EP2145265A1 - Évaluation sur la base d'une agrégation d'intérêts utilisateurs - Google Patents

Évaluation sur la base d'une agrégation d'intérêts utilisateurs

Info

Publication number
EP2145265A1
EP2145265A1 EP08744674A EP08744674A EP2145265A1 EP 2145265 A1 EP2145265 A1 EP 2145265A1 EP 08744674 A EP08744674 A EP 08744674A EP 08744674 A EP08744674 A EP 08744674A EP 2145265 A1 EP2145265 A1 EP 2145265A1
Authority
EP
European Patent Office
Prior art keywords
items
cluster
user
clusters
collection
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
EP08744674A
Other languages
German (de)
English (en)
Other versions
EP2145265A4 (fr
Inventor
James D. Chan
Kushal Chakrabarti
George M. Ionkov
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Amazon Technologies Inc
Original Assignee
Amazon Technologies Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US11/694,821 external-priority patent/US7966225B2/en
Priority claimed from US11/694,815 external-priority patent/US7743059B2/en
Priority claimed from US11/694,720 external-priority patent/US8095521B2/en
Priority claimed from US11/694,707 external-priority patent/US7689457B2/en
Application filed by Amazon Technologies Inc filed Critical Amazon Technologies Inc
Publication of EP2145265A1 publication Critical patent/EP2145265A1/fr
Publication of EP2145265A4 publication Critical patent/EP2145265A4/fr
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/951—Indexing; Web crawling techniques

Definitions

  • the present disclosure relates to data analysis and clustering methods for assessing item preferences of users, and for generating more reliable item recommendations for such users. Also disclosed are methods for organizing item recommendations for presentation to users.
  • Web sites and other types of interactive systems may implement recommendation services for recommending items stored or represented in a data repository. These services can operate, for example, by receiving an input list of items (optionally with associated item weights), and by outputting a ranked list of items that are collectively similar or related to the input set.
  • the items included in the input set are referred to herein as "source items " or "sources.”
  • recommendation services involves recommending products for purchase, rental, subscription, viewing or some other form of consumption.
  • e-commerce web sites commonly provide services for recommending products to users based on their respective purchase histories, rental histories, product viewing histories, and/or item ratings.
  • Recommendation services are also commonly used to recommend web sites, articles, users, music and video files, and other types of items.
  • the set of source items should ideally consist of items the target user likes. Otherwise, the recommendations may be of limited utility.
  • the task of reliably identifying such items without requiring explicit user input can be difficult.
  • a user's purchase history as maintained by an e-commerce web site is typically very useful for generating recommendations, this purchase history may include items purchased by the user for others as gifts. Unless the user actually designated these items as gifts at the time of purchase, these items may be difficult to identify and filter out.
  • the purchase history may include purchases made by multiple family members that share a home computer and account. The task of identifying appropriate source items is similarly difficult when the recommendations are based, e.g., on the item viewing histories, item rental histories, or item download histories of users.
  • some web sites allow users to view and "edit 7* their respective purchase histories, item viewing histories, and/or other item collections on an item-by-item basis, such as by rating, deleting, and/or tagging particular items. These edits are then taken into consideration in generating recommendations for the user. As one example, a user may delete from his or her purchase history all gift purchases, or may otherwise mark these items to indicate that they should not be used to generate recommendations. As another example, a user may tag the purchases that correspond to a particular family member or interest, and then request tag-specific recommendations that are based specifically on those purchases. In addition, some systems enable users to explicitly rate individual items that are recommended to them as "not interested.”
  • Various computer-implemented processes and features are disclosed for using item clustering techniques to assess user interests and to improve the utility of item recommendations provided to users. These processes may be implemented individually or in combination within a given computer system, such as, but not limited to, a web-based electronic catalog system.
  • One process involves applying a clustering algorithm to a user's purchase history or other collection of items. Information about the resulting clusters is then used — optionally in combination with other criteria — to select items to use as recommendation sources. For instance, items falling in a relatively small cluster may be excluded as sources on the basis that they likely represent gift purchases, or otherwise represent items falling outside the areas of interest of the user.
  • Another process involves displaying the clusters of items to the user via a collection management interface that enables the user to attach cluster-level metadata, such ratings or tags, to entire clusters of items.
  • the resulting metadata may be used to improve the recommendations generated by a recommendation engine. For example, a user may explicitly or implicitly indicate that an entire cluster of items should be excluded from consideration in generating recommendations.
  • Another process involves forming clusters of items in which a user has indicated a lack of interest. These clusters are used to filter the output of a recommendation engine so as to remove items that may represent poor recommendations. Another process involves applying a clustering algorithm to the output of a recommendation engine to arrange the recommended items into cluster-based categories for presentation to the user.
  • the calculated distances may be used as a basis for item clustering.
  • One aspect of the disclosure is a computer-implemented method, comprising: identifying a collection of items associated with a target user; applying a clustering algorithm to the collection of items to subdivide the collection into multiple clusters of items, wherein the clustering algorithm generates said clusters based, at least in part, on calculated distances between the items; selecting a subset of the items in the collection to use as recommendation source items based, at least in part, on one or more attributes of the clusters of items; using the selected recommendation source items as inputs to a recommendation engine to generate a set of recommended items for the target user; and outputting a representation of at least some of the recommended items to the target user.
