Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 39 views

    Research Challenges in Recommender Systems: 1 Modest Goals

    Uploaded by

    azertytyty000
    This document summarizes key challenges for recommender system research presented at a tutorial, as identified by practitioners from Netflix and Yahoo. The challenges discussed include: developing transparent recommendations that explain their reasoning to users; balancing exploration of new items with exploitation of known user preferences; enabling guided navigation of complex item spaces; accounting for how user preferences and item relevance can change over time; interpreting implicit signals from user actions; evaluating recommender systems based on real-world metrics like purchases and customer retention. Additional challenges mentioned but not covered in detail include balancing an "information diet", increasing serendipitous recommendations, privacy, limiting "balkanization", and combining content-based and collaborative filtering approaches.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Research Challenges in Recommender Systems: 1 Modest Goals

    Uploaded by

    azertytyty000
    39 views4 pages
    This document summarizes key challenges for recommender system research presented at a tutorial, as identified by practitioners from Netflix and Yahoo. The challenges discussed include: developing transparent recommendations that explain their reasoning to users; balancing exploration of new items with exploitation of known user preferences; enabling guided navigation of complex item spaces; accounting for how user preferences and item relevance can change over time; interpreting implicit signals from user actions; evaluating recommender systems based on real-world metrics like purchases and customer retention. Additional challenges mentioned but not covered in detail include balancing an "information diet", increasing serendipitous recommendations, privacy, limiting "balkanization", and combining content-based and collaborative filtering approaches.

    Original Description:

    Riedl Statement

    Original Title

    Riedl Statement

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document summarizes key challenges for recommender system research presented at a tutorial, as identified by practitioners from Netflix and Yahoo. The challenges discussed include: developing transparent recommendations that explain their reasoning to users; balancing exploration of new items with exploitation of known user preferences; enabling guided navigation of complex item spaces; accounting for how user preferences and item relevance can change over time; interpreting implicit signals from user actions; evaluating recommender systems based on real-world metrics like purchases and customer retention. Additional challenges mentioned but not covered in detail include balancing an "information diet", increasing serendipitous recommendations, privacy, limiting "balkanization", and combining content-based and collaborative filtering approaches.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    39 views4 pages

    Research Challenges in Recommender Systems: 1 Modest Goals

    Uploaded by

    azertytyty000
    This document summarizes key challenges for recommender system research presented at a tutorial, as identified by practitioners from Netflix and Yahoo. The challenges discussed include: developing transparent recommendations that explain their reasoning to users; balancing exploration of new items with exploitation of known user preferences; enabling guided navigation of complex item spaces; accounting for how user preferences and item relevance can change over time; interpreting implicit signals from user actions; evaluating recommender systems based on real-world metrics like purchases and customer retention. Additional challenges mentioned but not covered in detail include balancing an "information diet", increasing serendipitous recommendations, privacy, limiting "balkanization", and combining content-based and collaborative filtering approaches.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 4

    Research Challenges in Recommender Systems

    John Riedl

    Modest Goals

    This position paper lays out a set of challenges for the recommender research
    community. These challenges are specifically intended to be areas of research
    that are (a) not receiving much attention from active researchers at present;
    and (b) of importance to the practitioner community. The intent is to help researchers understand problems they might investigate that might lead to practically important results.
    Of course, researchers should not solely or even primarily focus on problems of merely practical importance. After all, these problems will receive plenty
    of attention from those who stand to gain market share, increase customer loyalty, and make more money if the sought-after improvements are achieved. Part
    of the reason for having a professional research community is precisely to focus
    on problems that are not of short-term commercial interest. Nevertheless, it is
    valuable to have a flow of information back and forth between practitioners and
    researchers. At some points in a research project, understanding the needs of
    the eventual consumers of the research may be helpful, and in some cases even
    inspiring.
    This paper arises from an invited tutorial that I was asked to lead at the
    Recommender Systems conference in New York in October 2009. ACM RecSys
    was first held in Minnesota in October 2007, and has now grown to 280 attendees. The attendees are very interested in the needs of the industrial practitioners, at companies like Google, Yahoo!, Netflix, and Amazon. The practitioners
    sometimes feel that the researchers are continuing to work on boring problems
    just because they are well-defined, and that more innovative and exploratory
    research would be more interesting and useful for them. Hence the tutorial, at
    which two leaders from industry (Jon Sanders of Netflix, and Todd Beaupre of
    Yahoo!) agreed to provide a critique of research directions in the field, and to
    suggest and motivate alternatives. My role was to help the presenters organize
    their thoughts in the months leading up to the conference, and to emcee the
    proceedings. I also chose, on my own initiative, to steal Jon and Todds good
    ideas, and share them with you. Any flaws in this writing are likely in my
    recollection or recounting, not in the original ideas.

