Merging Latent Factors and Tags to Increase Interactive Control of Recommendations Tim Donkers Benedikt Loepp Jürgen Ziegler University of Duisburg-Essen University of Duisburg-Essen University of Duisburg-Essen Duisburg, Germany Duisburg, Germany Duisburg, Germany tim.donkers@uni-due.de benedikt.loepp@uni- juergen.ziegler@uni- due.de due.de ABSTRACT 2. CONCEPT & PROTOTYPE We describe a novel approach that integrates user-generated A range of techniques is available for integrating supple- tags with standard Matrix Factorization to allow users to mental data into MF which has been shown to increase accu- interactively control recommendations. The tag informa- racy. These techniques, however, are typically very limited tion is incorporated during the learning phase and relates regarding user control. In addition, after being learned, the to the automatically derived latent factors. Thus, the sys- factors often exhibit no interpretable association with the tem can change an item’s score whenever the user adjusts a supplemental information, which thus cannot be accessed tag’s weight. We implemented a prototype and performed by the user. In contrast, in [1], the data is explicitly used to a user study showing that this seems to be a promising way establish a content-related association: Using a regression- for users to interactively manipulate the set of items recom- constrained formulation, the factors are considered as func- mended based on their user profile or in cold-start situations. tions of content attributes. We initially follow this approach closely for incorporating item-specific tag relevance infor- Categories and Subject Descriptors mation: For a set T of tags we define uA ∈ R|U |×|T | and i A ∈ R|I|×|T | representing their relationship with users U H.3.3 [Information Storage and Retrieval]: Information and items I, and redefine the original MF model: Search and Retrieval—information filtering, search process R ≈ PQT = uA uΘ( iA iΘ)T , (1) Keywords with uΘ and iΘ being the factor-tag matrices corresponding to users and items, respectively. However, explicit supple- Recommender Systems; Interactive Recommending; Matrix mental information may only be available either for users Factorization; Tags; User Interfaces or for items. Generally, we act on the assumption that tag- item relevance scores have been calculated separately and 1. INTRODUCTION i A is known a priori. Specifically, we exploit tag relevance Optimizing the objective accuracy of algorithms that gen- scores [5]: ait ∈ [0, 1] describes the extent to which tag t is erate recommendations has led to considerable advances, but relevant for item i. In contrast, the corresponding matrix does not necessarily increase user satisfaction [3]. Hence, let- for users, uA, is considered to be unknown. Thus, we treat ting users influence the recommendation process is increas- the whole term uA uΘ implicitly at this step by just learning ingly considered an important goal in Recommender Systems the user-factor matrix P as known from standard MF. With (RS) research. Interactive RS have been proposed that use this constrained equation, we formulate the minimization metadata such as user-provided tags for this purpose [5]. problem as in [1] and apply gradient descent. This has the advantage of using concepts that are mean- A user’s u (calculated) interest in a particular factor f is ingful to users without requiring explicit item descriptions. numerically expressed by entry puf of P while entry qif of Using tags to express user preferences thus seems promis- Q describes the extent to which item i possesses this factor. ing to improve user control and comprehension. However, Although in our case tag relevance scores are only known for attempts to increase interactivity (e. g. [5]) are typically in- items, we can establish a relation between users and tags as dependent of conventional Collaborative Filtering (CF) tech- well: Under the assumption that f reflects a certain charac- niques and consequently do not consider existing user pro- teristic which has the same semantic meaning for users and files based on e. g. previous ratings. Moreover, the availabil- items [4], we extend the approach of [1] and transfer the ity of precise and efficient algorithms such as Matrix Fac- learned relationship between tags and latent factors to the torization (MF) [4] is not exploited. What is lacking, thus, user side. In fact, we assume that the regression coefficients are techniques that combine the accuracy-related benefits stored in iΘ are equivalent to the implicitly assumed entries of model-based RS with the easy-to-understand semantics of the corresponding matrix uΘ, such that: uΘ = iΘ =: Θ. of tags. We therefore propose an interactive recommending Thus, according to (1) we solve for uA: approach that integrates latent factors derived by standard P = uAΘ ⇔ P = uAUΣVT ⇔ SVD-like MF with tags users provided for the items. (2) u A = PVΣ+ UT ⇔ uA = PΘ+ Since Θ is generally not a square matrix, we first calculate Copyright is held be the author(s). its pseudo-inverse Θ+ using SVD. Regarding the regression- RecSys 2015 Poster Proceedings, September 16-20, 2015, Vienna, Austria. constrained approximation of R in (1), this gives us: 3. EVALUATION & DISCUSSION u Ti T u T T Ti