  • Another aspect of the disclosure is a computer system, comprising: a computer data repository that stores a collection of items associated with a user: a clustering component that applies a clustering algorithm to the collection of items to divide the collection into multiple clusters of items; a source selection component that selects items from said collection to use as recommendation sources, wherein the source selection component selects said items based, as least in part, on information regarding said clusters of items; and a recommendation engine that uses the source items selected by the source selection component to generate personalized item recommendations for the user.
  • Another aspect of the disclosure is computer-implemented method, comprising: identifying a collection of items associated with a user, said collection represented in computer storage; applying a clustering algorithm to the collection of items to subdivide the collection into a plurality of clusters of items, wherein the clustering algorithm generates said clusters based, at least in part, on calculated distances between the items; for at least a one cluster of said plurality of clusters, assessing whether the cluster represents an interest of the user, said cluster comprising multiple items; and storing a result of said assessment in computer storage.
  • Another aspect of the disclosure is a computer-implemented method of facilitating user management of a collection of items.
  • the method comprises: maintaining an item collection of a user in computer storage; applying a clustering algorithm to the item collection to subdivide the collection into multiple clusters of items, said clusters generated based, at least in part, on calculated distances between the items; displaying the multiple clusters to the user via an interface that enables the user to associate metadata with specific clusters of items, and receiving cluster-level metadata from the user, said cluster-level metadata corresponding to a cluster selected by the user, said cluster comprising a plurality of items; using the cluster-level metadata, in combination with data regarding the collection of items, to generate personalized item recommendations for the user; and outputting a representation of the personalized item recommendations for presentation to the user.
  • Another aspect of the disclosure is a computer system comprising: a computer data repository that stores a collection of items associated with a user; a clustering component that applies a clustering algorithm to the collection of items to divide the collection into multiple clusters of items; a collection management interface that includes functionality for the user to view the clusters of items, and to associate metadata with a selected cluster of items, said cluster including multiple items; and a recommendation system that uses said metadata, in combination with data regarding the collection of items, to generate personalized item recommendations for the user.
  • Another aspect of the disclosure is a computer-implemented method, comprising: maintaining a collection of items in association with a user, said collection comprising items rated by the user; applying a clustering algorithm to said collection of items to subdivide the collection into multiple clusters, wherein the clusters are generated based, at least in part, on calculated distances between the items; obtaining a set of recommended items for the user from a recommendation engine; filtering the set of recommended items based, at least in part, on distances between the recommended items and one or more of said clusters of items, wherein filtering the set of recommended items comprises filtering out at least one item from the set of recommended items to generate a filtered set of recommended items; and outputting the filtered set of recommended items for presentation to the user
  • Another aspect of the disclosure is a computer system comprising: a computer data repository that stores a collection of items associated with a user; a clustering component that applies a clustering algorithm to the collection of items to divide the collection into multiple clusters of items; a recommendation engine that generates a set of personalized item recommendations for the user; and a filtering component that filters out selected items from the set of personalized item recommendations based, at least in part, on distances between said items and particular item clusters generated by the clustering component.
  • Another aspect of the disclosure is a computer-implemented method, comprising: generating a set of recommended items for a user; dividing at least a portion of the set of recommended items into multiple clusters of recommended items, wherein the clusters are generated based, at least in part, on calculated item distances; assigning a category name to each cluster of recommended items; and exposing each cluster of recommended items to the user in association with its respective assigned category name to facilitate viewing of the set of recommended items.
  • Another aspect of the disclosure is a computer system comprising: a recommendation system that generates personalized item recommendations for a user: a clustering system that uses distance calculations to divide at least a portion of the personalized item recommendations into multiple clusters of recommended items, and assigns category names to said clusters of recommended items; and a display system that exposes the clusters of recommended items to the user in association with said category names to facilitate viewing of the recommended items.
  • Figure 1 illustrates a clustered representation of a user's purchase history or other item collection.
  • Figure 2 illustrates a process for selecting recommendation source items from the user's item collection.
  • Figure 3 illustrates one example of a collection management user interface.
  • Figure 4 illustrates a cluster-based process for filtering the output of a recommendation engine.
  • Figures 5 and 6 illustrate processes for organizing a set of recommended items into cluster-based categories for presentation to users.
  • Figure 7 illustrates a portion of a web page showing an example browse cloud interface that may be used to organize a set of recommended items into cluster-based categories.
  • Figure 8 illustrates an example hierarchical browse structure.
  • Figure 9 illustrates one example how the various cluster-related features may be implemented in the context of a web-based electronic catalog system. DETAILED DESCRIPTION OF SPECIFIC EMBODIMENTS
  • the processes are described primarily in the context of a system that recommends catalog items to users of an e-commerce web site. As will be apparent, however, the disclosed processes can also be used to recommend other types of items, such as but not limited to web sites, news articles, blogs, travel destinations, service providers, other users, and events. In addition, the disclosed processes need not be implemented as part of, or in conjunction with, a web site.
  • Section I describes a process in which a user's collection of items, such as items in the user's purchase history, is subdivided into multiple clusters. These clusters are then analyzed to assess the user's interests, and are used (optionally in combination with other criteria) to select the "besf ' items to use as recommendation sources.
  • Section II describes a collection management interface, and associated methods, for enabling users to manage their item collections at a cluster level, hi the illustrated embodiment, the interface enables the user to rate and/or tag entire clusters of items, and to request recommendations that are specific to a particular cluster.
  • Section III describes a process in which clusters are formed of items in which the user has indicated a lack of interest. These clusters are then used to filter the output of a recommendation engine.
  • Section IV describes a process for using clustering to organize the output of a recommendation engine by category for display to a user.
