Identification of relevant product attributes is critical to the success of any marketing campaign. This task can be conceptualized as an attribute recommendation problem based on the product's content or features, where the goal of a solution would be to automatically recommend relevant features to the marketer for highlighting in a campaign. In this research, we try to solve this problem by using preference mapping, a powerful technique for associating feature preferences with users. We perform preference mapping with sentiment scores associated with product attributes mined from user reviews on the Web. As a result of this process, we are able to visualize a set of compared products and the appropriateness of the attributes on the same two-dimensional space, enabling us to easily recommend important features to a marketer. Finally, we show that expert recommendations or ratings for product features do not necessarily correlate with preference maps based on user sentiments.
Information retrieval [Retrieval tasks and goals]: Recommender systems
Motivation. Product manufacturers are always faced with the dilemma of identifying which attribute(s) of their products they should highlight in their targeted marketing campaigns. For example, a digital camera has several defining aspects like power of zoom, size of display and image size in megapixels. A release of a new camera model by a manufacturer like Nikon will usually be followed by a marketing 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 Canodp/oyrraigfheet. R2e0q1u4estfpoerrmtihsseionisndfriovmidpuearmlispsaiopnesr@sacbmy.otrgh.e paper’s authors. CCBoRpyeciSnygs ’p1e4r,mOictttoebderf6o,r20p1r4iv,aStielicaonndVaallceayd,Cemaliifcorpniuar,pUoSsAes.. This volume is pCuobpylirsighhetd2a01n4dAcCopMy9ri7g8h-1te-4d5b03y-2it2s57e-d7i/t1o4r/s0.7 ...$15.00.
ChtBtpR://edcxS.dyosi.2o0rg1/140,.O11c4t5o/b26e0r064,228.0216409,4S7i8li.con Valley, CA, USA.
campaign to potential customers that will try to highlight
certain aspects or attributes of the model. This attribute
recommendation problem is critical to the success of the
campaign. Focusing on features that do not appeal to users
can result in a loss of large amount of ad spend and potential
losses in product revenue for a manufacturer. In this paper,
we address this challenge by proposing a principled
technique called preference mapping [
Related research. Alpert [
Approach. In this research, after specifying our product and attribute set, we acquire sentiment scores of user reviews that mention attributes for the products in our set. Following this, we associate user sentiments with the attributes mentioned in the reviews (instead of the product as a whole) and average them over reviewers who have written reviews concerning the attributes. We perform preference mapping on this processed dataset involving products, attributes and average sentiment scores and generate a biplot visualization that can be used for attribute recommendation. Finally, we compare our recommendations with expert opinion and show that there is no perfect correlation with what experts believe to be good features and what consumers like in a marketed product.
Organization. The rest of this paper is organized as follows. In Sec. 2, we describe our method of applying preference mapping to this situation. Next, we describe our data in Sec. 3 followed by experimental results and discussion in Sec. 4. Finally, we summarize our contribution and provide directions for future work in Sec. 5.
We analyze a set of products p and a set of product
attributes k. Customers who have bought these products
often go to the product or retailer website to provide feedback
about the product in the form of textual reviews. Most
of these reviews generally contain mentions of product
attributes. Further, positive or negative sentiments usually
accompany the above mentions of the attributes. In our
approach, we collect reviews where each sentence talks about
only one attribute. Appropriate anaphora resolution is
performed for review sentences when the attribute name is not
directly mentioned [
A preference mapping is then performed with the revieweraveraged scores of each of the various attributes for the different products. We now explain how this is performed. As the first step, sentiment scores for all the product attributes are scaled to the same range so that variances are comparable across attributes of each product. Consider X = (X1, X2, . . . , Xp)T as the matrix of the reviewer-averaged scores for the p products (say, di↵erent camera models) and the k attributes (like battery life, size of display and shutter delay). Thus each Xi is a vector with its elements as Xij, which is the reviewer-averaged sentiment score for attribute j of product i. The principal component (PC) transformation of the feature vector X is the linear transformation Y = T (X µ) where µ = E(X) and ⌃ = V ar(X) = 0. The transformation is such that V ar(Y ) is maximized and the following holds: 1 2 . . .
p where, V ar(Yj) = j, j = 1, 2, . . . , p, E(Yj) = 0 and Cov(Yj, Yi) = 0 when i 6= j.
