Diese Präsentation wurde erfolgreich gemeldet.
Die SlideShare-Präsentation wird heruntergeladen. ×

DualDiv: Diversifying Items and Explanation Styles in Explainable Hybrid Recommendation (RecSys 2019)

Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige

Hier ansehen

1 von 1 Anzeige

DualDiv: Diversifying Items and Explanation Styles in Explainable Hybrid Recommendation (RecSys 2019)

Herunterladen, um offline zu lesen

Presented at the 13th ACM Conference on Recommender Systems (RecSys 2019) on September 16th 2019.

Presented at the 13th ACM Conference on Recommender Systems (RecSys 2019) on September 16th 2019.

Anzeige
Anzeige

Weitere Verwandte Inhalte

Weitere von Kosetsu Tsukuda (16)

Aktuellste (20)

Anzeige

DualDiv: Diversifying Items and Explanation Styles in Explainable Hybrid Recommendation (RecSys 2019)

  1. 1. DualDiv: Diversifying Items and Explanation Styles in Explainable Hybrid Recommendation Motivational examples How to diversify explanation styles? Experimental results NonDiv ItemDiv RandDiv DualDiv Recall@50 0.115 0.111 Recall@10 0.190 0.173 Recall@20 0.311 0.275 AILD@50 0.680 0.753 AILD@10 0.715 0.776 AILD@20 0.745 0.804 EILD@50 0.391 0.433 0.644 0.731 EILD@10 0.454 0.480 0.632 0.699 EILD@20 0.517 0.537 0.628 0.686 WEILD@50 0.117 0.0985 0.159 0.184 WEILD@10 0.122 0.100 0.141 0.157 WEILD@20 0.128 0.0993 0.122 0.135  DualDiv can improve AILD without largely decreasing Recall.  DualDiv can greedily select the three most appropriate ESs regardless of the number of candidate ESs (cf. RandDiv).  DualDiv can successfully select dissimilar sets of ESs, especially between similar artists (cf. WEILD). National Institute of Advanced Industrial Science and Technology (AIST)Kosetsu Tsukuda and Masataka Goto Explanation: not diversified 1st rock The Beatles band Artist-based Popularity-based 2nd rock Muse band 3rd rock Queen band I want to listen to a pop artist now. rock band pop 90s pop punk I do not think artist-based and tag-based explanations are persuasive. rock band pop 90s pop Avril Lavigne punk I think a friend-based explanation is persuasive, so I listen to Avril Lavigne! By diversifying both the items and explanation styles, we aim to increase the probability of a user accepting at least one recommended item. ES We recommend The Beatles because: 𝑒𝑒𝑒 Last.fm users Anna, Tom, and John with whom you share similar tastes, listen to The Beatles. 𝑒𝑒𝑒 People who listen to your preferred artist Bob Dylan also listen to The Beatles. 𝑒𝑒𝑒 The Beatles has similar tags with your preferred artist The Kinks. 𝑒𝑒𝑒 According to Last.fm, artist The Beatles is similar to your preferred artist Queen. 𝑒𝑒𝑒 Artist The Beatles is tagged with rock and band that are in your profile. 𝑒𝑒𝑒 Your friend Lisa likes artist The Beatles. 𝑒𝑒𝑒 Artist The Beatles is very popular in the Last.fm database with 3.67 M listeners and 516.17 M playcounts. Each recommended artist has at most 7 personalized recommendations generated from explanation styles (ESs). rock band 𝑒𝑒𝑒 𝑒𝑒𝑒 𝑒𝑒𝑒 pop 90s 𝑒𝑒𝑒 𝑒𝑒𝑒 𝑒𝑒𝑒 pop punk ? ? ? 1st 2nd 3rd ESs for 𝑎𝑎3: 𝑒𝑒𝑒, 𝑒𝑒𝑒, 𝑒𝑒𝑒, 𝑒𝑒𝑒, 𝑒𝑒𝑒 𝑒𝑒𝑒 is selected because it is the only ES not included in either 𝑎𝑎𝑎 or 𝑎𝑎𝑎. STEP 01 pop 𝑎𝑎𝑎 punk ? ? 𝑒𝑒𝑒 is selected because it is included in only dissimilar artist 𝑎𝑎𝑎. STEP 02 pop 𝑎𝑎𝑎 punk ? 𝑒𝑒𝑒 𝑒𝑒𝑒 𝑒𝑒𝑒 𝑒𝑒𝑒 is selected because 𝑒𝑒𝑒 and 𝑒𝑒𝑒 are included in both artist 𝑎𝑎𝑎 and 𝑎𝑎𝑎. STEP 03 pop 𝑎𝑎𝑎 punk 𝑒𝑒𝑒 𝑒𝑒𝑒 𝑒𝑒𝑒 Our method generates a diversified ranked list of ESs for each artist by a greedy algorithm. Criterion: similar sets of ESs should not be displayed for similar artists. Metrics Results  AILD/EILD: a high value indicates that artists/ESs in the recommendation list are diversified well.  WEILD: a high value indicates that dissimilar sets of ESs are selected between similar artists. The Beatles Spice Girls Avril Lavigne The Beatles Spice Girls Artist-based Artist-based Popularity-based Artist-based Popularity-based Artist-based Artist-based Tag-based Tag-based Artist-based Popularity-based User-based Artist-based Friend-based Tag-based Item: not diversified Explanation: not diversified Item: diversified Explanation: diversified Item: diversified 𝑎𝑎𝑎 𝑎𝑎𝑎 𝑎𝑎𝑎

×