  • Section V displays how item distances (used for clustering) can be calculated based on locations of the items within a hierarchical browse structure, such as a directed acyclic graph.
  • Sections VI-VIII describe specific examples of clustering methods that may be used to implement the processes described in Sections I-IV.
  • Section IX describes one example of a system architecture for implementing the various processes in the context of a web site that hosts an electronic catalog.
  • the term "item” may refer to an item itself (e.g., a product that can be purchased or a web site that can be accessed), or to an identifier or other representation of that item in a computer (e.g., a product or web site identifier or description stored in a database). In some cases, the term may be used collectively to refer to both.
  • Figure 1 illustrates a target user's collection of items as organized into six clusters, C1-C6, and will be used to explain how clustering can be used to assess user interests and to select items to use as recommendation sources.
  • the processes described in this section may be implemented by a computer system that implements, or makes calls to, a recommendation engine or service.
  • the primary goal of these processes is to automatically identify the items that likely do not fall within one of the target user's areas of interest (e.g., items purchased as gifts), and to refrain from using these items as recommendation sources. By excluding these items, the quality or utility of the recommendations can be significantly improved.
  • the collection consists of items purchased by the user, and each point represents a purchased item.
  • the collection may also include items rated, but not necessarily purchased, by the user.
  • the collection may additionally or alternatively be based on other types of item "selection 1" actions ⁇ e.g., rentals, views, downloads, shopping cart adds, wish list adds. subscription purchases, etc.).
  • selection 1 actions ⁇ e.g., rentals, views, downloads, shopping cart adds, wish list adds. subscription purchases, etc.
  • the distance between each pair of items (points) in Figure 1 represents the calculated degree to which the items are similar, with relatively small distances representing relatively high degrees of similarity.
  • any appropriate distance metric(s) can be used for item clustering.
  • each item may be represented as a vector of the browse nodes or categories in which it falls.
  • a browse node represents a particular category or other grouping of items, and is typically associated with a particular category name.
  • the respective vectors of two items can then be compared to compute the distance between the items.
  • the amount of weight given to each browse node is preferably inversely proportional to the number of items falling under that browse node, such that relatively narrow or low-level browse nodes are given more weight.
  • a specific example of how a hierarchical browse structure can be used to calculate item distances is provided below in Section V.
  • the distances between the items may additionally or alternatively be calculated based on other criteria.
  • the distance between two items, A and B may be calculated based on any one or more of the following: (a) the relative frequency with which A and B co-occur within purchase histories of users, (b) the relative frequency with which A and B co-occur within item viewing histories of users, (c) the relative frequency with which users tag A and B with the same textual tag, (d) the relative frequency with which A and B co-occur within results of keyword searches, (e) the degree to which A and B contain or are characterized by common keywords.
  • the foregoing are merely examples; numerous other criteria may be used to calculate the item distances.
  • the six circles shown in Figure 1 represent the cluster centers, and the lines identify the cluster to which each item is assigned.
  • the clusters may be generated using any appropriate type of clustering algorithm that uses item distances to cluster items. Examples include K-means, IsoData, nearest neighbor, and hierarchical type clustering algorithms. The clusters formed by such algorithms are mutually exclusive, meaning that each item is assigned to a single cluster. Specific examples of suitable clustering algorithms are described below in Sections Vl-VlIl.
  • the user's purchase history is characterized by three relatively large clusters, Cl, C2 and C5, and three relatively small clusters. C3, C4 and C6. Because C3, C4 and C6 are relatively small and are somewhat distant from the larger clusters, an inference may be drawn that some or all of the items in these clusters represent gift purchases, or represent purchases that otherwise fall outside the user's areas of interest. (Such an inference may not be suitable, however, if the small cluster is dominated by recent activity, suggesting a possible new interest for the user.) Thus, these "outlier " items, or items of outlier clusters, may be ignored (not used as sources), or possibly given less weight, when generating recommendations for the user. By excluding these items, the quality or utility of the recommendations can be significantly improved.
  • the outlier items and/or clusters may be identified and excluded programmatic ally based on one or more selected criteria. For instance, an item may be treated as an outlier and excluded if some or all of the following conditions are met: (a) the item falls in a cluster having less than some threshold number of items, such as 5; (b) this cluster is significantly smaller than the largest cluster (e.g., less than 10% of its size); (c) the item is some threshold distance from the nearest non-outlier cluster, (d) the item falls in a cluster that consists primarily of items rated below a selected threshold by the target user; (e) the item falls within a cluster having a cluster score that falls below some threshold, where the score generally represents the likelihood that the cluster represents an interest of the user. Other cluster attributes, such as scatter and homogeneity, may also be taken into consideration. In one embodiment, the assessment of outlier interests is performed only at the cluster level, and not the item level.
  • the outlier items may be subjected to greater scrutiny to determine whether they should be excluded. This may be accomplished by, for example, analyzing additional information regarding the particular user and/or affiliated users. The following are examples.
  • J0053 Comparison with known gift purchases. If the user has explicitly designated certain purchases of gifts, a determination can be made whether these known gift purchases correspond to particular clusters. This can be accomplished in various ways. For example, the known gift purchases can be excluded during the clustering process, and the distances between the resulting clusters and the known gift items can be computed and compared. If the known gift purchases tend to correspond to one or more particular clusters, these clusters more likely (but not necessarily) represent gift purchases that should be ignored. As another example, the known gift purchases may be included with the user's other purchases during cluster formation. If any resulting cluster primarily contains known gift purchases, all items in that cluster may be excluded or given less weight, or the cluster may be analyzed based on other criteria to assess whether it represents an interest of the target user
  • clustering can be applied to the purchase history of the known gift recipient. If one of the target user's clusters is close to (e.g., within a threshold distance of) one of the recipient ' s clusters, the target user ' s cluster may be treated as likely representing undesignated gift purchases for this same recipient. The items in that cluster may thus be excluded ⁇ not used as sources).
