Referenced from the paper A Survey on Session-based Recommender Systems
This paper first introduces the motivation and background of conversation-based recommendation,
then explains its difference from sequential recommendation,
and gives the subfields of conversational recommendation
and the classification framework of various methods currently used, and
finally proposes the future outlook in the current field.
—— cinquain
General Recommendation (RS) vs Session-Based Recommendation (SBRS)
general recommendation
Recommender systems can help users find related items of interest. In general recommendation systems, all historical interactions between users and items are usually used to learn users' long-term, static preferences for items. Such an operation actually implies an assumption: all historical interaction information is equally important for predicting the user's current preference.
But this assumption is not correct in actual scenarios because:
- Users not only have long-term past preferences, but also short-term preferences. These short-run preferences can be characterized by recent interaction information. (eg: Although I have always listened to classical music, there are one or two pop songs that are very popular recently, so I will also listen to them)
- User preferences also change over time and are dynamic rather than static. (eg: Although I used to like to listen to classical music, but now I don't like it, I have already thrown myself into the embrace of pop music)
Session-Based Recommendations
Therefore, in order to bridge these gaps, Session-based Recommender Systems (SBRS) have been proposed in recent years. SBRS uses each session as the basic input unit, which can capture the user's short-term preferences and dynamic preferences reflected by the interest transfer between sessions , thereby improving the accuracy and timeliness of recommendations.
A broad SBRS may have the following subfields:
- Predict the next interaction (next-interaction) eg: predict the next product, song/movie, POI, webpage, news, etc.
- Predict the interaction of the remainder of the current session (next partial-session recommendation) eg: predict the next item, session/basket completion
- Predict the next session*(next-session)*eg: Next basket recommendation, Next bundle recommendation
In the narrow sense, SBRS refers to the first type. (Based on the current session, predict the next interaction)
Sequential Recommendation (SRS) vs Session Based Recommendation (SBRS)
(Ambiguity: Inter-session: refers to the interior of a session, or between sessions? It is the latter, but it needs to be explained)
Sequential Recommendation Systems are recommendations based on sequence data , while session recommendations are based on session data. The two are very related and superficially similar, which can cause a lot of confusion.
First give the definition of session data and sequence data:
- session : A session is a sequence of user-item interactionswith distinct start and end boundaries . These interactions within a session may be ordered (ordered) or unordered (unordered). Session data consists of manysessions that occur at different times and are separated by many boundaries . There are different time intervals between sessions (time interval).
- sequence : A sequence is an ordered collection of interactions. All interactions of a user form a sequence, so there is only one boundary .
There is a boundary, which implies that there are co- occurrence -based dependencies between items in a certain period of time. This co-occurrence-based dependency forms the basis of SBRS. (Especially when the items in the session are out of order.) (eg: next-bastket recommendation? Yes, it can be regarded as a subfield of SBRS)
In the experimental part of each SBRS model, there is an operational description of segmenting user-item interaction sequences into sessions. Usually by specifying a time interval threshold (eg: 30 minutes / 8 hours). If this threshold is exceeded and the user still has no new activity, divide the previous interaction into a session.
Now, we can distinguish between session-based and sequential recommendations:
- Session-based recommendation : Predict the remaining interactions/next sessions of the next interaction/session, and exploit co-occurrence-based dependencies. In principle, SBRS does not require user-item interactions to be ordered , so it does not use sequential dependencies. Of course, if the session is ordered, we can also use it.
- Sequence Recommendation : Based on sequence data, predict the next item. Take advantage of sequence dependencies.
It is easy for us to confuse session recommendation and sequence recommendation, because many SBRS researches are based on sequenced sessions and predict the next interaction (compare the second and third rows of the table below).
The essence of the distinction is: whether to use co-occurrence-based dependencies
Subfields for Session-Based Recommendations
Session-based Recommendation System (SBRM) session-based recommendation algorithms can be divided into three categories:
- Conventional SBRS approaches
- Latent representation based approaches (latent representation model)
- Deep neural network based approaches (method based on deep neural network)
This article first introduces the first type of Conventional SBRS approaches in detail. In recommendation, early traditional methods usually use techniques such as data mining and machine learning to capture the dependencies in sequence data, including the well-known Item-KNN, FPMC and other baselines models that are often used for comparison. Understanding this part of the content is of great help in getting started. The four traditional recommendation algorithms will be introduced in detail below.
Patten/Rule Mining-based SBRSs (pattern/rule-based mining)
Two algorithms are included:
Association Rule Mining (Frequent Pattern/Association Rule Mining-based Approaches):
Divided into three steps: ——Evaluation: The idea should be introduced first, not the algorithm steps. Summarize as a whole.
(1) Find frequent itemsets and association rules
In all sessions (considered as user-item interaction sequences), first find all itemsets that meet the minimum scale threshold, these itemsets are called frequent itemsets FP //calculate P(AB) frequent itemsets that meet the minimum confidence requirement, called association rules (Association Rule) //calculate P(B|A) (2) Generate candidate items: For items in the item set, if they form FP together with items in the current session, then (3) Generate recommended items: For a candidate item, if its FP meets the confidence requirement (assostion rule), it can
be added to
the
recommendation
list
Comments: You can see that this method does not require the session to be ordered, but it does not take into account the sequence. To put it bluntly, only the co-occurrence probability can be calculated. The obtained association rules can only be interpreted as the probability that these two items are bought together is high.
Sequential pattern mining-based approaches
Sequential pattern mining takes sequentiality into account. According to the last purchase behavior (session), recommend the items of the next session starting from this session (as candidate items) that appear in the sequence mode.
The input of sequential pattern mining is different from the above frequent itemset mining: sequence set VS item set. The input of
sequence pattern mining is such data: the data of one line is all sessions of a user. The
input data of frequent itemset mining is like this: one line is a session. Probably rows 1,2,3 are all from the same user (regardless of user). Although the input is ordered, the order between items is not considered when recommending, only the co-occurrence is considered.
It can be seen that if the length of each session of the user is 1, sequential pattern mining can be regarded as frequent itemset mining.
Sequence mining step: similar to association rule mining, it is equivalent to changing an element from an object to a sequence. That's all
(1) Sequence pattern mining
(2) Sequence matching
(3) Generate recommendations: Candidate items: In the sequence rules that are mined, the items in the next sequence that match the last sequence. Confidence is then calculated for each item.