Research on serendipity has become more active in the last few years, but there is still no universal definition of this concept within the field of recommender systems. While the general consensus sees serendipity as a property of recommendations that are both unexpected and valuable, in practice, different researchers have a diverse understanding of the concept, which can lead to different implementations. This may derive from the word that was chosen for this concept: the term “Serendipity” has a long history both in literature and pop culture [113] and is sometimes seen to have an almost magical quality [114]. It is important to avoid being influenced by the poetic aura of the word and instead be guided by the needs of recommender systems when dealing with serendipity; therefore, finding a more unifying definition should be one of the goals of researchers in this field.

Another open challenge for serendipity is that of evaluation. As noted by prior reviews on serendipity [73], it is not easy to evaluate this specific aspect of a recommender system, and some researchers fail to do this to a satisfactory degree. Offline evaluation has become easier since the release of datasets specifically aimed at serendipity [115], but it is not entirely clear how well offline evaluation can capture the concept of serendipity since many of the metrics for serendipity require proxies for the concepts of unexpectedness. These are usually based on novelty and diversity or other concepts related to long tail exploration, but it might be necessary to define different approaches that better capture the user-based notion of unexpectedness. Online evaluation can be implemented by simply asking users to evaluate how serendipitous they find the recommended items (or how unexpected they found items they enjoyed) [116], but to have evaluations that can be compared across different systems, it is necessary to determine a reproducible evaluation processes and, before that, to have an agreed definition for serendipity.

Both the problem of providing clear definitions and that of evaluation are problems that are known in the field of computational creativity (http://computationalcreativity.net/ (accessed on 17 February 2025)), which attempts to obtain creative behaviors from computational systems. As noted by Lu and Chung [92], there is a strong relationship between the concepts of serendipity, which we describe as unexpectedness and usefulness, and that of creativity, which is usually described as novelty and value [117,118], although they are not the same (it must be noted that this concept of novelty is not the one used in recommender systems but rather describes the fact that creative artifacts are new). We propose that it is possible to learn from the past experience of computational creativity practitioners for some useful lessons regarding serendipity research. One such lesson is understanding where and when serendipity occurs. Is it a property of the system, capable of producing serendipitous recommendations, of the recommendations, that are serendipitous, of the user’s perception of the recommendations when they find them serendipitous, or of the researchers/system owners that deem certain interactions serendipitous? These viewpoints roughly correspond to the four perspectives of creativity: process, product, person, and press [119]. Understanding these different perspectives can perhaps provide researchers with inspiration for different approaches to serendipity and to its evaluation. Creativity is understood as an ex-post property [120]: a property that is attributed to something after its creation and, as such, cannot be imposed by construction. Arguably, serendipity is similarly a property that cannot be constructed but is found. For this reason, it is important to employ evaluation paradigms that go beyond the simple quantitative evaluation of the produced recommendations but rather evaluate the recommendation process itself and its capability of potentially producing serendipitous recommendations [121]. The literature on computational creativity offers many paradigms for evaluation [122] and for the description of how a conceptual space is explored [123,124], and we believe that it would be possible to employ many of those ideas in the context of serendipitous recommendations.