  • each cluster or item (or each outlier cluster or item) can be scored based on multiple criteria, including some or all of the criteria listed above. The resulting scores can then be used to select particular items to include or exclude as recommendation sources.
  • This process (including application of the clustering algorithm) may be performed in real time when a user requests recommendations, or may be performed off-line in response to certain events, such as purchase events.
  • the criteria used to analyze the clusters may vary significantly based on the type of collection being analyzed. For example, the analysis performed on a collection that includes or consists of items rated but not purchased by the user will typically be different from the analysis performed on a collection that consists of purchased items. The criteria used may also depend largely on the types of items involved (e.g., physical products, web sites, articles, etc.).
  • Figure 2 illustrates a generalized sequence of steps that may be performed by a computer system, such as one or more physical servers of a web site system, to implement the foregoing process.
  • the relevant item collection for the target user is initially retrieved. This collection may, for example, include or consist of items the target user has purchased, rented, viewed, downloaded, rated, added to a shopping cart, or added to a wish list.
  • the items may be products represented in an electronic catalog, or may be some other type of item (e.g., web sites) that is amenable to clustering.
  • the computer system may then determine whether the collection is sufficiently large (e.g., at least 40 items) to apply a clustering analysis. If it is not, the entire collection may be used as the recommendation sources, or some other criteria (e.g.. purchase dates, item sales ranks, etc.) maybe used to select the sources from the collection. Assuming the collection is sufficiently large, an appropriate clustering algorithm is used to subdivide the collection into multiple clusters (block 36). As part of this step, distances between the items may be computed using any appropriate metric.
  • the collection is sufficiently large (e.g., at least 40 items) to apply a clustering analysis. If it is not, the entire collection may be used as the recommendation sources, or some other criteria (e.g.. purchase dates, item sales ranks, etc.) maybe used to select the sources from the collection. Assuming the collection is sufficiently large, an appropriate clustering algorithm is used to subdivide the collection into multiple clusters (block 36). As part of this step, distances between the items may be computed using any appropriate metric.
  • the resulting clusters are analyzed optionally in combination with other user data such as purchase dates, item ratings, and/or known gift purchases — to select particular items to include or exclude as sources. This may be accomplished in numerous ways. As one example, a score may be generated for each cluster, and these scores may be used to select the clusters from which the source items are obtained. The cluster scores may be based on a variety of factors, such as some or all of the following: (1) the number of items in the cluster.
  • the sources may, for example, be selected from the highest scored clusters only, with additionally item-specific criteria optionally used to select specific items from these clusters. Alternatively, the sources may be selected such that the probability that a given item will be selected is directly proportional to the score of the cluster in which that item falls.
  • a score may be assigned to each item in the collection (or only to those identified as outliers), and these scores may be used on an item-by- item basis to select items to include/exclude.
  • the item scores may, for example, be based on both the cluster-based and non-cluster-based criteria described above. If all of the items are scored, some pre-selected number (e.g., 64) of the most highly scored items may be selected for use as the sources, with this number being selected so as to regulate the load placed on the recommendation engine/service.
  • the selected source items are used to generate recommendations for the target user.
  • this step involves passing a list of source items, optionally together with item weights, to a recommendation engine.
  • the item weights, if provided, may be based on the purchase dates of the items (if applicable), the user ' s ratings of the items (if rated), and/or some other criteria.
  • the recommendation engine uses this list to generate and return a ranked list of recommended items.
  • the recommendations may. for example, be generated as described in U.S. Patent No. 6,912,505, which is hereby incorporated by reference in its entirety, although other types of recommendation processes can be used.
  • the ranked list of recommended items, or an appropriately filtered version of this list is then presented to the target user (block 42).
  • the entire process depicted in Figure 2 may optionally be performed in real time in response to a user request for recommendations.
  • the process may be executed when a user clicks on a link which reads "view your recommendations," or when the user accesses a particular page that is configured to display personalized recommendations.
  • the user may be given the option to manage the collection at the cluster level.
  • the system may include a user interface that displays each cluster and allows the user to exclude an entire cluster (i.e., all of its items) from consideration in generating recommendations.
  • the user interface may additionally or alternatively enable the user to tag an entire cluster (and thus all of its items) with an appropriate label, such as a label that represents the corresponding interest or that identifies a particular family member.
  • the system also preferably enables the user to request "tag-specific'" recommendations as described in U.S. Appl. No. 1 1/281,886, filed November 17, 2005, the disclosure of which is hereby incorporated by reference in its entirety.
  • FIG. 3 illustrates a web page that represents one example of such a "collection management" user interface.
  • the left pane 50 of the web page displays clusters of items purchased by the user (three clusters shown, each referred to as a '"category").
  • the clusters may be generated by applying an appropriate clustering algorithm to the user's purchase history or other collection, as described above.
  • Each cluster/category has a name, which may initially be set by the system to the name of the browse node most closely associated with the cluster, and which may be modified by the user. Where a given cluster is too large to be concurrently displayed on the page, a scroll bar 52 is provided for scrolling through the cluster horizontally. Additional clusters/categories can be viewed by selecting the arrow 54 at the bottom of the pane 50.