    Key Challenges

    In this section we discuss the key challenges presented in the tutorial. For each
    challenge we introduce the challenge in a paragraph, and then present a hint of
    some directions in another paragraph or two.
    Transparency. The worst recommendation I ever made was in my senior
    year of college. I suggested to my friend Art that he would love the movie A
    Clockwork Orange. I still maintain that the recommendation was a good one
    though we are both certain it was a poor choice for his first date with Angela.
    If I had taken the time to explain to him what he would like about the movie
    the scary vision of a fascist futuristic society the disaster would have been
    avoided. The goal of transparency in recommenders is to provide these sorts of
    insight.
    Transparency can mean different things in different applications. Sometimes
    it means to help the user make a good decision, by providing him with selected
    subsets of information. Other times it means to help the user decide whether
    a particular recommendation fits a goal or mood. The challenges of developing
    effective transparency include (a) choosing whether to trace the workings of an
    algorithm, or to put together an independent argument that justifies the results
    of the algorithm; (b) what data to use, including other items, other users, or
    metadata; and (c) how to present the results to the user.
    Exploration versus Exploitation. One of the longest-standing challenges
    in recommenders is how to deal with the cold start problem for new items and
    new users, for which the recommender does not yet have enough information
    to form reliable recommendations. If information is not collected to form recommendations, the items will never be shown to any users, so information will
    never be collected, ...
    One approach is to take turns presenting the user with items he is likely
    to like (exploitation), and low information items to get his input so they can
    be effectively recommended to other users (exploration). The problem can be
    modelled by estimating the value provided to this user by a good recommendation, and comparing it to the value obtained for the community by getting this
    users input on a low information item. (Generally the value of additional input
    on a high information item is low, because of diminishing returns.) Of course,
    a key component of the estimate must be the risk of driving away the user if he
    is shown too many items he is not interested in.
    Guided Navigation. Rather than passively consuming a set of recommended
    items, users often want to be in control of their interaction with the items in a
    system. The recommendations are often best used as guides through a complex
    item space, letting the user choose in what directions to move at each step.
    One set of key implementation challenges is providing the user with navigation options along dimensions that she cares about, such as: mood, tags, and

    familiarity versus freshness. Another challenge is to incorporate conversation


    in the interface, so the system helps the user by suggesting key questions she
    might answer.
    Time Value. In some item spaces, such as books or movies, the relationships
    between users and items and users changes slowly. In other item spaces, such as
    daily news, items change in relevance rapidly, and different users have different
    taste in items. For instance, one user might value above all always knowing the
    most recent information, while another user might prefer the deeper insights of
    careful reporting that takes days or weeks to complete.
    Key question to explore include:
    How should user input decay in value over time, as the user changes preferences in response to earlier experiences?
    How can short-term (e.g., planning a trip to Hong Kong) and long-term
    (e.g., new job) preference changes be distinguished?
    How can user models recognize different preferences for timeliness?
    What should be done about evolving user tastes in long-term items, such
    as books, movies, or wines?
    User Action Interpretation. Explicit ratings are a valuable signal of user
    interest. However, in most systems at least an order of magnitude more information is available in the implicit signals of user interest, such as what items
    he clicks on, how long he reads them, which items are added to a wish list or
    shopping cart, etc.
    One deep challenge in these data are interpreting negative choices in addition
    to positive choices. What information is contained in a user not choosing to
    click on a news item? In systems with effective recommenders, users are usually
    presented with items they will like, so most ratings are high. The Best Paper
    at RecSys 2009 analyzed the ways this distortion in the available preference
    data causes problems for the recommender algorithms, and suggests statistical
    techniques to correct for the distortion.
    Evaluating Recommenders. How should recommender systems be evaluated? At the bottom line, companies that deploy recommenders care most
    about (a) purchase behavior; and (b) customer retention. Incorporating purchase behavior in recommender evaluation is relatively straight-forward (though
    the problems of the previous section must be dealt with).
    Incorporating customer retention is more difficult, because it takes weeks
    or months to measure. For this reason, companies seek proxy measures they
    believe reflect eventual retention, such as click-through rates, or time on site.
    These measures often show very different results from offline algorithm tests.
    In practice, interface changes often lead to much more dramatic changes in

    customer behavior than algorithmic changes. The most effective way to compare alternatives is through A/B testing, in which customers are separated into
    groups, which are subject to different treatments. Carrying out such studies
    can be difficult for researchers who do not have access to recommender systems
    with a steady flow of visitors.

    Missing Challenges

    In preparing for the tutorial we also discussed a number of other relevant issues
    that we did not have time to present in detail. The most interesting are:
    Information Diet: Users often want a mixture of items, not a set that includes
    only items of one type. For instance, a user might only want to see one
    story on Paris Hilton a day.
    Serendipity: Traditional algorithmic measures often evaluate how well a recommender does at finding items a user already knows about. Often the
    most valuable recommendations in practice are for items a user will like,
    but has no idea they even exist.
    Privacy: Can recommenders be developed that use privacy protected data?
    Can they prove to a user that the data is safe?
    Balkanization: Do recommenders lead to increasingly focused sub-cultures on
    shared values and interests? Is this dangerous? If so, can algorithms be
    developed that moderate the balkanization?
    Driving Traffic Strategically: Companies often have external goals they wish
    to meet with their recommenders. For instance, they might want to only
    recommend items that are in stock, or, given items of nearly equal value
    to the user, recommend the one with a higher profit margin.
    Combining Content and Collaborative Recommenders: Recommenders
    work in a world with rich metadata and rich data about users. How can
    these diverse data be combined most effectively?

    Closing Thoughts

    Yes, this paper is parochial in its narrow focus on recommender systems, which
    are but a small part of the big picture. Nevertheless, I hope these ideas have
    helped inspire some relevant thoughts about the area of social computing you
    find most interesting. Though most of these problems are under-researched, I
    can suggest citations to work that will provide a starting point if you ask.

    You might also like