Functions V ar(·), E(·) and Cov(·) refer to the variance, expectation, and covariance functions, respectively, and the j’s represent the eigenvalues of the matrix X. These eigenvalues have the corresponding eigenvectors as 1, 2, . . . , p (the number of eigenvectors is equal to the rank of the matrix X). Then the ith PC for each product is the weighted sum of the scores of the product across the attributes, the weights being obtained from the ith eigenvector. A biplot graph can be plotted for PC1 and PC2 with the weighted scores of each of the products and the eigenvector values for each attribute. The resultant graph provides an easily interpretable visualization that shows how products compare among each other based on customer reviews and the relative proximity of each attribute to their respective products with respect to associated positive user sentiment. Based on this multivariate visualization, marketing contents can be designed, highlighting favorable attributes for products. A schematic of the steps a marketer will undergo to utilize statistical analysis of social reviews to design product specific marketing campaigns is shown in Figure 1. Relevant steps have been explained in this section. Specific details about our dataset and experimental setup will be provided in the next section.
We test our approach on a dataset consisting of 1309 reviews related to four digital camera models (Canon G3, Canon Powershot SD500, Canon S100, and Nikon Coolpix 4300), having a total of 13 distinct attributes. These attributes (or features) that we analyzed are: flash, zoom, battery, auto (quality of automatic mode), photo quality, view (quality of view through the viewfinder), delay (delay between photos), look, start (startup speed), color, night (quality of night photos), lens and resolution. The reviews are pre-processed to identify mentions of camera attributes within their texts. The 13 attributes are mentioned a total of 583 times in the product reviews that we collected.
Expert ratings. It is an interesting exercise to compare our attribute recommendation system to expert opinion. To this end, we went through popular digital camera review sites dcresource1 and imaging-resource2 for extracting expert ratings on the thirteen attributes for our 1http://www.dcresource.com, Accessed 11 July ’14. 2http://www.imaging-resource.com, Accessed 11 July ’14.
Canon G3
Canon PowerShot SD500 Canon S100
Nikon coolpix 4300 four camera models. Since none of the popular camera review sites provide direct numeric ratings for attributes, we mapped expert opinion to a score of 1 or 2 depending upon the comments provided. For example, comments containing words like exceptional, excellent and good about an attribute were mapped to two, and weak and worst were assumed to be a one rating. The data that we collected has been made publicly available at http://goo.gl/v8BGj4.
We assign a sentiment score to each sentence in each review in our dataset with the Alchemy API3 and transfer the score to the attribute mentioned in the sentence. The higher the magnitude of the score, the stronger is the strength of the associated sentiment. Following this, the positive and negative sentiment scores of all the 52 (= 13 ⇥ 4) cameraattribute pairs were averaged together over all the reviewers who mentioned the pair in his/her reviews, the neutral sentiments contributing zero to the sum. The missing observations are assumed to be neutral sentiments and hence the scores in such cases are assumed to be zero. These average sentiments for each camera over all attributes are shown in a radial chart in Figure 2. As a specific example, the battery of the Canon S100 was mentioned in 13 reviews, with seven, one, and five review(s) showing positive, negative and neutral scores respectively. While the numbers of positive and negative mentions seem comparable, the average positive and negative sentiment scores were found to be 1.3461 and 0.3569 respectively, indicating that the strength of the negative sentiment was not as strong as the positive sentiment. In our experiments, the two values were averaged to obtain 0.8515.