  • the interface enables the user to drag-and-drop items to move them between clusters/categories. If a user reassigns an item to a new cluster/category, this reassignment may be maintained if/when the user's collection is later re-clustered. This may be accomplished by utilizing one of a variety of constraint-based clustering algorithms to maintain the user ' s manual assignments while augmenting them with the automatically computed clustering. The ability to move items between categories may alternatively be omitted.
  • the web page shown in Figure 3 also includes controls for rating each category/cluster on a scale of 1-5 or with a "not interested" rating.
  • the category-level ratings applied via these controls preferably do not override any pre-existing item-level ratings assigned by the user. If the user marks a category as "not interested,” none of the items in that category will be used as sources. If the user rates a category on a scale of 1-5, the category-level rating may be used as a default rating for any unrated items in that category/cluster.
  • the resulting item ratings may be used both to select items to use as recommendation sources (with highly rated items generally given preference), and to determine how much weight to give to particular source items when generating the recommendations. Instead of providing a "not interested" cluster rating option, a checkbox may be provided for the user to explicitly indicate that a particular cluster should not be used to generate recommendations.
  • the collection management user interface of Figure 3 also enables the user to tag an entire category/cluster to more efficiently add tags to the items in that category.
  • the tags "classical music' " and "guitar” have already been assigned to the first category, and the user can click on "add a tag " to add a new tag, or on the 'x * to the right of the tag to remove it.
  • Tags can also be added at the item level, and a given item can have any number of tags.
  • the user can request recommendations that are specific to a given tag. For example, if the user has assigned the tag "books for my kids" to
  • cluster-level ratings, tags, and names that are assignable via the UI are all examples of cluster-level metadata that can be attached to the item collection. As illustrated by the above examples, the ability to attach cluster-level metadata improves the system's ability to provide useful recommendations to users.
  • the web page in Figure 3 also displays a corresponding "recs based on these items " button. Selection of such a button causes the system to immediately generate and return recommendations that are based solely on these items.
  • the items in the cluster are passed to the recommendation engine/service as the recommendation sources, if some of these items have been rated by the user, these ratings may be used by the recommendation service as item weights. Because the recommendations are specific to a particular cluster or interest selected by the user, they are very likely to be useful.
  • the right pane 56 in Figure 3 displays, for each category/cluster, the user ' s level of interest in the category as a function of time. This gives the users insight into their own purchasing patterns over time.
  • the levels of interest may, for example, be based on the user ' s purchase activity.
  • the right pane 56 displays a marker indicating the location (time) of purchase in the associated graph.
  • the graphs can be generated by computing the normalized distribution of items in the respective clusters for each point in time, giving the users an idea of their relative interest at that time.
  • Another possibility for the graph would be to generate box and whisker plots for each of the clusters, based on the timestamps of the items in the cluster. Such a graph would show abrupt shifts in interests, for example.
  • the display of such an interface may be optional or configurable by the user.
  • Another feature of some embodiments involves the use of clustering to filter the recommendations generated by a recommendation service or engine.
  • the goal of this feature is to improve the quality of the recommendations by filtering out items similar to those in which the user is "known " to lack interest. This is preferably accomplished by clustering together items the user has marked/rated as "not interested,” and then filtering out any recommended items that are similar/close to the resulting "not-interested" cluster or clusters.
  • FIG. 4 illustrates one embodiment of this process.
  • an appropriate collection of items associated with the target user is initially passed to a clustering engine 60.
  • This collection preferably includes or consists of items marked/rated by the target user as "not interested," although other indicia of the user's lack of interest may be used (e.g., ratings falling below a threshold level).
  • the clustering engine 60 applies an appropriate clustering algorithm to this collection, and returns a set of data describing the resulting clusters.
  • the clustering engine 60 may, but need not, use the distance metric and clustering algorithms described in Sections V-VIIl below.
  • the clusters returned by the clustering engine 60 are the not-interested clusters used for filtering. If, on the other hand, the collection includes or consists of other types of items (e.g., items purchased and/or rated on a scale of 1-5 by the user), the resulting clusters are analyzed to identify the clusters, if any, that can be classified as a not-interested clusters. For example, a cluster may be treated as a non-interested cluster if it consists primarily of items rated as "not interested” or rated at less than a threshold level. The methods described in Section VIII for reducing "high entropy" clusters may optionally be used to generate the not-interested clusters.
  • a set of source items is passed to a recommendation engine 62.
  • the set of source items may, for example, include or consist of items purchased or rated highly by the user, and preferably does not include any of the not- interested items.
  • the set of source items may, but need not, be selected using the process described in Section I.
  • the recommendation engine 62 may use a number of recommendation methods, such as, but not limited to, those described in U.S. Pat. 6,912,505, to select items to recommend.
  • a recommendation engine that uses traditional collaborative filtering methods may also be used.
  • the next phase of the process shown in Figure 4 involves the use of a filtering component or system 64 to filter out any recommendations that are similar to any of the not-interested clusters. This may be accomplished by, for example, measuring the distances between the recommended items and the cluster centers of the not- interested cluster(s). and removing any recommended items that fall within a threshold distance of one of these cluster centers. By removing these items, the likelihood that the recommendations will be useful to the user is significantly increased.
  • the distances between these items and the centers of the user's "interest clusters " may also be considered.
  • the decision whether to filter out a recommended item may be based on both (1) the distance between that item and the center of the closest not- interested cluster, and (2) the distance between that item and the center of the nearest interest cluster.
