We introduce HRS, a recommender that exploits user reviews and identifies the features that are most likely appealing to users. HRS incorporates this knowledge into the recommendation process to generate a list of top-k recommendations, each of which is paired with an explanation that (i) showcases why a particular item was recommended and (ii) helps users decide which items, among the ones recommended, are best tailored towards their individual interests. Empirical studies conducted using the Amazon dataset demonstrate the correctness of the proposed methodology.
Suggestions generated by existing recommenders are not always
personalized and diverse enough to expose users to a wide range
of items within their realm of interest, not just popular ones [
In this paper, we present the initial research conducted to attempt
to solve the issues mentioned above. We created Honest
Recommendation System (HRS), a novel recommender system
that shows items in their real light. In developing HRS, we focus
our efforts in using collected information from users’ reviews to
generate personalized suggestions with their corresponding
explanations. By incorporating into the recommendation process
the feature preferences of an individual user (inferred from his
reviews), we can get to know the user better than by simply
considering his rating patterns. We strive for the development of a
recommender system a user trusts by providing information he is
interested in, no matter if it has a positive or negative connotation.
Our main contribution is the increased effectiveness and
satisfaction on a domain independent recommender. This is
accomplished by giving users information they care about, which
helps them make the best decision, in terms of selecting the most
adequate item among the recommended ones. Users' overall
satisfaction with a recommender is related to the perceived quality
of its recommendations and explanations [
In this section we discuss HRS overall recommendation process. (Parameters used by HRS were empirically determined. However, details are omitted due to page constrains.) Identify User’s Interest on Items. Consider a user U, who is a member of a popular site, such as Yelp or Amazon, which archives U’s reviews and rating history. Given that we aim to provide U with information he values and needs to choose among suggested items, we examine reviews written by U and identify the set of features (i.e., traits) that U cares the most about. Since we know features are mainly expressed as nouns, we perform semantic analysis on reviews1 and consider the frequency of occurrence of nouns U employs in his reviews. We rely on WordNet-based similarity measures (using WS4J java library, specifically WuPalmer algorithm) to find and cluster similar 1 Using Stanford CoreNLP http://nlp.stanford.edu/software/. terms, as different nouns can be used to express similar meaning. Each cluster would contain most frequent terms together with its closest words among the ones U uses in reviews. We do this to learn what items’ traits U most frequently mentions in his reviews and use that knowledge to predict which candidate items would be of U‘s interest. The top-2 most frequently-used term clusters are treated as U’s preferred features. Note that each cluster is labeled using the most representative2 cluster term.
U’s historical data (i.e. rated items) and employ the well-known
matrix factorization strategy [
reviews for each candidate item I and following the same process defined for identifying features of interest to U, we identify the top-2 features most-frequently mentioned in reviews pertaining to I. Thereafter, we generate a ranking score for I, which shows the degree to which U’s preferred feature are addressed in I’s reviews. This score is computed by averaging the degree of similarity (defined based on WordNet using the RitaWordnet library) between all the words in the term clusters generated for U and I. This score represents the level of U’s interest in I and is used for ranking U’s candidate items, such that the top-k ranked candidate items are selected as the items to be recommended to U. Generate Explanations. We generate the corresponding explanation for each recommended item I by showing why I is likely appealing to U. We do so by extracting the descriptions other users provided on U’s preferred features pertaining to I from archived reviews. We identify sentences in reviews pertaining to I that include terms exactly-matching (or highly-similar as determined using WordNet) to each of the labels generated for U’s clusters. In the explanation of each recommended item, HRS includes 3 sentences for each label. In doing so, HRS provides U with sufficient information about the recommendations without overwhelming U with too much information to read related to the recommended items. As previously stated, we do not emphasize the sentiment of the features, since our intent is not to make U like one option more than another, but save U’s time in identifying information important for him.
We conducted initial experiments using the Software3 domain in
the Amazon Review dataset [
0.704 0.748
We developed a new recommendation system that takes advantage of ratings and reviews, to create personalized suggestions. The first version of HRS, which showcases the main idea and purpose of the system, generated promising results, yet there are opportunities to explore in the future that will enhance its performance. Even though HRS did better than SVD, we plan to provide deeper examination and comparisons with other baseline and state-of-the-art recommendation strategies. We will also analyze the effect of considering only candidate items with rating scores above 3, which we anticipate will improve the overall performance of HRS. We will also extend the performance evaluation by conducting online user studies to further verify the fact that HRS helps users in making appropriate choices among provided suggestions. One of the limitations of the current design of our recommender is that only nouns extracted from reviews are treated as features which cause losing rich information from adjectives and verbs. To address this issue, we will conduct more in-depth analysis on part-of-speech and type dependencies on sentences in reviews. We are aware that HRS, in its current state, does not entirely solve the “cold start” problem. We will consider adopting a hybrid recommendation strategy that considers general item metadata, along with the popularity of items, in addition to examining alternative ways to extract information from reviews and further work towards eradicating the cold start problem.