Sequence-Aware Recommender Systems

Characterization. Adopting the formalisms of [3], we can describe the problem at a more formal, abstract level as follows. Let C be a set of users and I a set of recommendable items. In contrast to matrix-completion problems, we are not interested in predicting a utility value for each i ∈ I and for each c ∈ C , but in computing an ordered list of objects L of length k for each user, where each element of l ∈ L corresponds to an element of i ∈ I . Technically, each sequence L is an element of the set of all permutations up to length k of the powerset of I , i.e., L ∈ Sk (P(I )). We denote this latter set of possible lists as L∗. Letu be a function that returns a utility score of a given sequence L for a user c , i.e.,u : C×L∗ → R. The sequence-aware recommendation problem then consists of determining the sequence l ′ c ∈ L∗ that maximizes the score for the user, i.e., ∀c ∈ C, l ′ c = argmax l ∈L∗ u(c, l) (1) The main problem in recommender systems is to learn or extrapolate the utility function u from some given data. In the matrix completion problem, which underlies the work in [3], the input is a sparse matrix of user-item ratings. In sequence-aware recommender systems, we in contrast assume that the underlying data is a dataset D consisting of sequence1 of user actions where each user action A ∈ D has a number of attributes. A sequence dataset D can be considered as an enriched log of actions of a user community, where the attributes of each action A includes some sort of user ID2 and additional optional attributes like the action type (e.g., an item view or click event) or a timestamp. Overall, our function u is not limited to characterizing utility scores for individual items, but for entire ordered lists of items. This makes it possible to consider additional aspects of utility in sequence-aware recommendation problems, including the diversity of the set as a whole, the quality of the ordering itself, e.g., in terms of transitions between objects, or the degree of fulfillment of weak or strict order constraints in L. How these quality factors are considered within existing algorithms will be discussed later in this work in Section 4. Generally, the design of the utility function u depends on the specific type of value the recommender system should provide to the user, or its purpose in the sense of [50]. In the literature on recommender systems, researchers often do not explicitly discuss the underlying purpose of the system, which could be information filtering but also discovery support. Instead, they focus on optimizing an abstract computational task like predicting a hidden rating. The situation in the 1A sequence, as usual, is considered as an ordered set of objects. 2In practice, the user ID can either refer to a known user or it is created from a cookie in an ongoing user session. ACM Computing Surveys, Vol. 1, No. 1, Article 1. Publication date: February 2018. Sequence-Aware Recommender Systems 1:5 context of sequence-aware recommenders is often similar with the difference that the computational task is mainly to predict the hidden elements of a session given the session beginning. While the performance of an algorithm that predicts the next hidden user action can be assessed with standard measures from information retrieval (such as precision and recall), in other cases specific measures (e.g., diversity metrics) are required in the evaluation process. 2.2 Relation to Other Areas Implicit-Feedback Recommender Systems. Our characterization of the sequence-aware recommendation problem mainly targets scenarios in which we observe the individual and collective behavior of a user community over time instead of asking for explicit item ratings. A number of research works exist that focus on implicit user feedback like purchase events. The problem formulation is however often based again on matrix completion, where multiple interactions of one user with an item are not taken into account. Explicit item ratings, on the other hand, can also be taken into account in a sequence-aware recommender as one of several types of user actions. One potential problem in that context however is that the point in time when users provide a rating can be quite different from the point in time when they consumed or purchased an item (e.g., when registering for a movie recommendation service, users initially rate a bunch of movies they have watched in the past). The sequence and timestamp of the ratings might therefore mislead a sequence-aware recommender. Context-Aware and Time-Aware Recommender Systems. In some of the application scenarios discussed in the next sections, sequence-aware recommender systems represent a special form of context-aware recommender systems. In session-based recommendation, the users’ short-term intents, which can be estimated from their very last actions, can represent an important piece of context information to be taken into account when recommending [56]. Time-aware recommender systems (TARS) usually consider time information that is associated with past user actions to adapt the recommendations accordingly, see [17] for an overview. TARS share a number of commonalities with sequence-aware recommenders, e.g., in terms of how we can compare different approaches in offline settings. The focus of sequence-aware recommenders is however often less on the exact point of time of the past user interactions, but on the sequential order of the events. Furthermore, a number of proposals on time-aware recommenders mainly rely on the matrix completion problem setting when modeling temporal dynamics [68]. Research in Other Related Fields. Some aspects of sequence-aware recommender systems were finally explored also in neighboring fields. Examples are the problem of query suggestion in the field of information retrieval or the problem of interest drift in the more general field of user modeling. In this paper, we concentrate on works where the recommendation problem itself is the main focus, in contrast to works that, e.g., aim to develop methods to capture changes in the user preferences over time. When searching for papers to consider in our survey, we therefore used a corresponding search string and selection strategy when we queried a digital library, as will be described in more detail in Section 3.5. 3 A CATEGORIZATION OF SEQUENCE-AWARE RECOMMENDATION TASKS We identified four main goals in the academic literature that can be achieved with the help of sequence-aware recommender systems in different application scenarios: (1) Context Adaptation (2) Trend Detection (3) Repeated Recommendation ACM Computing Surveys, Vol. 1, No. 1, Article 1. Publication date: February 2018. 