We now have a matrix with four rows (corresponding to each camera model) and thirteen columns (corresponding to each model attribute). The cells of this matrix are the reviewer-averaged sentiment scores associated with each camera and attribute pair. A principal component analysis (PCA) is then performed on this matrix of camera-attribute pairs. The PC1 and PC2 for this example, cumulatively explain 85% of the variability in the data. We then produce the biplot of the weighted scores of the products and the eigenvectors of each of the attributes, as shown in Figure 3. 2 1 20 PC −1 −2
Canon G3 color lens
battery resolfuutniocntion asutatort phvoiteow Nikon coolpix 4300look zoom delay
flash −2
Canon PowerShot SD500 PC01 2 Canon S100
This graph provides a lot of information for design of marketing campaigns. First, in the graph, two attributes (in red) that are pointing towards the same direction, are attributes that tend to be highly positively correlated. A product that is in the same direction as an attribute, has a high value for this attribute. Thus, from the graph, we can conclude that attributes, which are closer and in the same direction as a product, are the ones that should be recommended for highlighting in marketing content for that particular model. For example, Canon G3 and Canon S100 received high sentiment scores on attributes like lens and color, while Nikon Coolpix 4300 and Canon PowerShot SD500 received high positive sentiments on low shutter delay and zoom quality. Thus, for example, lens and color should be recommended for designing marketing content in the campaign for Canon G3, rather than the zoom.
Second, this methodology also helps to contrast competing products simultaneously and provides competitive intelligence to the marketer. Thus, based on the given set of consumers’ reviews, one can deduce that Nikon Coolpix 4300 and Canon PowerShot SD500 are similar with respect to the attributes studied, as compared to Canon G3 and Canon S100. For example, if Nikon Coolpix 4300 and Canon Powershot SD500 are competing products, then it is meaninful to recommend only discriminatory features that add value to a particular product for its campaign. It is more sensible to recommend flash for Nikon Coolpix 4300 (more closer to the model than Canon 500) than the zoom, which is approximately equidistant from the both the products.
Analysis of expert opinion. From the data collected on expert comments (Sec. 3), we find that many of the discussed attributes are rated as 2, which implies that these attributes are “excellent” or “good” (Table 1). We assume that high expert score is analogous to high positive sentiment.
Table 2 shows the Kendall-Tau rank correlation coecients between the preference mapping technique and the plain average sentiment scores (which is the unweighted sum of the attributes as opposed to the weighted sum for each camera). For three cameras we have statistically significant (at 0.05 level) correlation between the methods and a moderExpert ratings were not available for all the attributes. So the sum of the values in a row may not add up to one ate correlation for Nikon Coolpix 4300. This shows that our method has high correlation with the intuitive understanding of the importance of the attributes and helps in further refinement. We could not observe any direct relation between the predictions based on the preference mapping and the attributes highly rated by experts.
The preference mapping technique, as described by us in
this research, recommends potentially “valuable” attributes
of products to marketers for highlighting in a marketing
campaign. Our method provides the marketer the ability
to design marketing content that can potentially increase
response rates. We have used sentiment scores for product
attributes, extracted from review texts to identify product
features to be highlighted in campaigns. By focusing on
attributes that are known to have received positive sentiments
of customers, the risk in the campaign is minimized.
Moreover, the comparison with the experts’ comments suggests
that sometimes, what customers value more about a
product may be di↵erent from attributes that experts consider
of high quality. So, designing marketing content taking into
account what a large section of consumers show positive
sentiments towards may help in engaging more e↵ectively with
a larger section of the consumers. The sentiment score in
our research is a continuous variable and PCA has been used
to identify appropriate attributes that have high scores. If
some or all the scores are categorical in nature, multi-factor
analysis [
Poor the potential customer and is likely to improve customer satisfaction.
As future work, we would like to cluster products using attribute sentiment scores as features and observe the correlation of the clustering output to the representation produced by our preference mapping technique. Also, the quality of the reviews can be improved by choosing relevant users by mapping them to specific customer segments. This can lead to better insights on the data and finer levels of control in the design of marketing content.
We thank Ritwik Sinha from Adobe Research Labs India for valuable inputs at various stages of this work. 6.