Transfer to Rank for Top-N Recommendation Wei Dai, Qing Zhang, Weike - - PowerPoint PPT Presentation

Transfer to Rank for Top-N Recommendation

Wei Dai, Qing Zhang, Weike Pan∗ and Zhong Ming∗

daiwei20171@email.szu.edu.cn, qingzhang1992@qq.com, panweike@szu.edu.cn, mingz@szu.edu.cn

National Engineering Laboratory for Big Data System Computing Technology and College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 1 / 27

Introduction

Problem Definition

Top-N Recommendation Input: Ratings in the form of (user, item, rating) triples. Goal: Recommend a personalized ranked list of items to each user u from the items that user u has not examined or rated before, i.e., I\Iu, where Iu = IE

u denotes the set of examined or rated items by

user u.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 2 / 27

Introduction

Challenge

How to exploit users’ explicit feedback more sufficiently?

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 3 / 27

Introduction

Overall of Our Solution

Coarse-to-Fine Transfer to Rank (CoFiToR)

1

We view the rating records from three different but related perspectives, i.e., examinations, scores and purchases.

2

We decompose a user’s shopping process into three stages, i.e., E-stage, S-stage and P-stage, which correspond to three specific questions, including (i) whether an item will be examined by a user, (ii) how an item will be scored by a user, and (iii) whether an item will finally be purchased by a user. Our CoFiToR progressively models users’ preferences from a coarse granularity to a fine granularity by transferring knowledge in the form of candidate item lists.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 4 / 27

Introduction

Advantages of Our Solution

1

It is able to model and simulate a user’s shopping process in a coarse-to-fine manner.

2

It is a generic, flexible and efficient transfer learning framework for top-N recommendation.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 5 / 27

Introduction

Notations

Table: Mathematical notations and explanations.

Symbol Descriptions n number of users m number of items d dimension of latent feature vector U = {u} user set I = {i} item set R = {(u, i, rui)} rating records in training data rui rating of user u to item i ¯ rui normalized rating of user u to item i IP

u

items purchased by u in training data IE

u

items examined by u in training data ˆ rui predicted preference of user u to item i Uu· ∈ R1×d user u’s latent feature vector Vi· ∈ R1×d item i’s latent feature vector bu ∈ R user u’s bias bi ∈ R item i’s bias

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 6 / 27

Background

Bayesian Personalized Ranking

The loss function of BPR can be formulated as follows, min

Θ

• u∈U
• i∈IP

u

• j∈I\IP

u

− ln σ(ˆ rui − ˆ ruj), (1) where ˆ rui = Uu·V T

i· + bi is the prediction rule, Uu· is user u’s latent vector,

Vi· is item i’s latent vector, and bi is the bias of item i. BPR orders items by modeling the distance between the preference on a purchased item and an un-purchased item, i.e., ˆ rui − ˆ ruj, via a pairwise loss function. Due to its excellence in effectiveness and efficiency, we choose BPR as an essential component in our three-staged recommendation framework to ensure that the candidate items are what the users are likely to examine.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 7 / 27

Background

Probabilistic Matrix Factorization

The loss function of PMF can be formulated as follows, min

Θ

• (u,i,rui)∈R

(rui − ˆ rui)2, (2) where ˆ rui = Uu·V T

i· + bu + bi + µ is the prediction rule, Uu· is user u’s

latent vector, Vi· is item i’s latent vector, bu is the bias of user u, bi is the bias of item i, and µ is the global average. PMF is a popular pointwise regression-oriented recommendation method that models a user’s preference as the inner product of two low-rank user and item feature vectors, and it scales well to large and sparse datasets. In our proposed solution, the goal of adopting PMF is to refine the candidate list of items from a complementary view of rating scores.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 8 / 27

Background

Transfer to Rank (1/2)

1

In the first stage, the general objective of global preference learning is as follows, Prob(E ∪ ER|Θg), (3) where E is the examination behavior, ER denotes the (user, item) pairs of the rating records, and Θg is the set of model parameters used to govern the generation of the combined examination behavior E ∪ ER.

