Hybridisation Techniques for Cold-Starting Context-Aware Recommender Systems

1. Hybridisation Techniques for Cold-Starting Context-Aware Recommender Systems Matthias Braunhofer ! Free University of Bozen - Bolzano Piazza Domenicani 3, 39100 Bolzano, Italy mbraunhofer@unibz.it RecSys - October 2014, Foster City, USA

2. RecSys - October 2014, Foster City, USA Outline 2 • Context-Aware Recommenders and the Cold-Start Problem • Related Work • Context-Aware Rating Prediction Models • Evaluation and Results • Conclusions and Open Issues

4. Context-Aware Recommender Systems • Context-Aware Recommender Systems (CARSs) aim to provide better recommendations by exploiting contextual information (e.g., weather) • Rating prediction function is: R: Users x Items x Context → Ratings RecSys - October 2014, Foster City, USA 3 3 ? 4 2 5 4 ? 3 4 1 ? 1 2 5 ? 3 3 ? 5 2 5 ? 3 5 ? 5 4 5 4 ? 3 5

5. Example: Google Now • “The right information at just the right time” RecSys - October 2014, Foster City, USA 4 Nearby photo spots Traffic & transit Nearby attractions

6. Example: South Tyrol Suggests (STS) • Our Android app that offers context-aware place of interest (POI) recommendations for the South Tyrol region of Italy Personality questionnaire Rating screen Suggestions screen RecSys - October 2014, Foster City, USA 5

7. Cold-Start Problem • CARSs suffer from the cold-start problem • New user problem: How do you recommend to a new user? • New item problem: How do you recommend a new item with no ratings? • New context problem: How do you recommend in a new context? RecSys - October 2014, Foster City, USA 6 1 ? 1 ? 2 5 ? ? 3 ? 3 ? 5 ? 2 5 ? ? 3 ? 5 ? 5 ? 4 5 4 ? ? 3 5 ? 1 ? 1 2 5 ? 3 3 ? 5 2 5 ? 3 5 ? 5 4 5 4 ? 3 5 ? ? ? ? ? ? 1 ? 1 2 5 ? 3 3 ? 5 2 5 ? 3 5 ? 5 4 5 4 ? 3 5

8. Our Solution: Hybrid CARS • Intuition: it is possible to adaptively combine multiple CARS algorithms in order to take advantage of their strengths and alleviate their drawbacks when predicting a user’s rating for an item given a particular cold-start situation • Example: RecSys - October 2014, Foster City, USA 7 (user, item, context) tuple CARS 1 CARS 2 Combination Final score Score Score Hybrid CARS

9. • Context-Aware Recommenders and the Cold-Start Problem RecSys - October 2014, Foster City, USA Outline 8 • Related Work • Context-Aware Rating Prediction Models • Evaluation and Results • Conclusions and Open Issues

10. RecSys - October 2014, Foster City, USA Related Work 9 Cold-starting CARSs … using additional data … better processing known data Active Learning (Elahi et al., 2013) Cross-domain recs. (Enrich et al., 2013) Implicit feedback (Shi et al., 2012) User / item attributes (Woerndl et al., 2009) Context similarities (Codina et al., 2013) Survey data (Baltrunas et al., 2012)

11. RecSys - October 2014, Foster City, USA Related Work 9 Cold-starting CARSs … using additional data … better processing known data Active Learning (Elahi et al., 2013) Cross-domain recs. (Enrich et al., 2013) Implicit feedback (Shi et al., 2012) User / item attributes (Woerndl et al., 2009) Context similarities (Codina et al., 2013) Survey data (Baltrunas et al., 2012) No unique optimal solution!

12. • Context-Aware Recommenders and the Cold-Start Problem RecSys - October 2014, Foster City, USA Outline 10 • Related Work • Context-Aware Rating Prediction Models • Evaluation and Results • Conclusions and Open Issues

13. MF Methods • Matrix Factorisation (MF) predicts unknown ratings by discovering some latent features that determine how a user rates an item; features associated with the user should match with the features associated with the item r q p 5 x 4 matrix 5 x 3 matrix 3 x 4 matrix RecSys - October 2014, Foster City, USA 11 r11 r12 r13 r14 r21 r22 r23 r24 r31 r32 r33 r34 r41 r42 r43 r44 r51 r52 r53 r54 a b c x y = z r42 = (a, b, c) · (x, y, z) = a * x + b * y + c * z ȓui = qiTpu

