Hybrid recommendation approaches

Hybrid recommender systems Hybrid: combinations of various inputs and/or composition of different mechanism Collaborative: "Tell me what's popular among my peers" Content-based: "Show me more of the same what I've liked" Knowledge-based: "Tell me what fits based on my needs"

Hybrid recommender systems Idea of crossing two (or more) species/implementations hybrida [lat.]: denotes an object made by combining two different elements Overcome some of the shortcomings (i.e., cold-start problem) Reach desirable properties that is not present in parent individuals Different hybridization designs Parallel use of several systems Monolithic exploiting different features Pipelined invocation of different systems

Monolithic hybridization design Consist of only single recommendation component Integrates multiple recommendation algorithms and combines multiple feature/knowledge sources (as input data). Monolithic recommendation component can be designed based on two strategies: feature combination hybrid feature augmentation hybrid

Monolithic hybridization designs: Feature combination Feature combination strategy Social/collaborative features: movies liked by user Content features: movie1 is Comedy, movie2 is drama Hybrid features: user likes many movies that are comedies, … Example C Community and product knowledge

Monolithic hybridization designs: Feature combination (cont.) Predict similar users to Alice based on Pure collaborative approach User1 and User2 are similar to Alice Feature combination approach User2 behaves in very indeterminate manners User1 and Alice are similar. Hybrid input features

Monolithic hybridization designs: Feature augmentation In feature augmentation, we seek to have the contributing recommender produce some features that augment the knowledge source used by the primary(actual) recommender. e.g. content-boosted collaborative filtering [Prem Melville et al. 2002] Content-Based Contributing / Collaborative Actual The content-based algorithm is trained on the user profile (e.g. Alice) e.g. which items would user like/dislike The collaborative algorithm retrieves e.g. candidate restaurants from similar users (to Alice) e.g. User1 likes Macdonald, User2 likes pizza hut These candidates (restaurant ) will be rated by the content-based classifier and those ratings are used to augment the user profile (e.g. Alice ); thus filling it out with more ratings. e.g. content-base classifier predict Alice would like pizza hut and would dislike Macdonald The collaborative recommendation process is performed again with a new augmented profile (e.g. measure the similarity of users again). Similarity degree of Alice and User1 and Alice and User2 would change with the new feature augmented profile

Parallelized hybridization design Output of several existing implementations combined to generate final recommendation Three hybridization strategies: Mixed hybrids: the output of all recommenders presented side by side to a user Weighted hybrids: Weights can be learned dynamically Switching hybrids: Extreme case of dynamic weighting is switching

Parallelized hybridization design: Weighted A weighted hybridization strategy combines the recommendations of two or more recommendation systems by computing weighted sum of their scores. Given n different recommenders reck and their associated weights , the final recommendation score of item i for user u is: where

Parallelized hybridization design: Weighted Recommender 1 Item1 0.5 1 Item2 Item3 0.3 2 Item4 0.1 3 Item5 Recommender 2 Item1 0.8 2 Item2 0.9 1 Item3 0.4 3 Item4 Item5 Recommender weighted(0.5:0.5) Item1 0.65 1 Item2 0.45 2 Item3 0.35 3 Item4 0.05 4 Item5 0.00

Parallelized hybridization design: Weighted BUT, how to derive weights? Estimate, e.g. by empirical bootstrapping Dynamic adjustment of weights Empirical bootstrapping Historic data is needed Compute different weightings Decide which one does best Start with for instance uniform weight distribution For each user adopt weights to minimize error of prediction

Parallelized hybridization design: Dynamic weighting Let's assume Alice actually bought items 1 and 4: (r1 =1 ,r4 = 1) Identify weighting that minimizes Mean Absolute Error (MAE) Absolute errors and MAE Beta1 Beta2 rec1 rec2 error MAE 0.1 0.9 Item1 0.5 0.8 0.23 0.61 Item4 0.0 0.99 0.3 0.7 0.29 0.63 0.97 0.35 0.65 0.95 0.41 0.67 0.93 0.47 0.69 0.91 MAE improves as rec2 is weighted more strongly

