PReFacTO: Preference Relations Based Factor Model with Topic Awareness and Offset Priyanka Choudhary∗ Maunendra Sankar Desarkar Indian Institute of Technology Hyderabad Indian Institute of Technology Hyderabad Hyderabad, Telangana, India Hyderabad, Telangana, India cs16mtech11011@iith.ac.in maunendra@iith.ac.in ABSTRACT of rating data containing ratings given to the items by the users, Recommendation systems create personalized list of items that Recommendation Systems try to predict the ratings of the items might interest the user by analyzing the user’s history of past pur- that are not yet rated/viewed by the user. Based on these predicted chases and/or consumption. For rating based systems, most of the rating values, ranked list of the items that can be of user’s interest traditional methods for recommendation focus on the absolute rat- are recommended to the users. Latent factor models [1, 8, 13] have ings provided by the users to the items. In this paper, we extend the been extensively used in the past for this purpose. traditional Matrix Factorization approach for recommendation and There are lot of recommendation systems where the user feed- propose pairwise relation based factor modeling. While modeling back come in the form of ratings. Majority of such recommendation the items in the system, the use of pairwise preferences allow in- systems use these absolute ratings entered by the users for model- formation flow between the items through the preference relations ing the users and items according to latent factor modeling, and use as an additional information. Item feedbacks are available in the those models for recommendation. Latent factor models like Matrix form of reviews apart from the rating information. The reviews Factorization [8] are commonly used to transform or represent the have textual information that can be really helpful to represent users and the items to latent feature spaces. These representations the item’s latent feature vector appropriately. We perform topic are helpful for explaining the observed ratings and predicting the modeling of the item reviews and use the topic vectors to guide the unknown ratings. These latent factors, e.g. in case of movie rec- joint factor modeling of the users and items and learn their final ommendations, can be genres, actors or directors or something representations. The proposed method shows promising results in un-interpretable. These factors try to explain the aspects behind comparison to the state-of-the-art methods in our experiments. the liking of the items by a particular user. The items are modeled in a similar fashion by representation of the hidden factors possessed KEYWORDS by them. This representation predicts the rating by possession of these factors in an item and affinity of users towards these hidden Recommendation System, Pairwise Preferences, Topic Modeling, factors. Latent Factor Models User feedback in the form of reviews along with the ratings is also ACM Reference Format: available for many online systems like Amazon, IMDb, TripAdvisor Priyanka Choudhary and Maunendra Sankar Desarkar. 2018. PReFacTO: etc. The review information can be really useful as it contains the Preference Relations Based Factor Model with Topic Awareness and Off- users’ perception about the items. There can be systems where set. In Proceedings of ACM SIGIR Workshop on eCommerce (SIGIR 2018 the item description is also available. There are algorithms [14] eCom). ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/nnnnnnn. nnnnnnn which consider the item description as additional input for latent factor modeling. However, the descriptions are often entered by the item producers or sellers. On the other hand, the feedback in 1 INTRODUCTION the form of reviews given by the user generally conveys these Users have access to large variety of items available online for pur- factors that are being liked or disliked in an item. An attempt to chase, subscription, consumption etc. Such a huge list of options include these textual information can be helpful for better modeling, often result in choice overload, where it becomes difficult to browse interpretation and visualization of the hidden dimensions [11]. through and/or select the items of interest. Recommendation Sys- An alternate form of recommendation system can be based on tems (RS) make this task of selecting appropriate items easier by pairwise preferences of the user among the items [3, 4, 7]. Given a finding and suggesting subset of the items that might be of interest pair of items (i, j), user u may give feedback regarding which of the to the user. Many traditional recommendation techniques use only item he prefers over the other. Such type of feedback is referred to as ratings to assess the users’ taste and behavior. Given a small subset pairwise preference or pairwise preference based feedback. A survey ∗ This is the corresponding author. in [6] shows that users do prefer comparisons through pairwise scores rather than providing absolute ratings. Although there is Permission to make digital or hard copies of part or all of this work for personal or Copyright © 2018 by the paper’s authors. Copying permitted for private and academic purposes. no available dataset where the pairwise preferences were directly In: J. Degenhardt, classroom use is G. Di Fabbrizio, granted withoutS.fee Kallumadi, providedM.that Kumar, Y.