14. MF Methods • Matrix Factorisation (MF) predicts unknown ratings by discovering some latent features that determine how a user rates an item; features associated with the user should match with the features associated with the item r q p 5 x 4 matrix 5 x 3 matrix 3 x 4 matrix RecSys - October 2014, Foster City, USA 11 r11 r12 r13 r14 r21 r22 r23 r24 r31 r32 r33 r34 r41 r42 r43 r44 r51 r52 r53 r54 a b c x y = z r42 = (a, b, c) · (x, y, z) = a * x + b * y + c * z Rating prediction ȓui = qiTpu

15. MF Methods • Matrix Factorisation (MF) predicts unknown ratings by discovering some latent features that determine how a user rates an item; features associated with the user should match with the features associated with the item r q p 5 x 4 matrix 5 x 3 matrix 3 x 4 matrix Item preference factor RecSys - October 2014, Foster City, USA 11 r11 r12 r13 r14 r21 r22 r23 r24 r31 r32 r33 r34 r41 r42 r43 r44 r51 r52 r53 r54 a b c x y = z r42 = (a, b, c) · (x, y, z) = a * x + b * y + c * z ȓui = qiTpu vector

16. MF Methods • Matrix Factorisation (MF) predicts unknown ratings by discovering some latent features that determine how a user rates an item; features associated with the user should match with the features associated with the item r q p 5 x 4 matrix 5 x 3 matrix 3 x 4 matrix RecSys - October 2014, Foster City, USA 11 r11 r12 r13 r14 r21 r22 r23 r24 r31 r32 r33 r34 r41 r42 r43 r44 r51 r52 r53 r54 a b c x y = z r42 = (a, b, c) · (x, y, z) = a * x + b * y + c * z ȓui = qiTpu User preference factor vector

17. Basic CARS Algorithms CAMF-CC (Baltrunas et al., 2011) • CAMF-CC (Context-Aware Matrix Factorisation for item categories) is a variant of CAMF that extends standard MF by incorporating baseline parameters for contextual condition-item category pairs kΣ Σ RecSys - October 2014, Foster City, USA 12 ˆ ruic1,...,ck = qi T pu +μ + bi + bu + btcj j=1 t∈T (i ) qi latent factor vector of item i pu latent factor vector of user u μ overall average rating bi baseline for item i bu baseline for user u T(i) set of categories associated to item i btcj baseline for item category-contextual condition tcj

23. Basic CARS Algorithms SPF (Codina et al., 2013) • SPF (Semantic Pre-Filtering) is a contextual pre-filtering method that, given a target contextual situation, uses a standard MF model learnt from all the ratings tagged with contextual situations identical or similar to the target one • Conjecture: addresses cold-start problems caused by exact pre-filtering • Key step: similarity calculation RecSys - October 2014, Foster City, USA 13 1 -0.5 2 1 -2 0.5 -2 -1.5 -2 0.5 -1 -1 1 -0.96 -0.84 -0.96 1 0.96 -0.84 0.96 1 Condition-to-item co-occurrence matrix Cosine similarity between conditions

24. Basic CARS Algorithms Content-based CAMF-CC • It is a novel variant of CAMF-CC that incorporates additional sources of information about the items, e.g., category or genre information • Conjecture: alleviates the new item problem of CAMF-CC kΣ Σ RecSys - October 2014, Foster City, USA 14 Σ T ˆ ruic1,...,ck = (qi + xa ) a∈A(i ) pu +μ + bi + bu + btcj j=1 t∈T (i ) qi latent factor vector of item i A(i) set of item attributes xa latent factor vector of item attribute a pu latent factor vector of user u μ overall average rating bi baseline for item i bu baseline for user u T(i) set of categories associated to item i btcj baseline for item category-contextual condition tcj

25. Basic CARS Algorithms Content-based CAMF-CC • It is a novel variant of CAMF-CC that incorporates additional sources of information about the items, e.g., category or genre information • Conjecture: alleviates the new item problem of CAMF-CC kΣ Σ RecSys - October 2014, Foster City, USA 14 Σ T ˆ ruic1,...,ck = (qi + xa ) a∈A(i ) pu +μ + bi + bu + btcj j=1 t∈T (i ) qi latent factor vector of item i A(i) set of item attributes xa latent factor vector of item attribute a pu latent factor vector of user u μ overall average rating bi baseline for item i bu baseline for user u T(i) set of categories associated to item i btcj baseline for item category-contextual condition tcj

