The time complexity of the online recommendation phase, which depends only on the number of available items and the maximum number of neighbours, is the same for user-based and item-based methods.

Two classes of techniques are used to preserve privacy:

1. Privacy at data collection time: In these techniques, the data collection approach is modified so that individual ratings are not collected. Rather, distributed protocols or perturbation techniques are used to collect the data only in a perturbed way or in the aggregate. The advantage of such an approach is that users are assured that no single entity has access to their private data, at least in the exact form.
2. Privacy at data publication time: In most practical settings, a trusted entity has access to all the rating data it has collected over time. In such cases, the trusted entity might wish to publish the data to the broader technical community to enable further advancements in the field of collaborative filter, so in such cases, models like k-anonymity are used to preserve privacy. Typically, such methods use group-based anonymization techniques in which records belonging to groups of a minimum size become indistinguishable. This is achieved by carefully perturbing selected attributes of the data records so that one cannot join such records with publicly available information to exactly identify the subjects of the data records. Such systems are more common and have wider applicability than the first scenario.

• The Naive Bayes model is a generative model, which is commonly used for classification.
• One can treat the items as features and users as instances to infer the missing entries with a classification model. The main challenge in using this approach for collaborative filtering is that any feature (item) can be the target class in collaborative filtering, and one also has to work with incomplete feature variables. These differences can be handled with minor modifications to the basic methodology of the naive Bayes model.
• Using the naive Bayes model approach computes the conditional probability of a rating, based on the ratings of other items. Therefore, this approach is an item-based Bayes approach.

Recommender systems based on neighborhood automate the common principle that similar users prefer similar items, and similar items are preferred by similar users. User-based Classification falls under this category.

• Neighborhood-based classification finds the most likely rating given by a user u to an item i, by having the nearest-neighbours of u vote on this value. The vote $v_{ir}$ given by the k-NN of u for the rating $r \in S$ can be obtained as the sum of the similarity weights of neighbours that have given this rating to i:
  $v_{ir} = \sum_{v \in N_i(u)} \delta(r_{vi} = r)w_{uv}$
  where $\delta (r_{vi} = r)$ is 1 if $r_{vi} = r$, and 0 otherwise. Once this has been computed for every possible rating value, the predicted rating is simply the value of r for which $v_{ir}$ is the greatest.

• For text inputs, many content-based recommender systems use simple retrieval models, such as keyword matching or the Vector Space Model (VSM). VSM is a spatial representation of text documents. In that model, each document is represented by a vector in an n dimensional space, where each dimension corresponds to a term from the overall vocabulary of a given document collection.
• The most commonly used term weighting scheme is TF-IDF (Term Frequency-Inverse Document Frequency) weighing.
• A similarity measure is required to determine the closeness between two documents. Many similarity measures have been derived to describe the proximity of two vectors; among those measures, cosine similarity is the most widely used:

• In content-based recommender systems relying on VSM, both user-profiles and items are represented as weighted term vectors. Predictions of a user's interest in a particular item can be derived by computing the cosine similarity.

• User independence: Content-based recommenders exploit solely ratings provided by the active user to build their profile. Instead, collaborative filtering methods need ratings from other users to find the nearest neighbours of the active user. Then, only the items that are most liked by the neighbours of the active user will be recommended.
• Transparency: Explanations on how the recommender system works can be provided by explicitly listing content features or descriptions that caused an item to occur in the list of recommendations. Those features are indicators to consult to decide whether to trust a recommendation. Conversely, collaborative systems are black boxes since the only explanation for an item recommendation is that unknown users with similar tastes liked that item.
• New item: Content-based recommenders are capable of recommending items not yet rated by any user. As a consequence, they do not suffer from the first-rater problem, which affects collaborative recommenders which rely solely on users’ preferences to make recommendations.