-C. copies areLin, notA.made Trotman, H. Zhao or distributed captured, many approaches in literature have induced pairwise (eds.): Proceedings for profit of the SIGIR or commercial 2018 eCom advantage andworkshop, 12 July, that copies bear2018, Ann Arbor, this notice andMichigan, USA, the full citation published at http://ceur-ws.org on the first page. Copyrights for third-party components of this work must be honored. preferences from absolute ratings [3, 7, 10] and used those relations For all other uses, contact the owner/author(s). for developing algorithms for recommendation. SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA The existing methods from the literature that are based on pair- © 2018 Copyright held by the owner/author(s). ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. wise preferences do not consider the item content information in https://doi.org/10.1145/nnnnnnn.nnnnnnn SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA Priyanka Choudhary and Maunendra Sankar Desarkar the modeling process. In this paper, we propose approaches that to improve the rating predictions and provide feature discovery. combine the pairwise feedback with the additional review data Different users give different weights to these features. For e.g., a available. We propose an algorithm to use Latent Factor modeling user who loves horror movies and hates romantic genre will have using the pairwise preferences to discover the latent dimensions, high weightage to "Annabelle" movie than the "The Notebook" in map users and items to joint latent feature vector space and produce contrary to a romantic movie lover. This weightage will affect the recommendations for the end user. The latent feature vector space overall scores and explain the rating difference. for the items are derived through topic modeling. In this approach, Recently researchers have shown keen interest in pairwise pref- we construct a proxy document for each item by considering the erences based recommendation techniques. In [2] suitable graphical reviews that it has got. If available, the descriptions of the items also interface has been provided to the user to mark his choices over can be used to populate this document. We performed probabilistic the pair of items. In [7] the pairwise preferences are induced from topic modeling on these documents representing items using Latent the available rating values of the items. Both implicit [12] and ex- Dirichlet Allocation (LDA). These topics are then used to guide the plicit feedback can be modeled using the pairwise preferences based factorization process for learning the latent representations of the latent factor models. In [3], the users motivate the use of prefer- users. We propose two different approaches for this purpose. One ence relations or relative feedback for recommendation systems. in which the LDA topic vectors for the items are directly used as Pairwise preferences have been used in [3, 4, 7, 10] in matrix fac- the latent representations of the items, and another where these torization and nearest neighbor of latent factor modeling settings LDA representations are used to initialize the item vectors in the to generate recommendations. However, in none of these works, factorization process. For the second approach, the item-latent off- the user reviews are taken into account. set is introduced alongside the LDA representations. The offset is learned throughout the factorization process and tries to cap- ture the deviations from the LDA representations of the items. We 3 METHODOLOGY call our approach as Preference Relations Based Factor Model with In this section, we present our proposed recommendation methods Topic Awareness and Offset or PReFacTO in short. Experimental that work with pairwise preference information from the user. evaluation and analysis performed on a benchmark dataset helps Apart from the pairwise feedback, we also consider the reviews that to understand the strengths of the pairwise methods and their are provided by the users for different items. The methods represent ability to generate efficient recommendations. We summarize the each user and item in a shared latent feature space, through factor contribution of our work below: modeling