The hotel may be close to the destination airport to the flight
Location for renting car may not be far away from the hotel.
Sequential Recommendation System takes the prior sequential interactions as a context to predict which items would be interacted in the near future.
Main Contribution of the paper:
analyze the key challenges caused by different data characteristics in SRSs.
summarize the current research progress in SRSs by systematically categorizing the state-of-the-art works from a technical perspective.
Data Characteristics and Challenges
Handling Long User-Item interaction sequences a. Challenge 1: Learning higher-order sequential dependencies : Dependency on large number of previous transactions for purchase order b. Challenge 2: Learning long-term sequential dependencies : Dependency of current purchase on the previous items that are far from here.
In the above sequence,
If we consider say n no. of previous purchase, it is n order sequential dependency.
the dependency of rose on vase after 3 interactions or purchases is the long-term sequential dependency.
Existing Solutions:
Higher-order Markov Chain
RNN
2. Handling user-item interaction sequences with a flexible order
In a shopping sequence = {milk, butter, flour}, - the order of buying the milk and butter doesn’t matter - But the union of milk and butter leads to higher probability of buying the flour next.
In the condition where there flexibility in the order of purchase, it is better to capture the collective sequential dependencies rather than point-wise dependencies
Existing Solutions:
CNN-based SRSs.
3. Handling user-item interaction sequences with noise
Some historical interactions are strongly relevant to the next interaction
While, others may be weakly or even irrelevant
Existing Solutions:
Attention models
4. Handling user-item interaction sequences with heterogeneous relations
Other than sequential dependencies, there can also be similarity based relations between the interacted items in terms of their features.
Hence another key challenges in building SRSs is how to effectively capture these heterogenouse relations embedded in the user-item interaction sequences respectively and to make them work collaboratively for the sequential recommendations.
Existing Solutions:
Mixture models are the only solution so far for this challenge
The mixture model integrates different types of relations modelled by different sub-models to collaboratively generate the sequential recommendations.
5. Handling user-item interaction sequences with hierarchical structures
a. Hierarchical Relation 1: hierarchical structure between the metadata and user-item interactions. For ex: the user’s demographics can determine the users’ preferences in some degree.
b. Hierarchical Relation 2: hierarchical structure between sub-sequences and user-item interactions. i. In some SRSs, one user-item interaction sequence includes multiple sub-sequences(also called sessions). These historical sequences can affect the current purchase.
