A Survey on Session-based Recommender Systems 39:7
Note that some actions may not be associated with specic items, e.g., a search action or a catalog
navigation action. But they may still provide useful information to an SBRS as discussed in [91].
3.4 Interaction and Interaction Properties
Interaction is the most basic unit in sessions. Let
𝑜
denote an interaction, which is a ternary tuple
consisting of a user
𝑢
, an item
𝑣
and the action
𝑎
taken by
𝑢
on
𝑣
, namely
𝑜 = ⟨𝑢, 𝑣, 𝑎⟩
. In the case
where the user information is not available, the interaction become anonymous, i.e.,
𝑜 = ⟨𝑣, 𝑎⟩
.
Moreover, in the case, where there is only one type of actions, e.g., clicks, the interaction
𝑜
can be
further simplied as
𝑜 = ⟨𝑣⟩
, namely it only consists of an item. All the interactions together form
the interaction set 𝑂.
3.5 Session and Session Properties
Session is an important entity in an SBRS. Let
𝑠
denote a session, which is a non-empty bounded list
of interactions generated in a period of continuous time which may be connected with some user-
(e.g., user ID) or session-specic (e.g., a session-ID or a cookie) information, i.e.,
𝑠 = {𝑜
1
, 𝑜
2
, ..., 𝑜
|𝑠 |
}
.
Note that here we use the concept "list" instead of "set" to indicate that there may be duplicated
interactions in one session. For example, a user listens to a song for multiple times in a listening
session. Each session is associated with a set of attributes, e.g., the duration of
𝑠
, which have
multiple corresponding values, e.g., 20 minutes or 40 minutes. Some other important attributes of a
session include the time and the day when the session happens. Next, we discuss ve important
properties of sessions that may have a great impact on SBRSs.
Property 1: session length. The length of a session is dened as the total number of interactions
contained in it. This is a basic property of sessions, which is taken as one of the statistical indicators
of experiment data in most literature [
49
,
126
]. Sessions of dierent lengths may bring dierent
challenges for SBRSs and thus lead to dierent recommendation performance. The session charac-
teristics related to session length together with the corresponding challenges for building SBRSs
are discussed in detail in Section 4.1.
Property 2: internal order. The internal order of a session refers to the order over interactions
within it. Usually, there are dierent kinds of order exibility inside dierent sessions, i.e., no order,
exible order and order. The existence of internal order leads to the sequential dependencies within
sessions which can be used for recommendations. The session characteristics related to internal
order and the corresponding challenges for building SBRSs are discussed in detail in Section 4.2.
Property 3: action type. In the real world, some sessions contain only one type of actions, e.g.,
purchase, while other sessions may contain multiple types of actions, e.g., click, purchase (cf. Fig. 2
(a)). The dependencies over dierent types of actions are often dierent. For instance, the items that
are clicked together in a session may be similar or competitive while the items purchased together in
one session may be complementary. Therefore, the number of action types in a session determines
whether the intra-session dependencies are homogeneous (based on a single type of actions) or
heterogeneous (based on multi-type actions), which is important for accurate recommendations.
The session characteristics related to action type as well as the corresponding challenges for
building SBRSs are discussed in detail in Section 4.3.
Property 4: user information. User information of a session mainly refers to the IDs of the users
in the session, and sometimes user attributes are also included. In this paper, the property of user
information refers to the availability of user information in a session. In the real word, the user
information of sessions is given in some cases, while it is not available in other cases (cf. Section
3.1) [
43
,
116
,
138
]. User information plays an important role to connect sessions from the same
user happening at dierent time and thus its availability determines the possibility to model the
long-term personalized preference across multiple sessions for a specic user. In practice, SBRSs
ACM Comput. Surv., Vol. 9, No. 4, Article 39. Publication date: May 2021.