2

In the second stage, ToR focuses on users’ local preference learning aiming at refining the candidate list of items Lg generated from the first stage. The general objective of local preference learning is as follows, Prob(R|Θℓ; Lg), (4) where Θℓ is the set of model parameters used to govern the generation of the rating behavior R.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 9 / 27

Background

Transfer to Rank (2/2)

The two sequential stages of ToR aim to answer the following two related questions: Whether a user will examine an item; How will a user like an item if he/she has examined it. Hence, ToR can simulate a user’s shopping process to some extent. But for different users, even he/she assigns a high score to an item, it doesn’t mean that the user will finally purchase the item and vice versa. That is to say, users not only differ in tastes of items, but also in purchase choices, which is decisive but neglected by ToR.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 10 / 27

Method

CoFiToR: Assumptions

1

If a user examines an item, it means that the user is interested in the item;

2

A user virtually or implicitly scores an item if and only if he/she has examined it;

3

If a user likes an item, there is a high probability that he/she will purchase it;

4

A user may not purchase the item even if he/she likes it and vice versa. Based on these four assumptions, we design a three-staged transfer learning framework to bridge the connection among the three aforementioned actions in a proper way.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 11 / 27

Method

CoFiToR: Illustration

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 12 / 27

Method

CoFiToR: Objective Function

The optimization problem is as follows, Prob(E, S, P|R), (5) where E, S and P denote the derived examinations, scores and purchases from the original rating records. We decompose it into three sequential sub-problems resulting in our three-staged solution, Prob(E|ΘE) → Prob(S|ΘS; LE) → Prob(P|ΘP; LS), (6) where ΘE, ΘS and ΘP are the model parameters to be learned, and LE and LS are the candidate lists of likely to be examined items and high-score items, respectively.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 13 / 27

Method

E-stage: Whether Will It Be Examined

Adopted model: Revised BPR Goal: Extract items that users are most likely to examine Input: Examinations E = {(u, i)|(u, i, rui) ∈ R} Output: Candidate list LE with NE items of each user The objective function is as follows, min

Θ

• u∈U
• i∈IE

u

¯ rui

• j∈I\IE

u

fuij, (7) where ¯ rui = (2rui − 1)/25 is a normalized version of the rating rui of user u to item i in the original rating records R, and IE

u is the set of items

examined by u in training data. Notice that the rating ¯ rui is expected to widen the gap between positive items and negative items, and distinguish positive items with different rating values.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 14 / 27

Method

S-stage: How Will It Be Scored

Adopted model: PMF Goal: Model users’ score preferences Input: Scores S = R, candidate list LE Output: Candidate list LS with NS items of each user (NS < NE) Notice that the objective function is the same as that in PMF. We bridge two types of user actions, i.e., examinations and scores, based on the aforementioned assumption, i.e., a user virtually or implicitly scores an item if and only if he/she has examined it. Notice that in ToR, the obtained candidate list LS in this stage will be used for recommendation immediately without a further refinement, which is one major difference between the two-staged solution ToR and our three-staged framework CoFiToR.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 15 / 27

Method

P-stage: Whether Will It Be Purchased

Adopted model: BPR Goal: Extract items that users are most likely to purchase Input: Purchases P = {(u, i)|(u, i, 5) ∈ R}, candidate list LS Output: Candidate list LP with NP items of each user (NP < NS < NE) Notice that the objective function is the same as that in BPR. We argue that the two stages in ToR may not simulate the reality of a shopping process well because they ignore the relationship between “score” and final “purchase”. The main reason why we reuse BPR in P-stage is that pairwise preference learning in BPR and pointwise preference learning in PMF have a strong complementarity as has been shown in ToR. The candidate list LP is our final recommendation for users.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 16 / 27

Experiments

Datasets

Table: Statistics of the first copy of the two datasets used in the empirical studies. Notice that n and m denote the number of users and the number of items, respectively, and |R|, |P(va.)| and |P(te.)| are the number of rating records in training data, the number of purchases in validation data, and the number of purchases in test data, respectively. We also use |R(5)| to denote the number of (u, i, rui) triples in R with the highest rating value rui = 5 for analysis of the performance of the one-class collaborative filtering algorithm in the experiments.

Dataset n m |R| |P(va.)| |P(te.)| |R(5)|/n MovieLens 10M 71,576 10,681 6,000,034 308,673 308,702 12.96 Netflix 480,189 17,770 59,443,266 4,556,505 4,557,413 28.47

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 17 / 27

Experiments

Baselines

PMF: Probabilistic Matrix Factorization BPR: Bayesian Personalized Ranking ToR: Transfer to Rank CoFiToR(E,P): CoFiToR without S-stage

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 18 / 27

Experiments

Parameter Configurations

For BPR, PMF, ToR and our CoFiToR, we first fix the learning rate as γ = 0.01, the dimension of latent feature vectors as d = 100, and then search the optimal hyper-parameter on the regularization terms α ∈ {0.001, 0.01, 0.1} and the iteration numberT ∈ {100, 500, 1000} via the performance of NDCG@5 on the validation data. We select the items with highest estimated preferences via BPR, PMF, ToR or CoFiToR from I\IE

u rather than from I\IP u , which

ensures a fair comparison from the perspective of candidate items. Notice that IE

u = Iu denotes the set of examined/rated items by user

u.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 19 / 27

Experiments

Evaluation Metrics

Precision@5 Recall@5 F1@5 NDCG@5 1-call@5

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 20 / 27

Experiments

Results (1/4)

Table: Recommendation performance of our proposed CoFiToR and three closely related algorithms, including PMF, BPR, and ToR, on MovieLens 10M and Netflix w.r.t. five commonly used ranking-oriented evaluation metrics. The best results are marked in bold.