T We performed an evaluation using a standard SVD-like R ≈ AΘΘ A ≈ AUΣV VΣ U A u T Ti T T (3) MF algorithm1 as a baseline, and extended this algorithm ≈ AUΣΣ U A ≈ GΨH according to our approach considering a number of the most |U |×|T | popular tags as additional training data. We used the well- G∈R basically stores all vectors for the users and summarizes uAU. Conversely, H ∈ R|I|×|T | holds the item known MovieLens 10M dataset for ratings and the Movie- vectors. Ψ ∈ R|T |×|T | is a diagonal matrix containing posi- Lens Tag Genome dataset for tag-item relevance scores. Fig- tive eigenvalues of ΘΘT . The general interest of a certain ure 2 shows the results of one of our offline experiments. user regarding all tags is now expressed by vector au of uA, 0.85 which is basically the counterpart of the tag-item relevance 0.80 scores. Since they also comprise the latent factors, the gu RMSE vectors can then be used to generate recommendations. 0.75 The previously abstract user-factor and item-factor vec- 0.70 Standard SVD-like MF Tag-supported MF (20 Tags) tors can now both be accessed in a much more comprehensi- 0.65 Tag-supported MF (50 Tags) Tag-supported MF (100 Tags) ble way. The tag concept is easily understood by users and 5 10 15 20 can be used to actively adjust their own user vector, i. e. Number of latent factors their profile. In particular, users can influence the recom- Figure 2: RMSE for different configurations depend- mendations by searching, selecting and weighting tags, thus ing on the number of latent factors. indirectly determining their preferences in the latent factor space. A weight vector wu ∈ [0, 1]|T | therefore holds the user In line with others (e. g. [1]), it seems beneficial to include feedback regarding the tags, where 0 means no and 1 very metadata into MF. However, we also performed a user study strong interest in a particular tag. Integrating the weights with 46 participants (33 female; age: M = 22.89, σ = 6.88) into the calculation of recommendations leads to: who had to interact with our prototype RS in different con- r̃ui = (gu + αwu )Ψhi , (4) ditions, with and without tags. We used a questionnaire where α ∈ R represents the extent to which the weight in- comprising among others items from [2]. Results from our formation should be considered. r̃ui is a combination of the and other offline experiments could be confirmed, as subjec- user’s general preference structure gu , with the operational- tive perception of recommendation quality was higher with ization of the user’s current mood or interest wu . Initially, (M = 3.65, σ = 0.69) than without (M = 3.16, σ = 0.73) tags all values of wu are set to 0. When users start to interact (t(45) = −3.98, p < .001), also prior to interaction. Outlin- with the system by manipulating the values of wu , for exam- ing some further results, participants were also very satisfied ple by means of sliders, the resulting set of recommendations with the movie they finally selected from the recommenda- is continuously adapted in realtime. tions (M = 4.35, σ = 0.09) and stated a good usability (78 on SUS). In general, users liked the interaction via tags while perceiving the interaction effort to be acceptable (M = 3.64, σ = 0.74). In the tag condition, initial preferences were elicited by only selecting a small number of tags instead of rating items first. Since this led to particularly promising re- sults in terms of e. g. perceived recommendation quality, our tag-supported approach seems also to be useful in cold-start situations. In future work, we plan to evaluate the users’ perception of differences between conventional MF and our prototype integrating tags in more detail, and to exploit the integration of additional data more extensively. 4. REFERENCES [1] P. Forbes and M. Zhu. Content-boosted matrix factorization for recommender systems: Experiments with recipe recommendation. In Proc. RecSys ’11, pages 261–264. ACM, 2011. Figure 1: A user has selected and weighted the tags [2] B. P. Knijnenburg, M. C. Willemsen, and A. Kobsa. A “Sci-Fi” and “Action”, and therefore receives match- pragmatic procedure to support the user-centric ing movie recommendations from our prototype sys- evaluation of recommender systems. In Proc. RecSys tem such as “Matrix” or “Fight Club”. ’11, pages 321–324. ACM, 2011. [3] J. A. Konstan and J. Riedl. Recommender systems: Figure 1 shows a web-based prototype movie RS we have From algorithms to user experience. User Model. implemented to demonstrate this approach: At the top, an User-Adap., 22(1-2):101–123, 2012. area is shown where the users can place the tags they select [4] Y. Koren, R. M. Bell, and C. Volinsky. Matrix and adjust their weight by manipulating the sliders attached factorization techniques for recommender systems. to them. Users can also search for tags with the input field IEEE Computer, 42(8):30–37, 2009. underneath. Below, the system shows some suggested tags. [5] J. Vig, S. Sen, and J. Riedl. Navigating the tag Alongside each recommendation the three most relevant tags genome. In Proc. IUI ’11, pages 93–102. ACM, 2011. for this movie are shown. In addition, users could also rate 1 the recommended movies to further adapt their profile. Mahout ParallelSGDFactorizer (20 fact., 40 iter., λ = .001)