Parallelized hybridization design: Dynamic weighting BUT: didn't rec1 actually rank Item 1 and item 4 higher? Be careful when weighting! Recommenders need to assign comparable scores over all users and items scores generated by different recommenders for the same items should not be high-variance Some score transformation could be necessary Stable weights require several user ratings Recommender 1 Item1 0.5 1 Item2 Item3 0.3 2 Item4 0.1 3 Item5 Recommender 2 Item1 0.8 2 Item2 0.9 1 Item3 0.4 3 Item4 Item5

Parallelized hybridization design: Switching Requires a mediator that decides which recommender should be used in a specific situation, depending on user profiles and the quality of recommendation results Special case of dynamic weights (all except one Beta is 0) Example: To overcome the cold-start problem for new user Use knowledge-based first, when users’ ratings collected, use collaborative filtering To overcome the cold-start problem for new item Use content-based filtering first, then use collaborative filtering when item’s ratings collected

Parallelized hybridization design: Mixed Combines the results of different recommender systems at the level of user interface Results of different techniques are presented together Recommendation result for user and item is the set of tuples for each of its constituting recommenders reck

Pipelined hybridization designs In pipelined hybridization several recommenders sequentially build on each other before the final one produces recommendation for the user Pipelined hybrids differs according to the type of output they generate for the next recommender Cascade : the preceding component deliver a recommendation list to subsequent component Meta-level: the preceding component preprocess input data to build a model for subsequent component

Pipelined hybridization designs: Cascade Refinement of recommendation lists Successor's recommendations are restricted by predecessor Where for all k > 1 Subsequent recommender may not introduce additional items Thus produces very precise results

Pipelined hybridization designs: Cascade Recommendation list is continually reduced There is a chance that no recommendation be made Cascade strategy should be combined by switching strategy to generate enough recommendation First recommender can only make a recommendation list e.g. knowledge-based recommender produce lists of unsorted items Subsequent (second) recommender assigns score or exclude an item Ordering and refinement (e.g. collaborative)

Pipelined hybridization designs: Cascade Recommender 1 Item1 0.5 1 Item2 Item3 0.3 2 Item4 0.1 3 Item5 Recommender 2 Item1 0.8 1 Item2 Item3 0.4 2 Item4 Item5 Ordering and refinement Recommendation list Recommender 3 Item1 0.70 2 Item2 0.00 Item3 0.85 1 Item4 Item5

Pipelined hybridization designs: Meta-level Successor exploits a model delta built by predecessor one recommender builds a (user) model that is exploited by the principal recommender to make a prediction Examples: Fab system [Balabanovic and Shham,1997]: Based on meta-level pipeline hybridization design Online news domain CB recommender builds user models based on weighted term vectors CF identifies similar peers based on these user models but makes recommendations based on users’ ratings

Limitations of hybridization strategies Only few works exist that compare different recommendation strategies, specially their hybrids Like for instance, [Robin Burke 2002] Most datasets do not have all the requirements to compare different recommendation paradigms i.e. ratings, requirements, item features, domain knowledge, critiques rarely available in a single dataset Thus few conclusions that are supported by empirical findings Monolithic: some preprocessing steps or minor modification in the recommender algorithms and data structure are required– to include more knowledge on the feature level Parallel: requires careful matching and weighting of scores from different predictors Pipelined: works well for two antithetic (mutually incompatible) approaches Netflix competition – "stacking" recommender systems Parallel hybrid design with adaptive weights/switching for recommenders Weighted design based on >100 predictors – recommendation functions Adaptive switching of weights based on user model, context and meta-features

Literature [Robin Burke 2002] Hybrid recommender systems: Survey and experiments, User Modeling and User- Adapted Interaction 12 (2002), no. 4, 331-370. [Prem Melville et al. 2002] Content-Boosted Collaborative Filtering for Improved Recommendations, Proceedings of the 18th National Conference on Artificial Intelligence (AAAI) (Edmonton,CAN), American Association for Artificial Intelligence, 2002, pp. 187-192. [Roberto Torres et al. 2004] Enhancing digital libraries with techlens, International Joint Conference on Digital Libraries (JCDL'04) (Tucson, AZ), 2004, pp. 228-236. [Chumki Basuet al. 1998] Recommendation as classification: using social and content-based information in recommendation, In Proceedings of the 15th National Conference on Artificial Intelligence (AAAI'98) (Madison, Wisconsin, USA States), American Association for Artificial Intelligence, 1998, pp. 714-720.