Dataset Method Pre@5 Rec@5 F1@5 NDCG@5 1-call@5 MovieLens 10M PMF 0.0230±0.0008 0.0179±0.0006 0.0164±0.0005 0.0305±0.0010 0.0906±0.0026 BPR 0.1300±0.0008 0.1664±0.0017 0.1218±0.0010 0.1804±0.0015 0.4636±0.0034 ToR 0.1248±0.0004 0.1656±0.0010 0.1183±0.0005 0.1788±0.0005 0.4552±0.0010 CoFiToR(E,P) 0.1432±0.0006 0.1864±0.0016 0.1345±0.0008 0.1992±0.0012 0.5009±0.0020 CoFiToR 0.1448±0.0005 0.1882±0.0014 0.1358±0.0007 0.2012±0.0010 0.5059±0.0026 Netflix PMF 0.0612±0.0008 0.0331±0.0005 0.0345±0.0005 0.0694±0.0009 0.2181±0.0025 BPR 0.1532±0.0004 0.1051±0.0002 0.0973±0.0002 0.1787±0.0003 0.4911±0.0007 ToR 0.1431±0.0008 0.0979±0.0006 0.0900±0.0006 0.1701±0.0008 0.4700±0.0022 CoFiToR(E,P) 0.1595±0.0003 0.1103±0.0004 0.1011±0.0003 0.1867±0.0004 0.5062±0.0012 CoFiToR 0.1606±0.0004 0.1098±0.0004 0.1012±0.0003 0.1875±0.0004 0.5065±0.0012

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 21 / 27

Experiments

Results (2/4)

Observations: Our proposed CoFiToR achieves significantly better performance than the baseline methods. The variant of our CoFiToR with only E-stage and P-stage, i.e., CoFiToR(E,P), performs slightly worse than the three-staged solution in most cases, which shows the necessity of S-stage in our proposed coarse-to-fine transfer learning framework. . . .

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 22 / 27

Experiments

Results (3/4)

ToR-E ToR-S CoFiToR-E CoFiToR-S CoFiToR-P 0.1 0.15 0.2

NDCG@5

MovieLens 10M

ToR-E ToR-S CoFiToR-E CoFiToR-S CoFiToR-P 0.1 0.12 0.14 0.16 0.18 0.2

NDCG@5

Netflix

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 23 / 27

Experiments

Results (4/4)

Observations: CoFiToR-E > ToR-E: the effectiveness of our rating modeling strategy (weighting different positive instances) in E-stage. ToR-S > ToR-E, CoFiToR-S > CoFiToR-E: the effectiveness of regression-oriented preference modeling in S-stage. CoFiToR > ToR: ToR merely considers the correlation between “examination” and “score” but ignores whether an item will finally be “purchased”, which is taken into account in P-stage in our CoFiToR. . . .

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 24 / 27

Related Work

Related Work

Memory-based methods for top-N recommendation

Regression-oriented memory-based methods Ranking-oriented memory-based algorithms

Model-based methods for top-N recommendation

Pointwise approaches Pairwise approaches Listwise approaches

Our CoFiToR is a model-based methods that adopts a pointwise approach and a pairwise approach in a progressive way.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 25 / 27

Conclusions

Conclusions

We propose a novel and generic transfer learning based recommendation algorithm, i.e., coarse-to-fine transfer to rank (CoFiToR), which contains three sequential stages based on a novel perspective with three different views of the explicit feedback, i.e., examinations, scores and purchases. This perspective is based on a reverse engineering and decomposition of users’ shopping processes, i.e., from examination to pre-purchase score to final purchase. In comparison with previous top-N recommendation approaches, our CoFiToR is able to efficiently model users’ shopping processes by capturing users’ preference information more accurately in a coarse-to-fine manner.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 26 / 27

Thank You

Thank You!

We thank the handling Associate Editor and Reviewers for their expert and constructive comments and suggestions. We thank Ms. Wanqi Ma for helpful discussions, and the support of National Natural Science Foundation of China Nos. 61872249, 61502307, 61836005 and 61672358.

Dai, Zhang, Pan and Ming (SZU) Coarse-to-Fine Transfer to Rank (CoFiToR) IEEE Trans. on Big Data 27 / 27