ºÝºÝߣ

Editor's Notes

1. This work was done with Ossi Mokryn, Christophe Diot And Nina Taft This is a paper about cold start context based hotel recommender systemSo we r building a hotel recommender systemSo why cold start?Unlike restaurants or movies for example, in the hotel domain we don¡¯t have enough ratings for hotels which leads to a user cold start problemWe r going to use the user context to overcome the user cold start problem
2. So we saw that Users are generous with the star ratings while expressing their real opinion in writingThink about someone that go to a hotel with your family and kids, then in different occasion she will go to the same hotel for a business trip, Her needs for this two trips probably will be different, and the review the review that she will write might be different Its logic to think that user have different needs for different context, and we want to identify those needs or each context groupSo the main question here is Can we couple text analysis and user context to yield a better recommendation?
3. So how r we going to do it? With the use of common traitsWe borrow this term from psychology, trait is a basic characteristic of a person for example there is a shay person, a kind person etc. Common traits are defined for groups, for example in the Chinese tradition theyear of birth determines the persons¡¯ traitsEveryone that was born this year, the year of dragon, share certain?characteristics, according to the tradition they will be for example brave and flexible
4. We defined a common trait for each context group as the typical words that appear in the text written by users from this groupFor example when we look at reviews that was written by a single travelers we can see that they write a lot about wifi (a lot more that the other groups), so it means that wifi is very important feature for this group so wifi is a common trait for this context groupWe extracted the hotel features from the reviews text for example service, parking, pool etc.Then if the average appearance of a feature in all the context groups, minus the frequency of it for context group c is larger then the standard deviation, then this feature is a trait for this group
5. In practice for each feature that we extracted we assign a weight that reflects its importance for each context groupThis is the weight function of feature f for context group cThe deviation is the average frequency of a feature in all the groups text minus the frequency of this feature for this groupWhen the deviation is higher then the standard deviation then the weight will be 1 + the deviation divided by the standard deviation
6. To make our recommendation better we also need to know what are the user preferences for the hotel aspectsFirst lets define aspect, as I told u before we extracted features from the reviews¡¯ text, aspect is a group of features with the same orientation for example the hotel aspect Room includes the features bathroom, bed, room size, tv etc.To find all the features that belongs to an aspect We built a graph in which each node corresponds to a feature and the links are the pmi of the two features in the reviewsWe used unsupervised community detection algorithm called spin glass from the area of statistical mechanicsThe algorithm minimize the function that we see here, The basics of this function is to reword for internal links that exists and external links that doesn¡¯t exists, and penalize for internal links that doesn¡¯t exists and external link that existsThen each community that we find will correspond to an hotel aspect
7. We found 6 different clusters, we labeled them manually, Each feature can only end up in one clusterNumber of communities and there size is determined by the algorithmThe identification of these clusters is important as it determine the particular hotel aspects that we chose to ask users their preferences forSuppose for example that a user species that room is of most importance to her. The room aspect identifies a large number of features that are often used to discuss things inside a hotel room; thus reviews in which these features occur frequently are more important to this user than a userwho cares more about location
8. So now we r going to build a personalized hotel score
9. Now I m going to do an overview on the system, in the preprocessing phase we extracted the features from the review text, then we assigned a weight for the features for each context group, we clustered the features to hotels aspects, and we built an opinion lexicon with orientationThat was the preprocessing Now, at run time, the user gives us her input Now think what happens when u go online to look for a hotel, first of all U filter the hotels with prices and other meta data, then u read the reviews about the hotels, when u read the reviews, imagine that u r wearing a personalized glasses and u look for the things that r important for u. we try to use this glasses to read the hotels reviews and we give hight weight to things that are important for u. So in real time we build a personalized hotel score and recommend u the hotels that match your needs
10. When the user uses the system she select the purpose of the trip, nationality and her preferences on each one of the hotel aspect
11. So how r we going to createthis glasses? with weight, we r going to give each feature a weight that reflect how important is this feature for the user. After the user selects her input we have three different weights for the features, weight for the trip intent, the nationality and the hotels¡¯ aspects preferences, We multiply those weights and we get a weight for this user for each featureThe weight are multiplied to allow differentiation of users within our groups
12. Lets see an example, lets take the feature bathroom, the base weight for all features is 1, now Alice and Bob with the same nationality uses the system, and bathroom is a trait for their nationality, so the weight for bathroom for there nationality is 2, for Alice the room aspect is the most important one, and for Bob the location, the feature bathroom belongs to the aspect room, so bathroom are more important for Alice and the weight of bathroom for Alice will multiply to 4In this example we saw how the users input influence the feature weight
13. The hotel orientation score is the average score of all the hotels¡¯ reviews. The review score is summarization all the sentences score and the sentence score is summarizing all the features score, The feature¡¯s score is the semantic orientation score multiply by the feature¡¯s weight for user uThe review score should reflect the relative importance of the given review for the user. Reviews that are both important and positive are deemed most relevant thereby receiving the highest scores
14. The last parameter is the bias adjustment, it is based on the hotel ratings, it¡¯s a rating for hotel for a specific user, it contains the bias of the hotel, the bias of the purpose of the trip and the bias of the nationality of the user. For example if Germans rate best western hotel 0.5 more then all the other nationalities then the nationality bias of this hotel will be 0.5The calculation of the bias are pretty strait forward, lets look at the bias of nationality for hotel, for each review that was written on hotel h by nationality n, we reduce the average score of all the hotels and the hotel bias, we summarize all that and divide it by the number of reviewsThe hotel orientation score is between -40 to 80 and the bias term is between 0-5, Bias are included primarily to break ties, or to differentiate hotels when their scores are very close.
15. The hotels score is the hotel orientation score plus the bias adjustment
16. The effect of considering the intent of the trip is clear intuitively and from our results We want to farther verify the usefulness of nationality and bias parametersWe tested the affect of those parameters on the search resultsWe made queries to the system with the tested parameter and without it, then we counted how many different hotels were founded between the resultsWe can see that nationality context group affect 15-16% of the resultsAnd the bias score affect only 8-9% of the results and indeed the bias score break ties and differentiate between hotels that the orientation score is very close
17. We used human evaluationWe implemented our system and made it available to the public for useThe user make a query and then We present the user a list of six hotels. Some are results from our system, and some are the top rated hotels from TripadvisorIn order to avoid biasing the users, the six hotels are presented in random order and thus the user is unaware of the source of the recommendationsWe obtained 150 evaluations
18. For each hotel the users were asked 3 different questions, They needed to rate in a the scale of 1-5 vow well this recommendation match your needs, The second question was would u select this hotelAnd the last one was what was the main reason motivating your decision
19. For more then 60% of the hotels recommended by our system users said that they will select this hotels compare to more then 50% from the start rating recommendationThe other important thing is that for almost 16% of the hotels recommended by our system users said they wouldn't go compare to more then 26% of the hotels in the star ratings recommendationU can clearly see that the users was more satisfied from the results from our system
20. This is the ratings that the evaluators gave to the hotels that was recommended to themU can see that the hotels recommended by our method received more 4 and 5 ratings then those the other method, and similarly our recommended hotels received fewer 1 and 2 ratings than the other method.People were more satisfy from our system, and the approval rate is higher
21. This work was done with Ossi Mokryn, Christophe Diot And Nina Taft This is a paper about cold start context based hotel recommender systemSo we r building a hotel recommender systemSo why cold start?Unlike restaurants or movies for example, in the hotel domain we don¡¯t have enough ratings for hotels which leads to a user cold start problemWe r going to use the user context to overcome the user cold start problem