: Hybrid Associations Models for Sequential Recommendation

M2: Mixed Models With Preferences, Popularities and Transitions for Next-Basket Recommendation

“…The primary difference among M 2 -gp 2 t, Dream and FPMC is that M 2 -gp 2 t explicitly models users' general preferences using the frequencies of the items that each user has interactions with, while Dream and FPMC implicitly model them using the hidden state of RNNs or user embeddings. Given the sparse nature of recommendation datasets (Table 2), it is possible that the learned hidden states or user embeddings cannot represent the user preferences well, as the signals of user preferences are smoothed out due to data sparsity during the recurrent updates, or by the pooling or weighting schemes used to learn user embeddings as some other work also noticed [18], [36], [37]. The superior performance of M 2 -gp 2 t over Dream and FPMC on all the datasets demonstrates the effect of explicitly modeling users' general preferences.…”

“…Table 3 shows that M 2 -gp 2 t significantly outperforms Sets2Sets on all the datasets except TMall in terms of both recall@k and NDCG@k. The primary differences between M 2 -gp 2 t and Sets2Sets are 1) M 2 -gp 2 t explicitly models the transition patterns among items using encoder-decoderbased ed-Trans, while Sets2Sets implicitly models the transition patterns using RNNs, and 2) when calculating the recommendation scores, M 2 -gp 2 t learns a single weight on each user (i.e., α in Equation 10), but Sets2Sets learns different weights for different items on each user. Given the sparse nature of the recommendation datasets, weights for different items on each user may not be well learned [18], [36]. Thus, such weights may not necessarily help better differentiate user general preferences over items.…”

“…Recently, Ma et al [17] developed a hierarchical gating network (HGN), which uses gating mechanisms to identify important items and generate recommendations. Peng et al [18] developed hybrid associations models (HAM), which adapt the pooling mechanisms to model the association patterns and synergies among items.…”

“…Previous studies have shown that users' interactions are significantly affected by their general preferences [1], [6], which are also known as the long-term preferences in the literature [6], [18]. For example, some users prefer items of low price, while others may like luxurious items that could be expensive.…”

“…The transitions among items is another important factor in inducing the next baskets of items that the users will be interested in [2], [14], [18]. For example, if a user purchased cat toys in a basket, she/he is likely to purchase cat food and treats in the next baskets compared to wine and beers, as there could be stronger transitions among cat items compared to from cat items to alcohols.…”