1:6 M. Quadrana et al. (4) Consideration of Order Constraints and Sequential Patterns We will discuss these four categories in more detail in the next sections and will then also look at typical application domains for sequence-aware recommenders. Note that all types of problem settings discussed next are based on the same formal problem characterization described in Equation 1, but require specific algorithmic approaches that use the sequence information in the input datasets (see Section 4). The problems are also not mutually exclusive and multiple aspects (e.g., trends and repetitions) can be considered in parallel, as was done, for example, in [62] for the e-commerce domain. 3.1 Context Adaptation In many domains, the relevance of a recommendable item not only depends on the users’ general preferences, but also on their current situation and their short-term intents and interests. Contextaware recommenders take such additional types of information into account. Typical contextual factors in the literature include the user’s geographical position, the current weather, or the time of the day [4]. Context factors like these are examples of what is called the representational context [28], which is defined by a predefined set of “observable” context variables. Contextual factors like the user’s current shopping intent in an e-commerce setting or their current mood are however not directly observable. These types of information, which represent what is called the user’s interactional context, therefore have to be derived from the users most recent actions and eventually on behavioral patterns of the user and the community as a whole [40, 87]. Considering interactional context factors is particularly important for systems where there are many new or anonymous users. Since no historical data is available about their past preferences, it is important to make full use of interactional context information, as representational context information can only help to partition anonymous users into coarse-grained categories, without any real personalization [32]. Overall, understanding the users’ situation and goals and making context-adapted recommendations from past interaction data represents a main goal of sequence-aware recommender systems. Categorization based on importance of longand short-term interactions. Depending on the availability of historical data for individual users and the importance of focusing on the most recent interactions, we can differentiate between the following types of context-adaptation situations. • Last-N interactions based recommendation: In these scenarios, only the last N user actions are considered. A typical problem setting is that of predicting the next location (or check-in) in a location-aware recommender system [21, 72, 76]. The reason to limit oneself to the last actions could be that not many past interactions of that type exist or that the other previous actions of the same type (e.g., check-in events) are not predictive for the next action. • Session-based recommendation: In this problem scenario only the last sequence of actions of a user is known and this sequence of actions is limited to a session, i.e., a limited period of time when the user interacted with the site. Typical application examples include news recommendation [32], e-commerce, video and classified advertisement recommendation [46]. • Session-aware recommendation: Finally, there are situations in which we have knowledge both about the users’ actions in the last session and about their past behavior. This type of problem setting occurs if we have returning customers that can be identified. In this situation, a sequence-aware recommender system can be based on a combination of longterm and short-term interest models, e.g., in e-commerce settings or for app recommendation [9, 40, 56, 91]. Note that our problem definition in Equation 1 covers all three scenarios, i.e., the output is a ranked list of items. In the case of a session-aware adaptation problem, the underlying sequence dataset D ACM Computing Surveys, Vol. 1, No. 1, Article 1. Publication date: February 2018. Sequence-Aware Recommender Systems 1:7 is however usually split into two components, where one that contains the older interactions is used for building a long-term model, and the other is used to consider short-term user intents. How the different models are learned or combined then depends on the specific algorithmic approach that is used to maximize the utility function u, which might for example return higher scores when recommendations are a mix of familiar and novel items for the user. On utility functions and recommendation purposes. Most of the papers on context-adaptation problems in the literature do not make explicit statements about the characteristics of the utility function in the application scenarios they considered. As mentioned above, they in most cases implicitly define the goal through the evaluation procedure and aim to predict hidden elements of a given user session. Thereby, they implicitly assume that this next action is in some sense the best recommendation for the given purpose. The task of recommenders in context-adaptation scenarios can most often be characterized as “find matching items” for a given session beginning, without any further explicit specification of what represents a good recommendation. In some works – and in practical environments – a number of more specific purposes can be identified, see also [50]. The task of a recommender can be, for example, to create a list of alternatives for the currently inspected items (similar items). In other applications, in contrast, the task can be to determine complements, e.g., accessories to a main shopping item in e-commerce. In yet other application domains the recommendations should represent suitable or logical continuations of either the current session (e.g., next-track music recommendations) or the user’s longer term behavior (e.g., next-basket recommendations). Finally, we can differentiate if the user is assumed to pick one of the recommendations (e.g., one alternative in e-commerce scenarios), or consider all of them together (e.g., playlist recommendation for audio and video streaming). This latter scenario was recently addressed in [101] for the news domain. In their work, the authors model the user’s expected utility of an item during the course of session and try to diversify the recommended content within a session accordingly.