26. Basic CARS Algorithms Demographics-based CAMF-CC • It is a novel variant of CAMF-CC that profiles users through known user attributes (e.g., age group, gender, personality traits) • Conjecture: alleviates the new user problem of CAMF-CC kΣ Σ +μ + b+ b+ Σ bi u tcj RecSys - October 2014, Foster City, USA 15 ˆ ruic1,...,ck = qi T (pu + ya ) a∈A(u) j=1 t∈T (i ) qi latent factor vector of item i pu latent factor vector of user u A(u) set of user attributes ya latent factor vector of user attribute a μ overall average rating bi baseline for item i bu baseline for user u T(i) set of categories associated to item i btcj baseline for item category-contextual condition tcj

27. Basic CARS Algorithms Demographics-based CAMF-CC • It is a novel variant of CAMF-CC that profiles users through known user attributes (e.g., age group, gender, personality traits) • Conjecture: alleviates the new user problem of CAMF-CC kΣ Σ +μ + b+ b+ Σ bi u tcj RecSys - October 2014, Foster City, USA 15 ˆ ruic1,...,ck = qi T (pu + ya ) a∈A(u) j=1 t∈T (i ) qi latent factor vector of item i pu latent factor vector of user u A(u) set of user attributes ya latent factor vector of user attribute a μ overall average rating bi baseline for item i bu baseline for user u T(i) set of categories associated to item i btcj baseline for item category-contextual condition tcj

28. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Y Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N Y N New context? Y N New item? New user? Content-CAMF-CC & Demogr.-CAMF-CC Y N Y N Final score Final score Final score Score Score Score Score

29. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Y Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N Y N New context? Y N New item? New user? Content-CAMF-CC & Demogr.-CAMF-CC Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple

32. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N Y N New context? Y N New item? New user? Content-CAMF-CC & Demogr.-CAMF-CC Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple Y

33. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N Y N New context? Y N New item? New user? Content-CAMF-CC & Demogr.-CAMF-CC Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple Y

34. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N N New context? Y N New item? New user? Content-CAMF-CC & Demogr.-CAMF-CC Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple Y Y

35. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N N New context? Y N New item? New user? Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple Y Y Content-CAMF-CC & Demogr.-CAMF-CC

36. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Final score Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N N New context? Y N New item? New user? Y N Y N Final score Final score Final score Score Score Score new user, new item, known context) tuple Y Y Content-CAMF-CC & Demogr.-CAMF-CC Score

37. Hybrid CARS Algorithms Heuristic Switching • Heuristic Switching uses a stable heuristic to switch between the basic CARS algorithms depending on the encountered cold-start situation • Conjecture: better tackles all kinds of cold-start situations found in CARSs New context? RecSys - October 2014, Foster City, USA 16 (user, item, context) tuple Demogr.-CAMF-CC Content-CAMF-CC CAMF-CC New item? N N New context? Y N New item? New user? Y N Y N Final score Final score Final score Score Score Score Score new user, new item, known context) tuple Y Y Content-CAMF-CC & Demogr.-CAMF-CC Final score

38. Hybrid CARS Algorithms Adaptive Weighted (1/2) • Adaptive Weighted adaptively weights each basic CARS algorithm based on its predicted accuracy for the user, item and contextual situation in question • Extends the two-dimensional adaptive RS presented in (Bjørkøy, 2011) • Conjecture: optimises adaptation of differently performing CARS algorithms Score Error RecSys - October 2014, Foster City, USA 17 (user, item, context) tuple CAMF-CC Weighted score Final score Error model SPF Error model Content-CAMF-CC Error model Demogr.-CAMF-Error model Score Error Score Error Score Error

39. Hybrid CARS Algorithms Adaptive Weighted (2/2) • Builds for each basic CARS algorithm a new user-item-context error tensor whose entries are the known deviations (errors) of the CARS predictions from the true ratings • Uses a separate CARS error prediction model for each of these error tensors to predict the errors (accuracies) on a particular (user, item, context) tuple Σ )T (pu + ycu Σ )+μ + bi + bu RecSys - October 2014, Foster City, USA 18 ˆeuic1,...,ck = (qi + xci ci∈IC cu∈UC qi latent factor vector of item i pu latent factor vector of user u IC subset of item-related contextual conditions xci latent factor vector of contextual condition ci UC subset of user-related contextual conditions ycu latent factor vector of contextual condition cu μ overall average error bi baseline for item i bu baseline for user u

