Subprofile-aware diversification of recommendations

Abstract

A user of a recommender system is more likely to be satisfied by one or more of the recommendations if each individual recommendation is relevant to her but additionally if the set of recommendations is diverse. The most common approach to recommendation diversification uses re-ranking: the recommender system scores a set of candidate items for relevance to the user; it then re-ranks the candidates so that the subset that it will recommend achieves a balance between relevance and diversity. Ordinarily, we expect a trade-off between relevance and diversity: the diversity of the set of recommendations increases by including items that have lower relevance scores but which are different from the items already in the set. In early work, the diversity of a set of recommendations was given by the average of their distances from one another, according to some semantic distance metric defined on item features such as movie genres. More recent intent-aware diversification methods formulate diversity in terms of coverage and relevance of aspects. The aspects are most commonly defined in terms of item features. By trying to ensure that the aspects of a set of recommended items cover the aspects of the items in the user’s profile, the level of diversity is more personalized. In offline experiments on pre-collected datasets, intent-aware diversification using item features as aspects sometimes defies the relevance/diversity trade-off: there are configurations in which the recommendations exhibits increases in both relevance and diversity. In this paper, we present a new form of intent-aware diversification, which we call SPAD (Subprofile-Aware Diversification), and a variant called RSPAD (Relevance-based SPAD). In SPAD, the aspects are not item features; they are subprofiles of the user’s profile. We present and compare a number of different ways to extract subprofiles from a user’s profile. None of them is defined in terms of item features. Therefore, SPAD is useful even in domains where item features are not available or are of low quality. On three pre-collected datasets from three different domains (movies, music artists and books), we compare SPAD and RSPAD to intent-aware methods in which aspects are item features. We find on these datasets that SPAD and RSPAD suffer even less from the relevance/diversity trade-off: across all three datasets, they increase both relevance and diversity for even more configurations than other approaches to diversification. Moreover, we find that SPAD and RSPAD are the most accurate systems across all three datasets.

Appendix: Hyper-parameters values

First, we will show the hyper-parameter values for the baseline recommender systems.

For pLSA, MF and FMBPR, we choose the number of latent factors (d) from $V = \{10, 30, 50, \ldots , 290, 310\}$. FMBPR’s learning rate (lr) and regularization parameters ($regW$ and $regM$) are chosen from $\{0.01, 0.001\}$, and MF’s confidence level ($\alpha$) is chosen from $\{1,2,\ldots ,10\}$. The values that get selected are as follows:

• pLSA: $d=50$ for MovieLens; $d=30$ for LastFM ; $d=270$ for LibraryThing.

• MF: $d=30, \alpha =1.0$ for MovieLens; $d=30, \alpha =1.0$ for LastFM; $d=330, \alpha =1.0$ for LibraryThing.

• FMBPR: $d=190$, $lr=0.01$, $regM =0.01$, $regW =0.001$ for MovieLens; $d=10$, $lr=0.01$, $regW =0.01$, $regM =0.001$ for LastFM; $d=270$, $lr=0.01$, $regM =0.01$, $regW =0.01$ for LibraryThing.

Second, we show the hyper-parameter values for the re-ranking and the subprofile detection methods.

All of the re-ranking approaches have hyper-parameter $\lambda$, which controls the balance between relevance and diversity (Eq. 1), whose value we select from $[0.1, 0.2, \ldots , 1.0]$.

For the subprofile detection methods, we select the values of $k_{ ind }$, $k_{ nn }$ and $k_{ IB }$ from V, and we select the value of cp from the set $[0.5, 0.6, \ldots , 1.0]$.

Table 13 shows the values that get selected.