approach. Before discussing our proposed methods in • We use relative preferences over item pairs in a factor mod- detail, we briefly describe the concepts of pairwise preferences and eling framework for modeling users and items. The models also about the way in which we handle the textual reviews available are then used for generating recommendations. for the items. • We incorporate item reviews in the factorization process. Pairwise Preferences: The ratings in recommendation systems • Detailed experimental evaluation is performed on a bench- are generally absolute in nature, often in the range of 1-5 or 1-10. mark dataset. Analysis of the results are performed to un- However, users have different behavior while rating the items. The derstand the advantages and shortcomings of the methods. same rating value entered by two different users might be due The rest of the paper is organized as follows. After discussing to two different satisfaction levels. Moreover, the absolute rating the related work, we present the proposed methods in Section 3. We entered by a user to an item may change over time, if the same user briefly talk about pairwise preferences and handling textual reviews is asked to rate the same item again. Motivated by observations and then provide detailed description about the methods being like this, pairwise preferences are introduced in modeling users proposed in this paper. In Section 4, we define the four evaluation and items in recommendation systems [3]. Pairwise relation based metrics used to measure the performance of the proposed methods approaches try to capture the relative preference between the items. with the baseline methods. We provide the detailed discussion and Such feedback, if directly obtained, removes the user bias that may analysis of the results obtained. The conclusion and the future work correspond to the leniency or strictness of the users while assigning of this paper has been summarized in Section 5. the absolute ratings. Although pairwise preference relations can address some of the 2 RELATED WORK problems with absolute ratings mentioned above, there is no dataset (publicly available) with directly obtained pairwise preferences. Traditional recommendation systems have extensively used latent In absence of such data, we consider in our work, datasets with factor based modeling techniques. Many researches have been done absolute ratings as user feedback, and induce relative ratings from that employ the use of Matrix Factorization(MF) [8, 9] techniques those absolute ratings. We then consider those relative pairwise for the prediction of unknown rating values of items not seen by the preferences as input to the proposed methods. user and providing recommendations by selecting top-N items. This Handling Textual Reviews - Topic modeling: If the item de- basic MF model which corresponds to the pointwise method used scriptions are available, then the system can identify more about the in this paper. It acts as a baseline model to compare the proposed attributes or aspects that the items possess. This information can methods presented in this paper. The works of [11, 14] have included be useful in making the recommendations. In fact, content-based the content based modeling to interpret the textual labels for the recommendation algorithms try to exploit these item attributes for rating dimensions. This justifies the reasons how the user assess generating the recommendations. the products. Similar kind of work has been done in [5]. It tries PReFacTO: Preference Relations Based Factor Model with Topic Awareness and Offset SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA Several systems allow the users to enter reviews for the items. Item reviews are very useful in making view/purchase decisions as they often contain reasons or explanations regarding why the item was liked or disliked by the user who wrote the review. The reviews often describe some additional details about the items, for example the aspects that they possess. An example review for a product from Amazon is given below. It seems like just about everybody has made a Christmas Carol movie. This one is the best by far! It seems more realistic than all the others and the time period seems to be perfect. The acting is also far better than any of the others I’ve seen; my opinion. Figure 1: Graph showing pairwise relation between the We hypothesize that even if item descriptions are not available, items as a function of sigmoid. then also, the reviews reveal a great deal of information about the different attributes (specified or latent) that might be contained in the items1 . These attributes can then be useful in modeling the was adopted in [4] for inducing pairwise preferences from absolute items, and can further aid in generating efficient recommendations. ratings. Based on this assumption, we use the reviews given by the users We take a different approach for converting the absolute rating to different items as an additional source of information for learning to relative preferences. If the ratings given by user u to the two the item representations. We use Latent Dirichlet Allocation (LDA) items i and j are rui and ru j respectively, then we define the (actual topic modeling technique to learn the topic representation of the or ground truth) preference strength for the triplet (u, i, j) as items. LDA is an unsupervised method, which, given a document exp (rui ) collection, identifies a fixed-number (say k, input to the algorithm) rui j = exp (rui ) + exp (ru j ) of latent topics present in the collection. Each document can then (1) 1 be represented as a k-dimensional vector in that topic space. LDA = 1 + exp (−(rui − ru j )) works on documents, so we need to represent each item as a docu- ment. For that purpose, we combine all the reviews assigned to an The value of rui j thus obtained can capture the strength of the item to create a proxy document for that item. If dui represents a preference relation as well. If the difference between rui and ru j review given by a user u for an item i, then we denote the proxy becomes larger, then the strength of this relation becomes stronger document (di ) for the item i as the concatenation of all the reviews as shown in Figure 1. given by the set of users U for i. Then, we can have a document We model the prediction of the unobserved rui j ’s as: collection d corresponding to the set of items I as d = ∪i (di ) where i = 1, · · · , |I |. exp (pu (qi − q j ) + (bi − b j )) rˆui j = 1 + exp (pu (qi − q j ) + (bi − b j )) 3.1 Preference Relation based Factor modeling (2) 1 = (Pairwise) 1 + exp (−(pu (qi − q j ) + (bi − b j )))) Between the pair of items (i, j), users can express their relative where the rating matrix R consisting of user-item interaction preference if such a provision exists. This would allow the user to gives access to the values of rui , indicating the rating given to indicate, for the item pair, which item he prefers more. The user item i by user u. The quantity bi represents the bias for the item. can also indicate if he favors both the items equally. The method models each user u by a vector pu . This vector space This pairwise preference can be captured through an interface measures user’s interest in the particular item based on affinity of where users mark their preferences over a small subset of data. user towards these factors. Similarly, each item i is represented by However, as mentioned earlier, we are not aware of the existence a feature vector qi . This latent factor representation defines the of any such system that allows the users to enter the pairwise degree to which these factors are possessed by the item. preferences directly. In absence of that, if the rating data is available, Given the training set, the mean-squared error (MSE) function pairwise preferences can be obtained as: rui j = rui − ru j . Here, rui on the training data (with suitable regularization) is used as the indicates the absolute rating given by user u to item i. If the sign of objective function. The optimization is generally performed using rui j is positive, we may consider that item i is preferred over item Stochastic Gradient Descent (SGD) and the algorithm outputs op- j by the user u. If the sign is negative we may consider that j is timized values of the rating parameters Θ = {B, P, Q } where B preferred over i. If the value of ru i j is zero, then it indicates that represents the bias values (bi ) for all the items i ∈ I , P represents both the items are equally preferable to u. Similar kind of approach the user latent feature vector (pu ) for all the users u ∈ U and Q 1 The dataset used in our experiments did not have the item descriptions, but contained represents the item latent feature vector (qi ) for all the items i ∈ I . the reviews The objective function is defined as : SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA Priyanka Choudhary and Maunendra Sankar Desarkar by Equation 10. The optimization variables (parameters) now be- come Θ = {B, P }. The solution to this objective function is obtained si j λ λ Õ  2 min rui j − + λ||pu || 2 + ||qi || 2 + ||q j || 2 + through Stochastic Gradient Descent. Θ 1 + si j 2 2 (u,i, j)∈T λ λ  si j 2 ||bi || 2 + ||b j || 2 Õ (3) min rui j − + λ||pu || 2 + 2 2 Θ 1 + si j (u,i, j)∈T where si j = exp (pu (qi − q j ) + (bi − b j )) λ λ ||bi || 2 + ||b j || 2 (10) 2 2 T represents the training set and λ is the regularization parameter. Here qi remains fixed throughout the learning process. Hence, we The update rules for optimizing the above objective function are do not have regularization term for qi in the objective function. given below: The update rules remain same for pu , bi and b j as in Equation 4, Update rules : 7 and 8 respectively. Personalized utility scores of the items are 2eui j si j (qi − q j )   pu ← pu + α − 2λpu (4) computed using Equation 9 and recommendations are generated. (1 + si j )2  2eui j si j pu  3.3 Pairwise Relation based Factor modeling qi ← qi + α − λq i (5) with Topics and Offset (PreFacTO) (1 + si j )2 In the previous method described in Section 3.2, the topic modeling 2eui j si j pu   qj ← qj − α + λq j (6) provides the seed information for the item latent vector representa- (1 + si j )2 tions obtained from the reviews. These representations were fixed 2eui j si j   throughout the learning process. In our next method, we allow bi ← bi + α − λb i (7) the item representations to take deviations from their LDA topic (1 + si j )2 vectors. If ϵi is the deviation of the item i’s representation from its 2eui j si j   bj ← bj − α + λb (8) topic vector qi , then the pairwise ratings can be modeled as: j (1 + si j )2 exp (pu ((qi + ϵi ) − (q j + ϵ j )) + (bi − b j )) s rˆui j = where eui j = rui j − (1+si j ) and α is the learning rate. 1 + exp (pu ((qi + ϵi ) − (q j + ϵ j )) + (bi − b j )) ij (11) After obtaining the model parameters through stochastic gradi- 1 ent descent, we can predict the personalized utility of the item i for = 1 + exp (−(pu ((qi + ϵi ) − (q j + ϵ j )) + (bi − b j )))) the user u as: The parameters for this model are Θ = {B, P, E}. As earlier, B ρui = bi + pu qi (9) and P are the collection of item-bias vectors and user vectors. E is The top-N items according to this predicted personalized utility the collection of deviations or offsets of the items from their LDA are recommended to the user. topic vectors. The objective function for this model can be written as: 3.2 Preference Relation based Factor modeling with Topics (Pairwise+Topic) si j λ λ Õ  2 min rui j − + λ||pu || 2 + ||bi || 2 + ||b j || 2 + As motivated in the previous section, the review comments about Θ 1 + si j 2 2 items can be useful in identifying the aspects that the items pos- (u,i, j)∈T sess. Moreover, it also helps to understand the reasons behind the λ λ ||ϵi || 2 + ||ϵ j || 2 (12) liking or disliking of the item by the user. Hence, we extend the 2 2 previous method to incorporate the reviews about the items. The where si j = exp (pu (qi + ϵi ) − (q j + ϵ j )) + (bi − b j )) and rui j is al- proxy documents for the items are passed through a Latent Dirichlet ready defined in Equation 1. Allocation (LDA) framework to identify the latent topics present in The model parameters are learned using Stochastic Gradient the documents. Descent. The update rules are given below: LDA is a probabilistic topic modeling technique that discovers 2eui j si j ((qi + ϵi ) − (q j + ϵ j ))   latent topics in the documents. It represents each document di by pu ← pu + α − 2λpu (13) (1 + si j )2 k-dimensional topic distribution θ i through Dirichlet distribution. The k-th dimension of the vector indicates the probability with 2eui j si j pu   ϵi ← ϵi + α − λϵ i (14) which the k-th topic is being discussed in the document. Each topic (1 + si j )2 is associated with the word distribution ϕ k which is the probability  2eui j si j pu  of the word-topic association. ϵj ← ϵj − α + λϵ j (15) (1 + si j )2 We pass the collection of documents D = ∪i ∈I di to LDA. As an s output, we get the topic vector qi corresponding to each document where eui j = rui j − (1+si j ) . ij di ∈ D. For each item i, the latent representation is now fixed at qi , The update rules for the bias terms remain same as specified in and these values of qi ’s are fed to the factor modeling technique Equations 7 and 8. After the optimized values of the parameters are used in Section 3.1. The objective function for this method is given obtained, personalized utility of the item i for user u is computed PReFacTO: Preference Relations Based Factor Model with Topic Awareness and Offset SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA using following equation and Top-N recommendations are made 4.3 Evaluation for each user. For evaluation of the models presented in Section 3, we compare ρui = bi + pu (qi + ϵi ) (16) those three algorithms with the baseline methods mentioned in Section 4.2. We use Precision@k, Recall@k, IRecall and URecall 4 PERFORMANCE EVALUATION as the evaluation metrics. We took k = 100. The IRecall and the 4.1 Dataset URecall metrics are described below. We use the Amazon product review dataset2 for our experiments. IRecall: IRecall of an item is computed using the following equa- This dataset contains reviews and ratings given to different items by tion: different users. We consider items from the Movies and TV category. |Rec(i) ∩ Rated(i)| All items in this category were released between 1999 to 2013. We IRecalli = , (17) |Rated(i)| divided this timeline into three blocks each consisting of 5 year where Rec(i) denotes the sets of users to whom item i is recom- span: (A) 1999-2003, (B) 2004-2008, and (C) 2009-2013. From each mended. Rated(i) denotes the set of users who have i in their test block, we removed the items which have less than 10 reviews in set. Thus this metric measures the algorithm’s ability to recom- that block and the users who have given less than 5 reviews in that mend items to the users who have actually rated it. IRecall for an block. After this filtering to remove these non-prolific users and algorithm is defined as the average of the item-wise IRecall values items, we have 3,513 items, 85,375 users, 725198 ratings and 725176 over the set of concerned items. reviews in our dataset. We have used 70% of this data for training URecall: URecall of a user is computed as: and the remaining 30% for testing purposes. |Rec(u) ∩ Rated(u)| U Recallu = , (18) 4.2 Baseline Methods |Rated(u)| We compare our preference relation based models to the following where Rec(u) denotes the sets of items that have been recommended baselines: to user u. Rated(u) denotes the set of items present in the test set of user u. (a) Absolute Rating based Factor modeling (Pointwise): For the experimentation and evaluation purposes, we have di- In analogous to the standard latent-model [8], we convert vided the items into bins. These bins are created based on the the absolute rating values using the sigmoid function. The number of reviews. For each block, we maintain item review count sigmoid function is then used to make predictions using the written by the user during that time span (block range). We define following objective function: two bins for each block as follows: Bin-0 consists of the items hav- ing review count less than 40 and Bin-1 contains the items having si λ λ Õ  2 min ρui − + λ||pu || 2 + ||qi || 2 + ||bi || 2 review count greater than or equal to 40. We consider the Bin-0 as Θ 1 + si 2 2 a collection of sparse items, and the items from Bin-1 as dense items. (u,i)∈T For each bin, we compute the average of the IRecall value of all where exp (rui ) the items present in the corresponding bin. Analogous to the items, ρui = 1 + exp (rui ) we divide the users as well into the bins based on the number of reviews given by the user. Also, we take the average of the URecall si = exp (pu qi + bi ) value of all the users falling into the corresponding bin. We then (b) Absolute Rating based Factor modeling with Topics compare the IRecall and URecall values of the different methods (Pointwise+Topics) : We combine the topic modeling mentioned in this paper with the baseline approaches. technique with the latent factor modeling. The latent vector representations of the items are drawn from the reviews (by 4.4 Experimental Analysis And Discussion passing the reviews of the items as an input to the LDA) and Setting the parameters for the proposed method: The model fed to latent factor model. Here the item representations will hyperparameters λ (regularization parameter) and k (number of remain fixed and the user-latent space will be learned using topics) need to be determined in order to produce best models the Stochastic Gradient Descent. for recommendation. Experiments were conducted with different (c) Absolute Rating based Factor modeling with Topics values of λ and k on a small subset of the data. From the experiments, And Offset (Pointwise+Topics+Offset) : Along with the combination of λ = 4E − 05 and k = 10 were found to be the the factor and the topic modeling, we introduce item latent best values for the parameters. Hence, we select these two values vector offset which captures the deviations of the item feature of the hyperparameters for further experimentation. Performance vector representations drawn from the LDA. The objective of the algorithm on the test set for different values of λ (keeping k function to model the system and learn the user-latent and fixed at 10) and different values of k (keeping λ fixed at 4E − 05 are the item-offset representations can be written as: shown in Table 1 and Table 2 respectively. si λ λ Õ  2 Comparison with other methods and discussion: For each min ρui − + λ||pu || 2 + ||bi || 2 + ||ϵi || 2 Θ 1 + si 2 2 method, we run the experiments for the three blocks, and compute (u,i)∈T the average value of each metric over these three blocks. These where si = exp (pu (qi + ϵi ) + bi ) average values are reported in Table 3. It can be seen from the exper- 2 http://jmcauley.ucsd.edu/data/amazon/ imental results that pairwise methods and in particular, PreFacTO SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA Priyanka Choudhary and Maunendra Sankar Desarkar Table 1: Values of the evaluation metrics for different values of λ. Number of topics were fixed at 10. Alpha Precision Recall IRecall(reviews<40) IRecall(reviews>40) URecall(reviews<40) URecall(reviews>40) 4.00E-02 0.0076 0.1045 0.0117 0.0673 0.1074 0.0863 4.00E-03 0.0122 0.1451 0.0013 0.0793 0.1448 0.1456 4.00E-04 0.0120 0.1398 0.0012 0.0789 0.1390 0.1435 4.00E-05 0.0125 0.1457 0.0012 0.0792 0.1448 0.1504 4.00E-06 0.0124 0.1448 0.0011 0.0797 0.1438 0.1495 Table 2: Values of the evaluation metrics for different values of k: the number of topics. The value of λ was fixed at 4.00E − 05. No. of Topics Precision Recall IRecall(reviews<40) IRecall(reviews>40) URecall(reviews<40) URecall(reviews>40) 5 0.0107 0.1229 0.0008 0.0781 0.1221 0.1302 10 0.0125 0.1457 0.0012 0.0792 0.1448 0.1504 15 0.0108 0.1246 0.0011 0.0778 0.1238 0.1324 20 0.0108 0.1244 0.0008 0.0784 0.1233 0.1331 Table 3: Comparing performances of different algorithms. The best values for each metric across the algorithms are marked in bold. Method Precision Recall IRecall(reviews<40) IRecall(reviews>40) URecall(reviews<40) URecall(reviews>40) Pointwise 0.0106 0.1267 0.0141 0.0635 0.1271 0.1210 Pointwise+Topics 0.0048 0.0555 0.0256 0.0551 0.0551 0.0568 Pointwise+Topics+Offset 0.0055 0.0650 0.0252 0.0514 0.0651 0.0632 Pairwise 0.0021 0.0254 0.0420 0.0312 0.0255 0.0252 Pairwise+Topics 0.0038 0.0485 0.0378 0.0399 0.0491 0.0448 PreFacTO 0.0125 0.1457 0.0012 0.0792 0.1448 0.1504 0.08 0.14 0.07 0.12 0.06 0.1 0.05 0.08 0.04 0.06 0.03 0.04 0.02 0.02 0.01 0 0 Block-1 Block-2 Block-3 Average Block-1 Block-2 Block-3 Average Pointwise Pairwise Pointwise Pairwise Pointwise+Topic Pairwise+Topic Pointwise+Topic Pairwise+Topic Pointwise+Topic+Offset PreFacTO Pointwise+Topic+Offset PreFacTO Figure 2: Comparison of IRecall values of different Figure 3: Comparison of IRecall values of different algo- algorithms taking into consideration the items having re- rithms with review count of the items greater than or equal view count less than 40. to 40. gives the best results compared to other algorithms for the com- method outperforms all other approaches. The IRecall values for plete dataset. Although the PreFacTO and pointwise are at par dense items shows that PreFacTO performs very well for dense based on their performance but the PreFacTO slightly surpasses items. The IRecall values for the sparse and dense items for different the pointwise in terms of overall precision and recall values. If blocks are compared in Figure 2 and Figure 3 respectively. There are we compare the IRecall values for the sparse items, the Pairwise four groups of columns in both the figures. The first three represent PReFacTO: Preference Relations Based Factor Model with Topic Awareness and Offset SIGIR 2018 eCom, July 2018, Ann Arbor, Michigan, USA the three blocks, and the last one represents the average value over Pairwise and PreFacTo and fuse the recommendations generated the three blocks. by them from sparse and dense zones to come up with the final The superior performance of Pairwise and worst performance of recommendations. It might also be possible to develop parame- PreFacTO in case of the sparse items might be due to the sparseness terized algorithms that automatically switch between Pairwise of the reviews. The LDA representation for the sparse items having (no consideration of reviews) and PreFacTo (considering the re- very less reviews and further learning in the form of deviations on views) depending on the availability of data for the item under top of the LDA vectors do not provide any additional benefit. On consideration during the modeling. the contrary, it might have led to overfitting. But on the other hand, Pairwise tries to model the system only through rating information. REFERENCES The preference relations provide some additional information to [1] Deepak Agarwal and Bee-Chung Chen. 2009. Regression-based latent factor models. In Proceedings of the 15th ACM SIGKDD international conference on the item in the process of comparing it with the other items. There Knowledge discovery and data mining. ACM, 19–28. is no overfitting in the process and modeling the system for the [2] Laura Blédaité and Francesco Ricci. 2015. Pairwise preferences elicitation and sparse items works well. If we look at URecall values for the sparse exploitation for conversational collaborative filtering. In Proceedings of the 26th ACM Conference on Hypertext & Social Media. ACM, 231–236. users, the PreFacTO actually performs well. [3] Maunendra Sankar Desarkar, Sudeshna Sarkar, and Pabitra Mitra. 2010. Aggre- However, in case of dense items, the PreFacTO outperforms gating Preference Graphs for Collaborative Rating Prediction. In Proceedings of every other method including Pointwise. Along with the pairwise the Fourth ACM Conference on Recommender Systems (RecSys ’10). ACM, New York, NY, USA, 21–28. https://doi.org/10.1145/1864708.1864716 preference based learning, the item vector representation from the [4] Maunendra Sankar Desarkar, Roopam Saxena, and Sudeshna Sarkar. 