推薦系統：Next Basket Recommendation

03-06

Next Basket Recommendation: A Survey

Next basket recommendation: Predict what a user most probably would like to buy next when his/her sequential transaction data is given.

1 Rendle, Steffen, Christoph Freudenthaler, and Lars Schmidt-Thieme. "Factorizing personalized markov chains for next-basket recommendation." Proceedings of the 19th international conference on World wide web. ACM, 2010.

Factorized Personalized MC (FPMC): Matrix factorization (MF) + Markov chains (MC)

Personalized transition graphs over underlying Markov chains.
An own transition matrix is learned for each user.
Factorize the transition cube with a pairwise interaction model.

Markov Chains

Order m=1

Personalized Markov Chains

Factorizing Transition Graphs

Use a special case of Canonical Decomposition (CD) that models pairwise interactions

Summary of FPMC

Item Recommendation

(Sequential BPR

)

2 Wang, Pengfei, et al. "Learning hierarchical representation model for nextbasket recommendation." Proceedings of the 38th International ACM SIGIR conference on Research and Development in Information Retrieval. ACM, 2015.

Hierarchical Representation Model (HRM)

A two-layer structure is employed to construct a hybrid representation over user and items from last transaction, which is used to predict the next purchased items.

define the probability ofbuying next item i given user u and his/her last transaction $T_{t-1}^u$ via a softmax function

the hybrid representation is obtained from the hierarchical aggregation

$f(cdot)$ : the aggregation operation

average pooling

max pooling

Learning and Prediction

Negative sampling

3 Feng, Shanshan, et al. "Personalized Ranking Metric Embedding for Next New POI Recommendation." IJCAI. 2015.

Personalized Ranking Metric Embedding (PRME)

Embedding POIs in latent space. Each POI $l$ has a position $X(l)$ in a K-dimensional space

transition probability

$Z$ is the normalization term

Ranking based metric embedding

Personalized Ranking Metric Embedding

Incorporating Geographical Influence

d: the geographical distance

Parameter Learning (BPR)

4 Yu, Feng, et al. "A dynamic recurrent model for next basket recommendation." Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval. ACM, 2016.

Dynamic REcurrent bAsket Model (DREAM)

Pooling operation on the items in a basket to get the representation of the basket. The input layer comprises a series of basket representations of a user. Dynamic representation of the user can be obtained in the hidden layer. Finally the output layer shows scores of this user towards all items.

$n_v$ indicates the latent representations of item v

$B^u_{t_i}$ is a basket of items purchased by user u at time ti

max pooling

average pooling

hidden layer

output

Objective Function (BPR)

5 He, Ruining, et al. "Vista: A Visually, Socially, and Temporally-aware Model for Artistic Recommendation."Proceedings of the 10th ACM Conference on Recommender Systems. ACM, 2016.

the transition of user u from item $S_{t-1}^u$ to item i can be explained from two aspects:

the interaction between user u and item i, which captures the long-term preferences of user u
the interaction between the previous item Sut-1 and item i, which captures the short-term or temporary interest of user u

$o_iin U$ : owner/creator of item i

Higher-Order Markov Chains

Incorporating Content-Based Features

$f_i$ : the explicit feature vector of item i

Objective Function (S-BPR)

6 He, Ruining, Wang-Cheng Kang, and Julian McAuley. "Translation-based Recommendation." Proceedings of the Eleventh ACM Conference on Recommender Systems. ACM, 2017.

TransRec

$gamma$ : item vector

represent each user u with a translation vector $t_u$ to capture us inherent intent or long-term preferences that influenced her to make these decisions

the inherent triangle inequality assumption plays an important role in helping the model to generalize well

add another translation vector $t$ to capture global transition dynamics across all users

transition probability

$d(x,y)$ : some distance metric

Ranking Optimization (S-BPR)

（本文持續更新）