1. Towards Personalized Context-Aware Recommendation by Mining Context Logs through Topic Models
Kuifei Yu, Baoxian Zhang, Hengshu Zhu,Huanhuan Cao, and Jilei Tian
Springer-Verlag Berlin Heidelberg 2012

2. Goal
To mine common context-aware preferences (CCPs) from many users’ context logs through topic models and represent each user’s personal context-aware preferences as a distribution of the mined common context-aware preferences.
first we extract bags of Atomic Context-Aware Preference (ACP) Features for each user from their historical context logs.
Then, we propose to mine CCPs from users’ ACP-feature bags through topic models.
Finally, we make recommendations according to the given contexts and CCP distributions of users.

3. User Context Logs
A real-world data set of context logs collected from 443 mobile phone users spanning for several months, which contains more than 8.8 million context records, 665 different interactions (activities) in 12 content categories.

4. Related Work
Only leverage individual user’s historical context data preferences
Do not take into account the problem of insufficient personal training data
A personalized recommender system to recommend travel related information
Location-based personalized recommender system by Bayesian Networks
Points-of-interest (POI) for users in an automotive scenario by leveraging a Multi-Criteria Decision Making (MCDM)
Based on rating logs of mobile users and the objective is to predict accurate ratings for the unobserved items under different contexts
Collaborative Filtering (CF) based approaches
leverage a classification rule of decision tree to understand users’ personal preference
The approach can model user, location and activity as a 3-dimensional matrix, namely tensor
modeled the rich contextual information with item by N-dimensional tensor, and proposed a novel algorithm to make tensor factorization
It’s easier to collect context logs which contain users’ historical context data and activity records than rating data in user mobile devices.

5. Preliminary
Capture the historical context data and corresponding activity records as contextual feature-value pairs：
contextual features (e.g., Day name, Time range, and Location)
corresponding values (e.g., Saturday, AM8:00-9:00, and Home)
Transform raw location based context data such as GPS coordinates or cell Ids into social locations by some existing location mining approaches ：
“Home” and “Work Place”
Transform raw activity records by mapping the activity of using a particular application：
Transform two raw activity records “Play Angry Birds” and “Play Fruit Ninja” to same activity records “Play action games”
User u with C prefers activity a ：
a = activity
C = {p} = Atomic context
p = independent context
z = a variable of CCP

6. Example of activity records “Play action games”
ACP-Feature1
ACP-Feature2
ACP-Feature3 n
User1
Play action game
Saturday, evening, Home
Play video
Use facebook
Monday, morning, WorkPlace …
User2
Play action game Sunday, evening, By the way
Plan timetable
Play song list Monday, afternoon, WorkPlace
P(z|u)
Games Topic (z1)
Business Topic (z2)
Music Topic (z3)
Social Topic (z4)
Others…
User1
38/56
3/56
5/56
8/56 …
User2
36/125
24/125
12/125
15/125
P(a,p|z) z1 z2 z3 z4
Others…
Games- Saturday, evening, Home
0.78
0.12
0.22
0.09 …
Games - Sunday, evening, By the way
0.77
0.11
0.17
0.03
Social network service - Monday, morning, WorkPlace
0.25
0.23
0.21
0.53

7. LDA (Latent Dirichlet Allocation)
Suppose you have the following set of sentences:
I like to eat broccoli and bananas.
I ate a banana and spinach smoothie for breakfast.
Chinchillas and kittens are cute.
My sister adopted a kitten yesterday.
Look at this cute hamster munching on a piece of broccoli.
LDA is a way of automatically discovering topics that these sentences contain.
For example, given these sentences and asked for 2 topics, LDA might produce something like
Sentences 1 and 2: 100% Topic A
Sentences 3 and 4: 100% Topic B
Sentence 5: 60% Topic A, 40% Topic B
Topic A: 30% broccoli, 15% bananas, 10% breakfast, 10% munching, … (at which point, you could interpret topic A to be about food)
Topic B: 20% chinchillas, 20% kittens, 20% cute, 15% hamster, … (at which point, you could interpret topic B to be about cute animals)
每個topic中字的分佈為β 每個文章中可以有多個topics

8. Explanation of LDA
Corpus D ： di = {w1, w2, …, wn} which is called bag of words. It is also the input of LDA. VOC = {w1, w2, …, wm} is a set of D and each word in VOC is different. The output vectors of LDA ： α = each di’s probability distribution of z-th topics θd = <pt1, pt2, …, ptz> β = each topic’s probability distribution of word φz = <pw1, pw2, …, pwm> Core idea of LDA ： p(w|d) = p(w|t)*p(t|d) initial θd and φz at first the word’s probability of topic z in di = pwm* ptz according to this probability, we can update the word in topic z finally, we can get a convergent result
Pti=nti/n, n is the number of word in di , nti is the word appeared in ti
pwi=Nwi/N, N is the number of this topic, Nwi is {wi} in VOC appeared in this topic times

