Learning complex relations for session-based recommendations

Publication Type:
Issue Date:
Full metadata record
In the era of big data, recommender systems (RSs) are a powerful engine to promote intelligent life by helping humans to make decisions concerning their daily necessities (e.g., food, clothes, and houses) much more efficiently and effectively, selecting from a large number of choices. Of the various types of recommender systems, session-based (SB) ones are of great value and significance, but they are not well studied. The value of session-based recommender systems comes from two fold. From the research perspective, a SBRS takes a session as the basic unit for data organization and thus keeps the intrinsic nature of the original transaction-like data. As a result, the system effectively retains and models the rich information (e.g., intra-session dependency) embedded in a session structure to produce a more reliable recommendation. This modelling cannot be achieved by other types of recommender systems because they usually break down the original session data into multiple pair-wised user-item interactions to fit the models. From the business perspective, session data for session-based recommender systems is much more readily available than either the rating data or the item attribute data required by other recommender systems including content-based or collaborative filtering ones. This actually makes session-based RSs much more applicable in real-world business. Though valuable, SBRSs are quite challenging. Generally, a hierarchical architecture consisting of five levels (cf. Figure 1.1) is built from the low-level feature values till to the high-level sessions in session data, as demonstrated in Chapter 1. The challenge arises mainly comes from three considerations: the heterogeneity of the elements in each level (e.g., there are both categorical and numerical features), the complex dependency within each level (e.g., the implicit inter-item relations), and the interactions between different levels (e.g., the inter-session dependency may affect the item occurrence). From my observation, the existing works mainly focus on the general item-level dependency modelling for session-based recommendations while ignoring other level relations as demonstrated in Chapter 3. To bridge the huge gaps between the existing works and great challenges mentioned above, I build a systematic framework consisting of dependency modelling from the three core levels, i.e., feature-level, item-level and session-level (cf. Figure 1.1), for session-based recommendations. To the best of my knowledge, this is the first framework to systematically address various levels of challenges in session-based recommendations. Particularly, due to the limitations regarding space, I address one or two critical challenges in each level, as shown below. In Chapter 4, to capture the implicit inter-item relations ignored by existing rule-based approaches, I proposed an implicit rule-based RS that first infers implicit rules and then applies the resultant rules for reliable rule-based recommendations, a basic approach for session-based recommendations. In Chapter 5, I continue to work on item-level dependency modelling and focus on the issue of item heterogeneity, referring to different items with different levels of relevance to the next choice of an item. To this end, I build an attentive transaction-embedding model to discriminatively integrate multiple items in a transaction context into a unified context-embedding for next-item recommendations. In Chapter 6, the feature-level dependency and feature-item interactions are modelled by a shallow neural network which takes both contextual items in a session and their corresponding features as the input. Accordingly, the cold-start issue in SBRSs has been well addressed. In Chapter 7, the session-level (i.e., transaction-level) dependency and session-item interactions are modelled. A hierarchical attentive transaction embedding model is built to jointly model the intra-session (item-level) and inter-session (session-level) dependency. Accordingly, the influence from previous sessions on a current session is incorporated for more accurate next-item recommendations. All these models are applied to real-world transaction data, like Tmall and Tafang and they clearly outperform other representative SBRSs. More importantly, this thesis proposes a systematic framework to explore the driving force behind SBRSs, which provides some insights into both the researchers and engineers in this domain.
Please use this identifier to cite or link to this item: