In this paper we propose a new method of recommending not only items of interest to the user but also the conditions enhancing user experiences with those items, such as recommending to go to a restaurant for seafood. This method is based on the sentiment analysis of user reviews, predicts sentiments that the user might express about the aspects determined in an application, and identi es the most valuable aspects of user's potential experience with the item. Furthermore, our method recommends the items together with those most important aspects over which the user has control and can potentially select them, such as the time to go to a restaurant, e.g. lunch vs. dinner, or what to have there, such as seafood. We tested our method on three applications (restaurants, hotels and beauty&spas) and experimentally showed that those users who followed our recommendations of items with their corresponding conditions had better experiences, as de ned by the overall rating, than others.
Over the last ve years, there has been much interest in
leveraging user reviews for providing personalized
recommendations based on these reviews [
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
CBRecSys 2016, September 16, 2016, Boston, MA, USA.
Copyright remains with the authors.
These approaches aimed at predicting and explaining ratings in terms of the user and item characteristics without taking into the consideration additional factors, such as circumstances of consuming the item and personal choices of the user. For example, consider the user choosing between doing a manicure or a haircut in a salon. Depending on what the user would choose to do during her visit, she can give di erent ratings to the salon. Therefore, user experience of a particular item can be improved by recommending some appropriate conditions of consuming that item, such as doing a pedicure in that speci c salon.
In this paper, we address this issue by recommending not only particular items, but also the conditions (i.e. the most important aspects) of potential experiences that the user can control, such as having haircut or pedicure in a salon. Furthermore, we can recommend certain actions to the management of an establishment (item) that would personalize experiences of the user when consuming the item (e.g., visiting the establishment). For example, we may recommend to the management of the Gotham Bar & Grill restaurant in New York to suggest the duck dish to a user because our method estimated that the user would particularly like that dish in that restaurant.
In this paper, we make the following contributions. First, we propose a novel approach of enhancing functionality of recommender systems by recommending not only the item itself but the item with the corresponding conditions enhancing user experience of the item. Further, we propose that the management of an establishment reviews the crucial aspects of the user's experience of the establishment that were identi ed by our method and uses them to personalize user experiences, thus improving the overall user ratings. Second, we developed a method for identifying the most valuable aspects of future experiences of the users that is based on the sentiment analysis of user reviews. Third, we tested our method on actual reviews and showed that users who followed our recommendations rate their experiences signi cantly higher than others. 2.
Over the last few years, several papers tried to improve
estimation of unknown ratings by using user reviews [
In contrast to these papers, in our work, we focus not only on rating predictions but also on determining the most important aspects having the highest impact on ratings measuring user's potential future experience of an item. Moreover, we provide recommendations not only of items of interest to the user but also the conditions enhancing user experiences with those items.
Furthermore, [
Finally, there are a few publications on explanations of
recommendations based on user reviews. In particular, [
In this section we present a method of identifying the most valuable aspects of future user experiences of the items and recommend the items together with the most important aspects over which the user has control and that she can potentially select. This method consists of sentiment analysis of user reviews, identi cation of relations between di erent aspects of the reviews, predicting sentiments that the user might express about these aspects, and calculating personal impact factors that each aspect contributes to the overall rating for the user. Moreover, we identify and recommend to the management of an establishment to consider potentially important aspects of the user experience with the establishment and use them to personalize user experiences there, thus improving the overall user rating of the establishment. The proposed method consists of 5 steps described below. (1) Extracting aspects from the reviews.
In this step we utilized the state-of-the-art
\industrialstrength" sentiment analysis system Opinion Parser [
We applied Opinion Parser to the set of reviews R for a given application (e.g. restaurants) in order to build a set of aspects A occurring in R. Furthermore, for each review r 2 R; Opinion Parser identi es a set of aspects Ar occurt ring in r and corresponding sentiments sui opinions of user u about aspects t 2 Ar of experience with item i. (2) Training the sentiment predicting model.
In a typical review, set Ar contains only a small part of all the aspects A, and thus for each particular aspect the matrix of known sentiments for (user, item) pairs is more sparse than matrix of ratings. Therefore, the problem of predicting sentiments of the aspects becomes even more challenging than prediction of the overall rating of review r.
Since various aspects of a review can be correlated, such as \desert" and \fruits", we propose an approach of using this correlation information to train the sentiment prediction model. In particular, in review r for each aspect t 2 fA t Arg we estimate its (unknown) sentiment sui as a weighted average of the explicitly speci ed sentiments of its k nearest neighbors. More formally, s^tui =
P1k wtj sui ;
j
P1k wtj where the weights wtj are computed as Spearman correlations between aspects t and j. In case if the user did not express sentiments about a certain number of the correlated aspects, we leave the sentiment value of the aspect t blank.
Next, for each aspect t 2 A we train the Matrix Factorization model in the following form: s^tui =
t + btu + bit + ptu qit: This approach takes into account user's personal preferences, items individual characteristics and their interaction for each particular aspect t 2 A: (3) Building regression model to predict ratings.
To build the regression model for predicting ratings, we rst ll in the remaining missing values of sentiments for those aspects t 2 fA Arg that were left blank in Step 2. We ll it with the average values of the aspects' sentiments avg(stui) taken across explicitly speci ed values in R. Note, that if the user did not mention the aspect in the review, we assume that she had the average satisfaction of the aspect and it did not a ect her overall rating signi cantly.
We next build the regression model predicting rating rui of the review based on the estimated sentiments s^tui for aspects from A. More speci cally we estimate the overall rating with the regression model in the following form: rui = (A + Bu + Ci) Sui (1) where, A = (a0; : : : ; an) is a vector of general coe cients, Bu = (b0u; : : : ; bun) is a vector of coe cients pertaining to user u, Ci = (ci0; : : : ; cin) is a vector of coe cients pertaining to item i, and Sui = (s^0ui; : : : ; s^uni) is a vector of estimated values of sentiments corresponding to aspects from A in the particular review. Further, we avoid over- tting by using the regularized model:
min a ;b ;c
X (rui rui)2+ (u;i)2R n X t=0 at2 + X(btu)2 + X(cit)2 u i !!
As it follows from Eq.(1), the proposed model estimates individual preferences of user Bu and the individual characteristics of item Ci. (4) Calculating impacts of aspects on rating.
In this step we apply the models built in Steps 2 and 3 for determining most important aspects of user's potential experiences with the establishment, where the importance of an aspect is determined by its weight in the regression model described in Step 3. More speci cally, for a new potential review we predict sentiment stui of each of the aspects t 2 A. Then we compute the impact of each sentiment being potentially explicitly expressed in the review as follows. By construction, our regression model takes an average sentiment value in case when the user did not express her opinion. Therefore, we calculate the aspects' impact as product of the di erence between the predicted value of sentiment and the average sentiment value for the particular aspect, with the corresponding coe cient in the regression model (Eq.(1)): impacttui = (at + btu + cit) (stui avg(stui)): (2)
These calculated impacts re ect the importance of each aspect on the overall rating and they might be positive or negative. (5) Recommending items and conditions.
Next, we identify two groups of aspects in A over which (a) the user has control and (b) the management of the establishment has control, in order to recommend most important of them in both cases. Furthermore, we identify the conditions that we want to recommend to the user or the management together with the item, where the condition is a suggestion to experience (positive) or not to experience (negative) a particular aspect. Finally, we recommend an item and its corresponding conditions to the user, or the most important conditions to the management.
For example, if our system identi ed aspect \ sh" as having high positive impact on the rating, we will recommend this restaurant and the condition of ordering sh in that restaurant to the user. Similarly, if aspect \desert" has strong negative impact on the rating, we may still recommend to the user to visit that restaurant under the condition not to order desert there. Similarly, we can recommend such conditions to the management. For example, we can recommend to the management of a health&beauty salon to provide a complementary drink to the user (since it will improve her overall experience) and don't chat to her too much while in session.
In summary, we proposed a method for identifying and recommending the most valuable aspects of potential user experiences of the items. This method consists of sentiment analysis of user reviews, identi cation of relations between di erent aspects of the reviews, predicting sentiments that the user might express about these aspects, and calculating personal impact factors that each aspect contributes to the overall rating for the user. In Section 4, we show the results of applying the proposed method to real data from three applications.
To demonstrate how well our method works in practice, we tested it on the Yelp dataset1 for restaurants, hotels and 1https://www.yelp.com/dataset challenge/dataset Meat beef meat bbq ribs veal pork
Fish cod salmon cat sh tuna shark sh
Dessert tiramisu cheesecake chocolate dessert ice cream macaroons
Money price dollars
cost budget charge check
Service bartender waiter service hostess manager sta
Decor design ceiling decor lounge window space beauty&spas applications for the reviews collected in several US cities over a period of 6 years. In this study we used the reviews of 24,917 restaurants produced by 384,821 users (1,344,405 reviews in total), 1,424 hotels produced by 65,387 users (96,384 reviews) and for the 6,536 beauty&spas produced by 71,422 users (104,199 reviews in total).
We applied the 5-step method presented in Section 3 to this Yelp data. As a result, we managed to extract 68 aspects for restaurants, 44 aspects for hotels, and 45 aspects for beauty&spas applications in Step 1 of our method using Opinion Parser. Table 1 presents several aspects pertaining to the restaurant application with examples of corresponding words.
Further, the set of reviews R in each application is partitioned into train and test sets in the ratio of 80% to 20%. After determining the sets of aspects in the reviews and aggregating their sentiments as described in Step 1, we lled in the missing sentiment values and trained the sentiment prediction models for each aspect, as described in Step 2.
We compared the performance of our approach with the standard MF models built only on the sets of explicitly expressed sentiments. Our results show that the proposed approach works better than standard one in terms of RMSE for most of the aspects across all the three applications. In particular, our method signi cantly outperformed standard MF for 43 aspects (out of 68) for restaurants, for 19 aspects (out of 44) for hotels, and for 33 aspects (out of 45) for beauty&spas. For all the remaining aspects, the di erences between our method and the standard MF approach are not statistically signi cant. As it was expected, our approach works better for those aspects that have several close neighbors frequently mentioned in the reviews. For example, in a restaurant application aspect \music" is close to \atmosphere" and \interior", and therefore, our approach outperformed the standard one in predicting sentiments of the \music" aspect by 7:8% in terms of RMSE.
In Step 3 of our method, we obtained the rating prediction model and compared its performance with the standard MF approach to rating predictions. The results of this comparison (in terms of RMSE) are presented in Table 2, from which we conclude that the two methods are very similar in terms of their performance. This means that our regression model (Eq.(1)) predicts ratings reasonable well (comparably to the MF approach).
Followed Other cases Not followed Other cases
Restaurants users managers 3.818 3.816 3.734 3.737 3.482 3.473 3.784 3.787
We next applied our models to the test data and determined the most important aspects for each user and item pair, as described in Step 4. Finally, we produced recommendations both for the users and the management, as described in Step 5.
We next compared the ratings of the users and the management who followed our positive recommendations (i.e., mentioned the recommended aspects in their reviews) against other cases (i.e., that did not mention them) across the restaurants, hotels and beauty&spas applications. Further, we also did similar comparisons for negative recommendations (such as, do not order the desert in restaurant X), i.e., we compared users or managers who did not follow our negative recommendations (i.e., mentioned the negatively recommended aspects in their reviews) with others.
The results of these comparisons are presented in Table 3, where numbers in cells represent average ratings for users and managers across the three applications. The rows in Table 3 represent di erent conditions of whether or not the users followed our recommendations. As Table 3 shows, our recommendations of items and conditions lead to higher evaluation ratings in those cases when users followed them vs. other cases, and all the di erences are statistically signi cant.
In this paper, we presented a new method of recommending not only items of interest to the user but also the conditions enhancing user experiences with those items. This method is based on the sentiment analysis of user reviews, predicts sentiments that the user might express about the aspects determined in an application, such as restaurants or hotels, and identi es the most valuable aspects of user's potential experience with the item. We tested it on three Yelp applications (restaurants, hotels and beauty&spas) and showed that our recommendations lead to higher evaluation ratings when users followed them vs. others.
This new approach to providing recommendations helps users to customize and enhance their experiences when consuming items, such as deciding what they should order in a particular restaurant.