  • the recommended item may be filtered out if its distance to the center of the closest not-interested cluster is both (a) less than a selected threshold, and (b) less than its distance to the center of the nearest interest cluster.
  • Various other factors, such as the sizes of these clusters, may also be considered.
  • the process of Figure 4 can be implemented without using "not interested" ratings to identify the items in which the user lacks interest.
  • the not- interested clusters can be identified based on items explicitly marked as gifts, and/or based on items to which the user has given a rating less than a designated threshold.
  • Another feature of some embodiments involves the use of clustering to organize a set of recommended items into categories for presentation to the user.
  • Each category may correspond uniquely to a particular cluster, and may automatically be named based on attributes of the items falling in that cluster.
  • This feature may, but need not, be used in combination with one or more of the features described in Sections I-III.
  • FIG. 5 illustrates this process according to a first embodiment.
  • a set of recommended items is generated for a user.
  • the recommendations may be generated using any type of recommendation engine and process, including the various processes described herein.
  • a clustering algorithm is applied to the set of recommended items.
  • the clusters may, but need not, be formed using the distance metric and clustering algorithms described in Sections V-VlIl.
  • a category name is assigned to each cluster based on attributes of the items in that cluster. This may be accomplished in various ways. For example, if the items are arranged within a hierarchical browse structure, the name of the lowest level category/browse node common to all items in the cluster may be used. As another example, if subject keywords or keyword phrases are available for the items, the subject keyword or keyword phrase that appears the most frequently within the cluster may be used. Additionally, if the user has manually chosen a name for the cluster via some UI such as the one proposed in Section II, it is preferable that the user ' s selected name have priority over any automatically generated one. The names may be assigned such that no two clusters are assigned the same name.
  • data may be obtained from a catalog or extracted from user tags.
  • Tags may be extracted from the target user's tags, or from other users who have similar clusters.
  • One method of choosing which terms to use is to run a type of part-of- speech tagging algorithm on the keyword/subject term phrases, and extracting only nouns. These extracted words can then be weighted based on the frequency of their occurrence in an item catalog or some other source. For example, in the case of books, the actual text of the books may be used to evaluate the relevancy of the keyword.
  • step 76 the recommended items, as arranged by cluster/category, are output to the user together with the associated category names selected in step 74.
  • the category names can be presented in a "browse cloud " interface in which each category name is displayed as a selectable link to the corresponding list of items, and in which the font size of each such name/link is directly proportional to the number of items in the corresponding category/cluster (see Figure 7, discussed below).
  • the recommended items can be displayed in a list format with category headings.
  • the clustering algorithm assigns each item to a single cluster
  • the categories generated by the process of Figure 5 are mutually exclusive, meaning that each item appears in only one of the categories. This is a significant benefit over existing browse cloud displays in which the same recommended item will commonly appear in multiple categories of the browse cloud.
  • Figure 6 illustrates a second embodiment of a process of arranging the recommended items into mutually exclusive categories or clusters.
  • a clustering algorithm is applied to an appropriate item collection of the target user.
  • this item collection may. for example, include or consist of items purchased and/or rated by the target user.
  • the item collection may, for example, include or consist of news articles viewed (or viewed for some threshold amount of time) by the target user.
  • step 82 the clusters resulting from step 80 are optionally analyzed to identify those that likely represent actual interests of the target user. Any of the methods and criteria described in the preceding sections may be used for this purpose. Any clusters identified as likely not representing an interest of the user are excluded from the subsequent steps. The desirability of performing this step 82 may depend on the nature of the item collection. For instance, if the item collection consists of items rated favorably by the target user, this step 82 may be omitted.
  • step 84 a category name is assigned to each remaining cluster. This may be accomplished using one of the methods described above for step 74 of Figure 5.
  • step 86 item recommendations are generated for the target user.
  • the recommendations may be generated using any type of recommendation engine and process, including the various processes described herein. If the recommendations are generated based on a set of source items, the item collection clustered in step 80, or a selected subset of these items, may be used as the sources.
  • step 88 the process attempts to match or "assign" each recommended item to one of the interest clusters, and thus to one of the category names, resulting from steps 80-84. This may be accomplished by, for example, calculating the distances between the recommended items and the cluster centers, and assigning each recommended item to the interest cluster whose distance is shortest. Other factors, such as the sizes and entropy levels of the interest clusters, may also be considered.
  • the effect of this step 88 is to subdivide all or a portion of the set of recommended items into multiple clusters, each of which corresponds to a previously-identified interest of the user.
  • Recommended items that are more than a threshold distance from the closest interest cluster may be filtered out (not displayed), or may be displayed under a category name (e.g., "more recommended items” or "all categories '" ) that does not correspond to any particular interest cluster.
  • category name e.g., "more recommended items” or "all categories '"
  • the filtering process shown in Figure 4 and described above may also be used to assess whether particular recommended items should be filtered out.
  • step 90 the recommended items (excluding any that have been filtered out), as arranged by interest cluster/category, are exposed to the user in association with the corresponding category names selected in step 84.
  • a browse cloud interface may optionally be used for this purpose.
  • Figure 7 is a screen display illustrating one example of a browse cloud interface that may be used.
  • the images and titles of the recommended items in all categories are displayed in a scrollable format in the upper portion of the display.
  • the user can click on one of the category names (e.g., "Action & Adventure, ' ' or "Humorous " ') in the lower portion of the display.