2012. Pref- rich-textual information of the reviews and learning the deviations erence relation based matrix factorization for recommender systems. In Inter- national conference on user modeling, adaptation, and personalization. Springer, from these item vectors help in better prediction with reasoning as 63–75. to why the item will be likeable or dislikable to the user. [5] Gayatree Ganu, Noemie Elhadad, and Amélie Marian. 2009. Beyond the stars: In any real recommendation system, there are sparse items, and improving rating predictions using review text content.. In WebDB, Vol. 9. Citeseer, 1–6. there are dense items as well. Depending on the exact system or [6] Nicolas Jones, Armelle Brun, and Anne Boyer. 2011. Comparisons instead of rat- domain, the ratio of sparse to dense items can vary. In this study, ings: Towards more stable preferences. In Proceedings of the 2011 IEEE/WIC/ACM we have explored few algorithms that consider pairwise feedback International Conferences on Web Intelligence and Intelligent Agent Technology- Volume 01. IEEE Computer Society, 451–456. instead of absolute ratings. Among the proposed methods, Pairwise [7] Saikishore Kalloori, Francesco Ricci, and Marko Tkalcic. 2016. Pairwise prefer- works well for the sparse items and PreFacTO works well for the ences based matrix factorization and nearest neighbor recommendation tech- niques. In Proceedings of the 10th ACM Conference on Recommender Systems. ACM, dense items. The experiments show the power of preference relation 143–146. based feedback for recommendation. However, it does not establish [8] Robert Bell Koren, Yehuda and Chris Volinsky. 2009. Matrix factorization tech- the superiority of any single algorithm that works across the entire niques for recommender systems. Computer (Long Beach, Calif.) 42 (Aug. 2009), 30–37. https://doi.org/10.1109/MC.2009.263 range of data (both sparse and dense zones). Nonetheless, we believe [9] Yehuda Koren and Robert Bell. 2015. Advances in collaborative filtering. In that it might be possible to design such algorithms that works well Recommender systems handbook. Springer, 77–118. for the entire range of data. It might be an interesting research [10] Shaowu Liu, Gang Li, Truyen Tran, and Yuan Jiang. 2016. Preference Relation- based Markov Random Fields for Recommender Systems. In Asian Conference on direction to develop hybrid methods that consider both Pairwise Machine Learning (Proceedings of Machine Learning Research), Geoffrey Holmes and PreFacTo for fusing the recommendations from sparse and and Tie-Yan Liu (Eds.), Vol. 45. PMLR, Hong Kong, 157–172. http://proceedings. mlr.press/v45/Liu15.html dense zones to generate the final recommendations. [11] Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. Proceedings of the 7th ACM conference on Recommender systems. ACM, 2013 (Oct. 2013), 165–172. https: 5 CONCLUSIONS AND FUTURE WORK //doi.org/10.1145/2507157.2507163 [12] Ladislav Peska and Peter Vojtas. 2017. Using implicit preference relations to We have presented the PReFacTO approach in this paper, which improve recommender systems. Journal on Data Semantics 6, 1 (2017), 15–30. aligns the latent factor modeling between the users and the item [13] Ruslan Salakhutdinov and Andriy Mnih. 2008. Bayesian probabilistic matrix pairs with the hidden topics in the reviews of the item. The pair- factorization using Markov chain Monte Carlo. In Proceedings of the 25th interna- tional conference on Machine learning. ACM, 880–887. wise relation adds significant information for the sparse items and [14] Chong Wang and David M Blei. 2011. Collaborative topic modeling for recom- provides better modeling of the user-item interaction, and the item mending scientific articles. In Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 448–456. hidden dimensions are effectively drawn from the reviews. The topic modeling based latent factors of the items along with the pairwise relation between these items (where the latent feature space of the items drawn from the LDA are allowed to change through offset during the learning process) provides significant improvement over the methods considered in isolation. Our algo- rithm runs very effectively on large dataset and comparable with the pointwise approach. In fact, PreFacTO method gives marginal improvements over the pointwise methods. It is also shown that Pairwise method works well for the sparse items and PreFacTO provides better performance in case of dense items. It was observed in the experimental results that Pairwise works well for sparse items and PreFacTO works well for dense items. It might be possible to develop hybrid methods that consider both