9. Mining Common Context-Aware Preferences through Topic Models
(Play action game, Sunday)
(Play action game, evening)
(Play action game, Home) …
(a, pl) = bag of words di
User1
User2
User3 …
Corpus D
<T1, C={Sunday, evening, Home …}, play action game>
<T2, C={Monday, morning, Work place …}, play facebook>
<T3, C={Saturday, morning, Home …}, browse web news> …
<Ti, C={p1, p2, p3, …,pl}, a>
Topic Z
Cpp1
Cpp2
Cpp3 …
(a,p),(a,p)…

10. PCR-LDA
The estimated values for two distributions {p(a, p|z)} and {p(z|u)}：
the number of times ACP-feature (a, p) has been assigned to CCP z + β
p(a, p|z) =
all the ACP-feature (a, p) in CCP z+ the number of ACP-features from u’s context log * β
the number of times a ACP-feature from user u’s context log that has been assigned to CCP z + α
p(z|u) = all the ACP-feature (a, p) in user u’s context log + the number of CCPs * α

11. Experiments Static Data Set
Utilize 10 activity categories ：
Web, Multimedia, Management, Games, System, Navigation, Business, Reference, Social Network Service (SNS), Utility
Contain 618 activities appear in total 408,299 activity-context records

12. Experiments benchmark methods
CPR (Context-aware Popularity based Recommendation)：
predict user preferred activities by the most frequent activities appear under C according to all users’ historical context logs
PCR-i (Personalized Context-aware Recommendation by only leveraging Individual user’s context logs)：
rank each activity a by probability = ACP-feature (a, p) / all ACP-features in the context log of u
Mean Average Precision at top K recommendation
Mean Average Recall at top K recommendation

13. Case Study

14. …… Training User1 User2 … User151 User152 user446 User151
(Watch Movie, Monday)
(Watch Movie, PM18:00-19:00)
(Watch Movie, Home)
(Play game, Monday)
(Play game, PM20:00-21:00)
(Play game, Home)
(Browse website, Monday)
(Browse website, PM21:00-22:00)
(Browse website, Home)
(Set timetable, Tuesday)
(Set timetable, AM6:00-7:00)
(Set timetable, Home)
(Use google map, Tuesday)
(Use google map, AM7:00-8:00)
(Use google map, On the way) …
TopicZ1 (media)
(Watch Movie, Monday) %
(Watch Movie, Saturday) %
(Watch Movie, Sunday) %
(Take picture, Monday) % …
(Watch Movie, Am:0:00-1:00) %
(Watch Movie, PM18:00-19:00) %
(Watch Movie, PM22:00-23:00) %
(Take picture, PM16:00-17:00) %
(Take picture, PM17:00-18:00) %
(Watch Movie, Home) %
(Watch Movie, On the way) %
(Watch Movie, Work place) %
(Take picture, On the way) %
TopicZ2 (internet)
(Watch Movie, Monday) %
(Watch Movie, Saturday) – 1.24%
(Watch Movie, Sunday) – 2.11%
(Browse website, Monday) – 6.12% …
(Watch Movie, Am:0:00-1:00) – 0.03%
(Watch Movie, PM18:00-19:00) – 0.14%
(Watch Movie, PM22:00-23:00) %
(Browse website, PM16:00-17:00) – 4.14%
(Browse website, PM21:00-22:00) – 5.56%
(Watch Movie, Home) %
(Watch Movie, On the way) – 0.01%
(Watch Movie, Work place) – 0.11%
(Browse website, Home) – 3.19% ……
Training

15. p(a,p|u) = p(a,p|z)*p(z|u)
User152
(Watch Movie, Monday)
(Watch Movie, PM21:00-22:00)
(Watch Movie, Home)
(Listen music, Monday)
(Listen music, PM22:00-23:00)
(Listen music,, Home)
(Browse website, , Monday)
(Browse website, PM22:00-23:00)
(Browse website, Home)
(Play online game, Tuesday)
(Play online game, AM9:00-10:00)
(Play online game, Home)
(Play action game, Tuesday)
(Play action game, AM10:00-11:00)
(Play action game, On the way) …
p(a,p|u) = p(a,p|z)*p(z|u)
The activity multimedia that user152 prefer is concerned by：
Probability of (a,p) appear in topic z1
Probability of topic z1’s distribution in user152
Testing