  • category names e.g., "Action & Adventure, ' ' or "Humorous " '
  • the text size of each such category name is generally proportional to the number of recommended items in the associated cluster. If, for example, the user clicks on "Action & Adventure, " the upper portion of the display will be updated to show only the items falling in the corresponding cluster.
  • This section describes one possible distance metric that can be used to calculate distances between items.
  • This distance metric may be used where the items being clustered are arranged in a hierarchical browse structure, such as a directed acyclic graph. As discussed above, the distances may additionally or alternatively be calculated based on other criteria. [0094] The distance metric is based on the degree to which two items fall under or ''share " the same browse nodes, with greater weight given to lower-level nodes (those containing smaller numbers of items). For example, suppose that the paths to two items, A and B, are as follows (it is assumed in this simple example that there is only one path to each item):
  • the "Business Skills" browse node i.e., the lowest level browse node shared by the two items
  • the "Products '" browse node would be given the least weight.
  • the weight given to each of the three shared browse nodes would be inversely proportional to the total number of items falling below that browse node. In a more typical scenario, multiple paths will exist to a given item, and each such path should be taken into consideration.
  • the numerator means that the more nodes the two items have in common, the more similar they are to each other.
  • the intersecting nodes are also weighted by the inverse probability of the individual nodes, w , Nodes of low probability (e.g., leaf nodes) are assumed to carry more information, and thus will increase the size of the numerator.
  • the denominator normalizes the score based on the size of the union, so that shallow parts of the browse hierarchy are not unduly penalized.
  • each item is initially represented as a set of all browse node IDs appearing in the item ' s ancestry.
  • Nl root node
  • each item it will be assumed that only one path exists from the root node, Nl, to each item..
  • the only path to 13 is ⁇ 1 , 3, 12> (where each number represents a browse node); that the only path to 14 is ⁇ 1 , 4, 12>; that the only path to 15 is ⁇ 1, 4, 8>; that the only path to 17 is ⁇ 1, 5, 10, 13>; and that the only path to 18 is ⁇ 1, 5, 1 1, 13>.
  • items 11 through 18 in Figure 8 are represented, respectively, by the following eight browse node sets or LL points:"
  • X 1 (1 ,2,7)
  • x 3 (1 ,3,12)
  • x 4 (1,4,12)
  • x 5 (1 ,4,8,12)
  • item Il could also appear under browse node N7, in which case it would be represented by x ⁇ ⁇ ⁇ 1 , 2, 6, 7>.
  • a table of node data values or "probabilities" is first computed.
  • the probability P(w) of any browse node w is preferably calculated as the number of times it occurs in the dataset divided by the number of points in the dataset ⁇ since each browse node can only occur once per point): ⁇ i if w e x
  • the distances between any two points in this dataset can be computed.
  • the distance between points x 4 and x 5 is computed as follows:
  • node 4 is given the most weight (i.e., has a greater effect at lowering the distance measurement) since it appears less frequently than nodes 1 and 32 in the dataset,
  • the distance function just presented represents one possible way to measure the distance between two arbitrary sets of browse nodes.
  • One variation to this method is to take into consideration the "conditional probabilities' " of the browse nodes. The benefits of using conditional probabilities become apparent when the relevant items belong to several parts of the browse hierarchy and some paths should have greater weight than others for a given item.
  • the distance metric is modified by adding conditional terms representing the probability of a browse node given an item, P(w j x e X) , where w is the browse node, and x is some item in the dataset X. Folding this into our distance function yields:
  • conditional is computed by summing the number of times each browse node appears in the item's ancestry, and dividing by the total number of paths in this ancestry.
  • conditional is computed by summing the number of times it occurs in all items in the cluster divided by the total number of ancestries in the cluster.
  • the amount of weight given to a particular browse node when calculating the distance between one pair of items may be different than the amount of weight given to this same browse node when calculating the distance between another pair of items.
  • more weight is given to the common portions of its ancestry. The more times a particular node appears in the different ancestry paths, the more weight it will contribute to the distance computation. This will discourage linking items together based on rather obscure assignments, making the distance computation more robust to some of the more questionable assignments in the browse tree.
  • the item distances may be calculated by a distance calculation system or service component that is part of a larger system. Once distances have been calculated and stored for all pair of items, the stored distance values may be used by an appropriate clustering algorithm to cluster together similar items.
  • IsoModes This section describes one possible clustering algorithm that may be used to implement the clustering phases of the various processes described above.
  • the algorithm is referred to herein as IsoModes, and is a variation of the well known IsoData clustering algorithm.
  • the IsoModes algorithm is one of many algorithms that can be used.
  • the IsoModes algorithm can be applied using the distance metric of Section V or any other distance metric.
  • IsoData stands for “Iterative Self-Organizing Data Analysis Technique.” It is self-organizing in that it differs from standard clustering techniques such as K-Means, where the number of clusters must be pre-specified. Since the number of clusters is generally unknown, IsoData attempts to automatically determine this number by optimizing other criteria, such as the ratio of inter-to-intra cluster scatter.
  • Intra-cluster scatter is defined to be the total sum of distances from points in a cluster to the cluster center. That is:
  • S mlenial is the internal scatter.
  • A is the cluster, and A c is the cluster center.
  • Inter-cluster scatter is the sum of the distances from cluster centers to the global mean. Intuitively, maximizing the ratio of inter-to-intra cluster scatter will favor groupings where the points within a cluster are similar to each other, and points from different clusters will be separated as much as possible.
