CS 378 Midterm 3 Part 3

Two broad groups of recommender system

1) Content based systems:examine the properties of what the user does eg. recommened 'Cowboy" movies 2)Collaborative filtering: based on similarity measures betwen users and times eg. recommend items preferrend by similar users

example of student specific parameters

student profieincy- account for apriori abilties

model tracing

action is compared to rules in applicable model and feedback proided

BKT imporvement

add student specific spped of learning parmaeters

knowledge tracing model advantages, disadvantges (ACT-R)

all students end up learning the programming rules, (studies LISP langaue_) however, students make 40% error on quiz problems guess paramters have no effect, but slip parameter has an affect on performance model is too dependant on strenght of rules- each rule is equal weight which is not how it is in real life authors abandoned an ideal student model to compare studetns against. models individual diddferecens between students in learning and perforamnce - advantage predicts performance on things directly outside of the model eg. the test

how to populate utlity matrxi

ask usres to rate items make inferences from user behavior

knowledge tacing

attempt to monitor the student's changing knowledge state during practice

obtaining item features from tags

can use user genrated tag descriptions to label image data

metric used to evaluate similarity of numerical features in recommender system

cosine

discovering features of docs

feature values are not apparent eliminate stop word (common words), clacluate TF.IDF score take n highest scored words and claculate either Jaccard distance or cosine distance between sets lexical similarity not important

how doyou estimate the utility matrix in collaboratiev filetering

find n users most similar to a user and average their ratings for the item I make sure to normalize the matrix Can also use item similarity to estimate the etnry for user U. Find m items and take the average rating among the m items. tradeoof between useing similairt items or users to estimate the matrix: users- only have the do the process onece items- multiple times but more relaible - easier to find items of the same genre better to precompute and fix matrix to lower computation time

how to use clustered utility matrix to predict a similar item for a user

find the clusters to which the user and item belong if entry not balnk, use this value if entry blank, estimate value by considering similar clusters.

collaborative filtering recommendation

focuses on teh simiality of user ratings for two itmes uses a column from the utlitty matrix for item attributes user preferences are represetned by rows

Jaccard distance collab filtering

focuses on the sets of items rated. good as a simple measure two user's data is represented a set

Repreting item profiles

goal is to create an item profile consisting of feature-value pairs and a perfect user profile of preferences to match the user to teh item some data can be represented iwth booleans eg. actor in movie or actor not in movie numerical data is repersetned using data, where similar values suggest that the items are similar

Bayesian Knoweldge Tracing Models

in the standard version, there are only skill spcific parameters this paper adds student-specific parameters into BKT leading to improvement

knowledge tacing paramertrs

intial learning- proabbility the rule is learned state before first oppritunity to apply rule acquisition- probability the rule will changed from unlearned to learned guess- prob, student will guess correclty if rule is in the unlearned state slip- probability the sutdent will make a mistka eif the rule is in thte larne dstate primary goal is to predict knowledge KUMON

how to model student knowledge

model it as a latent variable and update it based on the correctness of the test questions you throw at the student

Content based reccomendation- how it wors

need to construct item profiles eg. set of actors in movei

User profiles

need to create vectors that match items to indicate user preferences estimate user preferneces by aggregating the profiles of the items that the user likes if the utility matrix is not boolean then weight the matrix normalize by substracting the average value for all users

calcuate cosine distance colab. filtering

numerator: for each vector get product of all items, sum them togehter deno: sqrt(imtem1^2 + item2^2) * the same for vector two normalize data by subtracting the average of the user's ratings from their individual rating

Bayesian Knowledge Tracing

observations are represented as binary variables (right or wrong), knowledge is also represented as binary variables (know/dont know)

Bayestian Knowledge Trcaing drawback

only has skill specific parameters

pewoabbility f leanred state

posteriori probability of alreadt leanred + probablity of switching to leanre state used to tutor Lisp operator defun uses Baye's theorme

Clustering users and Items

select a dsitance measure to perform a clustering of items- improves reccomendations can revise the utltity matrix based on clusters- cols represent clusters of items and contain average user ratigns rows correspond to clusters of users now

mastery learning

students can achieve expeertise in a domain if 1) domanin knowledge is turned into heirrachy of component skills 2) learning experiecesn are sturctures so that pre-rquiistie skills are first mastered

disadvantage of classifer recommenders

take a long time to train the classifier only really works on small problem set sizes

collobartive filtering quick def

the process of identifying simialirt users and recommedning what simialr users like is clalled collaborative filtering

drawbacks of collaborative filtering

there is a differnce in bhevaior of usere and items items are simple to clasiify users are very comlpex similar items are easy to discover, similar users are not easy to find

Classification algorithms

treat recommendation as an ML problem. data is trainng set, build a classifier for each user to predict ratnigs use decision tree, user preferences are the branchs of the tree can create an ensemble of overfitted decison trees to imporve accuracy

knowledge tracing model characteristics

two state learning model: each codeing rule is in learned or unlearned sate move from larned to learned through readin the text or applying the rule in pracitce no such thing as forgetting problem: learned state- student makes mistake unlearned- student can guies correctly probability of learned state is mainted and updates

problems with the knowledge tacing model

understimates laearning performacnce for above average students, they are given more remedial exercises then neccasery overesimates learningn performacne for below average stuents that make many errors, recieve less remedials, perform worse

Recommending items ot users based on content

use the cosine distance to compare item matrix and user prefernece matrix use LSH to place users into buckets if a movie is similar to one that a user likes, there will be a SMALL cosine distnace

Utility matrix in Recommendation system

users- have prefernec for itmes need to find preferences represent these user item paris in a matrix. condense ratings in the matrix to an average after clustering- then it terns into a cluster cluster utility matrix then utilize the cluster-cluster matrix to find the clusters to which items belong. if there is no cluster-cluster lookup in original matrix

collaboratiev filtering: what measures does it use

uses either the Jaccard or cosine distance. If two user's vectors are close then theya re similar

long tail

what makes recommendation systems necessary recommend unlimited number of online items to users - the obscure ones