information retrieval systems literature that ranks the documents in order of decreasing probability of relevance(Robertson 1997). Note that, by definition, recommending the most highly predicted items is designed to help improve recommendation accuracy, but not recommendation diversity. Therefore, new ranking criteria are needed in order to achieve diversity improvement. Recommending less popular items intuitively should have an effect towards increasing recommendation diversity, as illustrated by the recommendation ranking criterion and its impact on the recommendation accuracy and diversif erity as a example in Section 1. Following this motivation, we explore the possibility to use item popularity as a 3.1 Re-Ranking Recommendation List By Item Popularity We define item popularity as the number of known ratings for each item, and item popularity-based ranking approach( denoted as rankitemPop )recommends the least popular items to users. The results in Figure 1 show that the item popularity-based ranking approach increased recommendation diversity from 385 to 1395 (i.e, 3.6 times); however, recommendation accuracy dropped from 89% to 69%, as compared to the standard ranking approach. Here, despite the significant diversity gain, such a significant accuracy loss(20%)would not be acceptable in most real-life personalization applications. Therefore, in the next subsection we introduce a general technique to parameterize recommendation ranking approaches, which allows to achieve significant diversity gains while controlling accuracy losses(e.g, according to how much loss is tolerable in a given application) 3.2 Controlling Accuracy-Diversity Trade-off: Parameterized Ranking Approaches Item popularity-based ranking approach as well as other ranking approaches proposed in this paper are parameterized with" ranking threshold" TRETH, Tmax](where Tmax is the largest possible rating on the rating scale, e.g., Tmax=5) to provide user the ability to choose a certain level of recommendation rank(i) ifR(u i ETR, Tmax] rank(,TR) here (TR)=(ElIR (u, itRI, I autrankstandard (0 V R()e[TH, R) an= max rank、() Simply put, items that are predicted above ranking threshold TR are ranked according to rank(i), while tems that are below TR are ranked according to the standard ranking approach rankstandard (i). In addition, all items that are above TR get ranked ahead of all items that are below TR (as ensured by au in the above formal definition). Therefore, choosing different TR values in-between TH and Tmax allows the user to set the desired balance between accuracy and diversity, as shown in Figure 1. For example, the item opularity-based ranking approach with ranking threshold 4. 4 could minimize the accuracy loss to 1. 32% but still could obtain 83% diversity gain(from 385 as6 to 703), compared to the standard ranking approach An even higher threshold 4.7 still makes it possible 09 fAns =44 to achieve 20% diversity gain(from 385 to 462) even when there are less than N items above the &a with only 0.06%of accuracy loss. Also note that, TR=41 anking threshold T, by definition, all the items TR=3.8 above TR are recommended to a user, and the 3as remaining top-N items are selected according to the a standard ranking approach. This ensures that all the a7fi-kten opidat-bawedraing worth ranking approaches proposed in this paper provide the same exact number of recommendations as their ot t $oo "piersit in-Topf2100 13001500 corresponding baseline technique, which is also very important from the experimental analysis point of MovieLens data, top-5 items, item-based CE, 50 neighbors view in order to have a fair performance comparison of different ranking techniques Figure 1. Performance of item popularity-based approach with its parameterized versions 19th Workshop on Information Technologies and Systemsinformation retrieval systems literature that ranks the documents in order of decreasing probability of relevance (Robertson 1997). Note that, by definition, recommending the most highly predicted items is designed to help improve recommendation accuracy, but not recommendation diversity. Therefore, new ranking criteria are needed in order to achieve diversity improvement. Recommending less popular items intuitively should have an effect towards increasing recommendation diversity, as illustrated by the example in Section 1. Following this motivation, we explore the possibility to use item popularity as a recommendation ranking criterion and its impact on the recommendation accuracy and diversity. 3.1 Re-Ranking Recommendation List By Item Popularity We define item popularity as the number of known ratings for each item, and item popularity-based ranking approach (denoted as rankItemPop) recommends the least popular items to users. The results in Figure 1 show that the item popularity-based ranking approach increased recommendation diversity from 385 to 1395 (i.e., 3.6 times); however, recommendation accuracy dropped from 89% to 69%, as compared to the standard ranking approach. Here, despite the significant diversity gain, such a significant accuracy loss (20%) would not be acceptable in most real-life personalization applications. Therefore, in the next subsection we introduce a general technique to parameterize recommendation ranking approaches, which allows to achieve significant diversity gains while controlling accuracy losses (e.g., according to how much loss is tolerable in a given application). 3.2 Controlling Accuracy-Diversity Trade-Off: Parameterized Ranking Approaches Item popularity-based ranking approach as well as other ranking approaches proposed in this paper are parameterized with “ranking threshold” TR∈[TH, Tmax] (where Tmax is the largest possible rating on the rating scale, e.g., Tmax=5) to provide user the ability to choose a certain level of recommendation accuracy. In particular, given any ranking function rankX(i), the ranking threshold TR is used for creating the parameterized version of this ranking function, rankX(i, TR), which is formally defined as: [ ] Standard [ ) , max , * ( ), ( , ) (, ) * ( ), ( , ) T T R T T H R rank i if R u i x rank i T x R rank i if R u i αu ∈ = + ∈ ⎧⎪ ⎨ ⎪⎩ , where )( )( * * * max },),(|{)( i x rank TIi Ru u Ru TiuRIiTI R ∈ = ∈= ≥ α . Simply put, items that are predicted above ranking threshold TR are ranked according to rankX(i), while items that are below TR are ranked according to the standard ranking approach rankStandard(i). In addition, all items that are above TR get ranked ahead of all items that are below TR (as ensured by αu in the above formal definition). Therefore, choosing different TR values in-between TH and Tmax allows the user to set the desired balance between accuracy and diversity, as shown in Figure 1. For example, the item popularity-based ranking approach with ranking threshold 4.4 could minimize the accuracy loss to 1.32%, but still could obtain 83% diversity gain (from 385 to 703), compared to the standard ranking approach. An even higher threshold 4.7 still makes it possible to achieve 20% diversity gain (from 385 to 462) with only 0.06% of accuracy loss. Also note that, even when there are less than N items above the ranking threshold TR, by definition, all the items above TR are recommended to a user, and the remaining top-N items are selected according to the standard ranking approach. This ensures that all the ranking approaches proposed in this paper provide the same exact number of recommendations as their corresponding baseline technique, which is also very important from the experimental analysis point of view in order to have a fair performance comparison of different ranking techniques. MovieLens data, top-5 items, item-based CF, 50 neighbors Figure 1. Performance of item popularity-based approach with its parameterized versions 81 19th Workshop on Information Technologies and Systems