The main advantages of neighborhood-based methods are:

• Simplicity: Neighborhood-based methods are intuitive and relatively simple to implement. In their simplest form, only one parameter requires tuning.
• Justifiability: Such methods also provide a concise and intuitive justification for the computed predictions.
• Efficiency: One of the strong points of neighborhood-based systems are their efficiency. Unlike most model-based systems, they require no costly training phases, which need to be carried at frequent intervals in large commercial applications.
• Stability: Another useful property of recommender systems based on this approach is that they are less affected by the constant addition of users, items, and ratings, which are typically observed in large commercial applications. For instance, once item similarities have been computed, an item-based system can readily make recommendations to new users, without having to re-train the system.

The data about user likes and dislikes can take on any of the following forms:

1. Ratings: In this case, users specify ratings indicating their preference for the item. Ratings can be binary, interval-based, or ordinal. The nature of the ratings has a significant impact on the model used for learning the user profiles.
2. Implicit feedback: Implicit feedback refers to user actions, such as buying or browsing an item. In most cases, only the positive preferences of a user are captured with implicit feedback but not negative preferences.
3. Text opinions: In many cases, users may express their opinions in the form of text descriptions. In such cases, implicit ratings can be extracted from these opinions. This form of rating extraction is related to the field of opinion mining and sentiment analysis.
4. Cases: Users might specify examples of items that they are interested in.

The main characteristics of user-user and item-item approaches it that they use only information from the user-item interaction matrix and they assume no model to produce new recommendations.

User-user

• To make a new recommendation to a user, user-user method roughly tries to identify users with the most similar interactions profile (nearest neighbors) to suggest items that are the most popular among these neighbors.
• The process of the user-user method is shown in the figure below:

Item-item

• To make a new recommendation to a user, the idea of the item-item method is to find items similar to the ones the user already “positively” interacted with. Two items are considered to be similar if most of the users that have interacted with both of them did it similarly. This method is said to be “item-centered” as it represents items based on interactions users had with them and evaluates distances between those items.
• The process of item-item method is shown in the figure below:

The recommendation process of a Content-Based Recommender System is performed in three steps:

• Content Analyzer: When information has no structure, some kind of pre-processing step is needed to extract structured relevant information. The main responsibility of the component is to represent the content of items coming from information sources in a form suitable for the next processing steps. Data items are analyzed by feature extraction techniques to shift item representation from the original information space to the target one.
• Profile Learner: This module collects data representative of the user preferences and tries to generalize this data, to construct the user profile. Usually, the generalization strategy is realized through machine learning techniques, which can infer a model of user interests starting from items liked or disliked in the past.
• Filtering Component: This module exploits the user profile to suggest relevant items by matching the profile representation against that of items to be recommended. The result is a binary or continuous relevance judgment, the latter case resulting in a ranked list of potentially interesting items.

• Recommender systems are the ones that try to recommend items to users, or predict how they might rate an item.
• Adaptation is defined as when a system adapts its behaviour to individual users based on information acquired about its users and its environment. A user-adaptive system is one that continuously or recurrently personalizes its behaviour based on interaction with a user.
• Personalization is defined as customization, consisting of tailoring a service or a product to accommodate specific individuals, as opposed to general groups. A personalization system creates a product for a particular person.

• Multi-armed bandit methods are particularly useful for computational advertising.
• It is noteworthy that multi-armed bandit methods are particularly useful in settings where; (a) new items enter the system all the time, and (b) the payoffs of selecting a particular strategy can be precisely computed.
• Computational advertising in a domain in which the items are extremely transient and therefore it is particularly important to use exploration and exploitation simultaneously.
• In many cases, the setting of computational advertising does not neatly fit into the traditional multi-armed bandit framework. For example, a publisher might present more than one advertisement at a time on a page, and a user might click on more than one advertisement presented to them. To handle this variation, the slate problem is proposed for multi-armed bandits.
• As a solution for the slate problem in computational advertising, different payoffs are associated with different placements of the advertisements on the Web page. For example, a higher placement in the ranked list will have a higher expected payoff than a lower placement.