Xing Zhao, Qingquan Song, James Caverlee and Xia Hu Department of - - PowerPoint PPT Presentation

Xing Zhao, Qingquan Song, James Caverlee and Xia Hu

Department of Computer Science and Engineering Texas A&M University, USA

1

Da Dataset Statistics cs

Items Quantity Proportion

Playlists 1,000,000 Unique Tracks 2,262,292 100% Unique tracks (freq ≥ 5) 599,341 96.05% Unique tracks (freq ≥ 100) 70,229 80.67% Unique albums 734,684 Unique artists 295,860

4

Track Appeared Times in Training Data

1 5 10 100 1000 10000 40000

Number of Remaining Tracks

#106 0.5 1 1.5 2 2.5

Cumsum Taking Up of Positive Samples

0.2 0.4 0.6 0.8 1

Therefore, in some part of our methods, we

• nly consider these tracks for training.

Ou Our Me Metho thod - Tr TrailMix

DNCF C-Tree CC- Title

5

Playlist Continuation: For Task 2 to 10 Cold Start: For Task 1

3 7 5 21 3 43 6 81 8 32 7 13 14 6 5

Tracks (2,262,292) Words (9,817)

Word list 1: Track list 1 Word list 2: Track list 2

Word list 3: Track list 3

… … …

Cluster Recommend

New title: e.g. Pop Punk 2018 Summer

Word list Tracks Word list Tracks Word list Tracks Word list Tracks Word list Tracks

Normalize Pre- process …

6

CC CC-Tit Title le: Co Context t Cl Cluster ering g us using ng Tit Title le

i j

Track i is existed in 6 playlists whose title contain word j

Items Quantity unique titles 92,944 unique normalized titles 17,381 unique non-stop normalized words 9,817 playlist without title after processing 22,921

7

Steps:

• 1. Preprocessing: stemming, stop words,

emoji, punctuation, etc.

• 2. Building word-track matrix of size

9817 x 2,262,292

• 3. Normalizing cells using ‘IDF’
• 4. Clustering words based on row

similarity

• 5. Recommend tracks in each cluster for

new title

CC CC-Tit Title le: Co Cont. t.

8

Highlight:

• 1. CC-Title could deal with large scale of matrix

computation with high efficiency.

• 2. In some cases (clusters), the performance is

very good.

CC CC-Tit Title le: Co Cont. t.

Pros:

• 1. Simple and Generic
• 2. Ensemble the advantages of

basic matrix factorization model and MLP. Cons: Computationally not efficient to be directly applied on the target problem due to the huge item scope and the matrix sparsity.

DNCF: DNCF: Dec Decorated ed Neu Neural Co Collaborati tive e Fi Filter ering

9

He et al. , “Neural Collaborative Filtering”. WWW, 2017.

Neural Collaborative Filtering

DNCF: DNCF: Co Cont. t.

10

Two modifications to address efficiency issue:

Training Phase: Constrained Negative Sampling. Testing Phase: Constrained Recommendation with Reordering.

• 2. Positive samples remain the whole

dataset during training to protect the feasible embedding and prediction of all the testing data. (Task 2-10)

11

• 1. Constrain the negative sampling

space to the space of the tracks appearing equal to or more than 100 times in the training data.

Track Appeared Times in Training Data

1 5 10 100 1000 10000 40000

Number of Remaining Tracks

#106 0.5 1 1.5 2 2.5

Cumsum Taking Up of Positive Samples

0.2 0.4 0.6 0.8 1

Training Phase: Constrained Negative Sampling.

DNCF: DNCF: Co Cont. t.

• 2. Reorder the predicted 500

tracks with an ensemble trick leveraging two types of predictions provided by the Word2Vec embedding.

12

• 1. Constrain the recommendation space by only recommending the

popular tracks (>=100 times) during testing phase towards a more targeted prediction.

Testing Phase: Constrained Recommendation with Reordering.

DNCF Word2Vec (1) Word2Vec (2)

L

1

L2 L3

φ1 φ2 φ3

φ1 \ L

1 ∪ L2 ∪ L3

DNCF: DNCF: Co Cont. t.

13

Highlight:

• 1. Results steadily increase with maximum performance at seed 25;
• 2. It performs better for playlists with random seeding tracks (R) than

sequential seeding tracks;

DNCF: DNCF: Re Result

C-Tree: ee: Construct cted Tree

A Playlist is:

• 1. Natural tree-structure: A playlist

consists of different tracks ,and these tracks always belong to a specific album of an artist;

• 2. Meaningful Cluster: A list of tracks

in a specific playlist always have latent similarity, such as genres, style, listening sense, etc.

14

Phylogenetic Tree.

(Source: https://www.creative-biostructure.com/custom- phylogenetic-tree-construction-service-399.htm)

15

A Real Example (PID: 11548):

• Playlist Title: Pop Puck
• 48 tracks belongs to 12

albums by 5 artists (2 rock bands and 3 pop punk bands)

Pop punk band Rock band

How do we compare the internal relationship? How do we compare it with another tree (external)?

C-Tree: ee: Co Cont. t.

16

Training Data: Complete Tree Testing Data: Incomplete Tree

External comparison

Incomplete Tree: A playlist

• nly contains partial of

tracks (seed), which is waiting for recommending.

C-Tree: ee: Co Cont. t.

Steps:

• 1. Building Forest: 1 million

complete trees;

• 2. Comparing and normalizing the

distance between the incomplete tree T-test and complete tree T-train;

• 3. Recommending the tracks

(leaves) from each T-train to the incomplete tree T-test, based on the score of each leaf.

17

Playlist 1 Playlist 2 Playlist 3 Playlist 4 Playlist n …

C-Tree: ee: Co Cont. t.

18

C-Tree: ee: Re Result

Highlight:

• 1. Results steadily increase with maximum performance at seed 25;
• 2. It performs better for playlists with random seeding tracks (R) than

sequential seeding tracks;

Tr TrailMix: En

Ensemble Mo Model el

CC-Title Final Recommendation ADNCF BDNCF AC-Tree BC-Tree

Num_handou t

Method 1 Method 2

19

Exp Experiment an and Re Result

Experiment Setting:

• Training 80%, testing 20%: cross-validation for hyper

parameter tuning

• Testing data strictly follows the rules designed by

RecSys 2018

20

Thank you!

21

Q&A

22