48. RecSys - October 2014, Foster City, USA Evaluation Used Datasets • 3 contextually-tagged rating datasets 20 STS (Braunhofer et al., 2013) LDOS-CoMoDa (Odić et al., 2013) Music (Baltrunas et al., 2011) Domain POIs Movies Music Rating scale 1-5 1-5 1-5 Ratings 2,534 2,296 4,012 Users 325 121 43 Items 249 1,232 139 Contextual factors 14 12 8 Contextual conditions 57 49 26 Contextual situations 931 1,969 26 User attributes 7 4 10 Item features 1 7 2

49. RecSys - October 2014, Foster City, USA Evaluation Evaluation Procedure • Randomly divide the entities (i.e., users, items or contexts) into ten cross-validation folds • For each fold k = 1, 2, …, 10 • Use all the ratings except those coming from entities in fold k as training set to build the prediction models • Calculate the Mean Absolute Error (MAE) and normalised Discounted Cumulative Gain (nDCG) on the test ratings for the entities in fold k • Advantage: allows to test the models on really cold entities • Disadvantage: can’t test for different degrees of coldness 21

50. Results Recommendation for New Users 1-nDCG@1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 RecSys - October 2014, Foster City, USA 22 MAE 2.4 2.2 2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 STS CoMoDa Music STS CoMoDa Music CAMF-CC SPF Content-based CAMF-CC Demographics-based CAMF-CC Average Weighted Heuristic Switching Adaptive Weighted

51. Results Recommendation for New Items 1-nDCG@1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 RecSys - October 2014, Foster City, USA 23 MAE 1.4 1.3 1.2 1.1 1.0 0.9 0.8 0.6 0.5 0.4 0.3 0.2 0.1 0.0 STS CoMoDa Music STS CoMoDa Music CAMF-CC SPF Content-based CAMF-CC Demographics-based CAMF-CC Average Weighted Heuristic Switching Adaptive Weighted

52. Results Recommendation under New Contexts 1-nDCG@1 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0 RecSys - October 2014, Foster City, USA 24 MAE 1.2 1.1 1.0 0.9 0.8 0.7 0.5 0.4 0.3 0.2 0.1 0.0 STS CoMoDa Music STS CoMoDa Music CAMF-CC SPF Content-based CAMF-CC Demographics-based CAMF-CC Average Weighted Heuristic Switching Adaptive Weighted

54. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) RecSys - October 2014, Foster City, USA Conclusions 26 likes SKIING FREERIDING ALPING SKIING likes MUSEUM MUSEUM likes

55. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes FREERIDING ALPING SKIING likes MUSEUM MUSEUM likes

56. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar FREERIDING ALPING SKIING likes MUSEUM MUSEUM likes

57. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes MUSEUM MUSEUM likes

58. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes Skiing MUSEUM MUSEUM likes

59. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes similar Skiing MUSEUM MUSEUM likes

60. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes likely likes similar Skiing MUSEUM MUSEUM likes

61. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes likely likes similar Skiing likes Wet MUSEUM MUSEUM weather Wet weather

62. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes likely likes similar Skiing MUSEUM MUSEUM likes similar Wet weather Wet weather

63. • Various cold-start situations require different CARS solutions • Hybridisation of several CARS techniques, each of which has its own strengths and weaknesses, allows to achieve best (cold-start) performance • First developed and tested hybrid CARS algorithms are able to outperform the state-of-the-art CARS algorithms (i.e., CAMF-CC and SPF) Skiing RecSys - October 2014, Foster City, USA Conclusions 26 SKIING 18-25 Male 18-25 Male likes similar likely likes FREERIDING ALPING SKIING likes likely likes similar Skiing MUSEUM MUSEUM likes likely likes similar Wet weather Wet weather

64. RecSys - October 2014, Foster City, USA Open Issues • Review additional knowledge sources which may be used to incorporate additional information about users, items and contextual situations • Check the availability of large-scale, contextually-tagged datasets with item and user attributes • Revise the used evaluation procedure and evaluation metrics • Identify the best-performing hybridisation method for cold-start situations • Design and execute a live user study 27

65. RecSys - October 2014, Foster City, USA Questions? Thank you.

Hybridisation Techniques for Cold-Starting Context-Aware Recommender Systems

More Related Content

What's hot (20)

Viewers also liked (12)

Similar to Hybridisation Techniques for Cold-Starting Context-Aware Recommender Systems (20)

Recently uploaded (20)

Hybridisation Techniques for Cold-Starting Context-Aware Recommender Systems