  • the optimization process is done by repeatedly assigning points to their nearest cluster and re-computing the cluster centers. At the end of each iteration, clusters which have very large scatter are split, and clusters which are too close to other clusters are broken up and reassigned to existing clusters. This procedure is repeated until the total scatter converges to within some margin of error, there is no change in the clustering in consecutive iterations, or some maximum number of iterations has been reached. If the distance metric of Section V is used, the cluster centers are preferably defined as sets of browse nodes, so the points and centers are fundamentally the same and are directly comparable.
  • IsoModes One significant difference between IsoModes and the classical IsoData algorithm is in the method used to split large clusters. In Euclidean space, a cluster with large scatter can be split in the direction of the largest eigenvector. This will maximize the separation of items when splitting the cluster. However, for nominal data, there are no notions of eigenvalues or eigenvectors, so the cluster split is not something that can be directly computed.
  • the 2-Modes algorithm is very similar to IsoModes, except that the number of classes is simply predetermined to be 2, giving us our desired 2-way cluster split.
  • IsoModes and IsoData Another minor difference between IsoModes and IsoData is in the stopping criterion; instead of stopping when the global scatter falls below some epsilon, it is stopped when the clustering stops changing. This is desirable because, among other reasons, it is very easy for the global mean to become degenerate when computing the center for nominal data.
  • N limit on maximum number of iterations
  • CC 1 computeClusterCenters( C 1 x )
  • the cluster centers are initialized by randomly selecting k points, and the points in the dataset are assigned to the closest centers.
  • the main iteration begins by recomputing the cluster centers from the previous assignment, and then re-computing the assignment until there is no change in the clustering, or the maximum number of iterations is reached.
  • the algorithm checks all clusters for any with large scatters, or which are too close to neighboring clusters. Clusters satisfying these criteria are split and reassigned as necessary.
  • Step 2. Compute an initial cluster assignment by assigning all of the data points to the nearest cluster centers.
  • Table 3 displays the computed distances from all points to all clusters. Assignments are denoted with (*). Ties are broken at random.
  • Step 3 Check to see if clusters have not changed or if we have exceeded the maximum number of iterations.
  • Step 4 Compute the cluster centers. This is done by taking all nodes that appear in greater than 50% of the points in each cluster computed in the previous step (e.g. Cluster 1 (C 1 ), was assigned (1,2,6) and (1 ,5,10,13) ). The results of this step are shown in Table 4.
  • Step 5 Re-compute the cluster assignments, with ties again broken at random. The results of this step are shown in Table 5.
  • Step 6 the clusters are evaluated to split clusters with very large scatters, or reassign clusters which are too close to other clusters. Even though the first two clusters in this example are too similar to each other, for brevity, we will assume no splits or reassignments occur. In the actual algorithm, one of the duplicate clusters would be removed, with the items randomly distributed among the remaining clusters.
  • Step 7. Re-compute the centers using assignments from previous iteration. The results of this step are shown in Table 6.
  • Step 8. Re-compute cluster assignments. The results are shown in Table
  • Step 9 Check again for scatter and minimum distance. In this case, no splits or reassignments occur.
  • Step 10 Re-compute the centers using assignments from previous iteration. The results of this step are shown in Table 8.
  • One drawback to the lsoModes algorithm is that it is very sensitive to the cluster center initialization.
  • the algorithm is essentially a form of hill-climbing, and as such, the starting point can dictate how many iterations are required to convergence, as well as whether converge to a global optimum is possible. If the wrong starting point is chosen, converge to a local optimum may occur, resulting in a suboptimal clustering.
  • the issue can be addressed by, among other solutions, performing a type of hierarchical sampling prior to clustering the entire dataset. For example, the system can initially choose a very small sample size of 50 items, run the algorithm, and then use the resulting clustering as a seed to progressively larger samples until the entire dataset is clustered. This technique results in faster convergence in comparison with random initialization.
  • Some of the clusters resulting from the lsoModes algorithm may have an entropy (in terms of user like/dislike) that is too high for a given application.
  • a high entropy signifies either that the user has a love/hate relationship with the cluster's items, or that the cluster is too broad and should be split.
  • a low entropy signifies that the user strongly likes or dislikes the items in the cluster.
  • high entropy clusters are re- clustered using the 2-Modes algorithm described above, except the initial seeding is based on a positive/negative classification of the items. Where sufficient ratings data is available for the particular user, this may be accomplished as follows.
  • Each item in the relevant cluster is initially classified as having a positive rating or a negative rating, or is removed from consideration if no such rating can be assigned. For example, if the user ratings are based on a 5-star plus "non -interested " rating system, items with a 1 or 2 star rating may be classified as negative: items with a 3 or 4 star rating may be classified as positive, and all other items may be disregarded.
  • the cluster ' s homogeneity may be determined by computing its entropy. This process is summarized by the following equations:
  • Figure 9 illustrates how the various features described above may be implemented in the context of a web-based system that provides functionality for users to browse and purchase items from an electronic catalog.
  • the various features of the disclosed embodiments can also be implemented in other types of systems, including, e.g.. music download systems that recommend music titles based on users" music download histories, video rental sites that recommend video/DVD titles based on users video rental histories, news sites that recommend news articles based on users' news browsing histories, browser toolbar based systems that recommend web sites and other network resources based on users' web browsing histories, web services for providing recommendations, checkout/wallet services that enable users to pay for goods from various participating merchants, and Internet advertising systems.
  • the system shown in Figure 9 includes one or more web server machines 100 that generate and serve pages of a host web site in response to page requests from user computing devices 102.
  • the web servers 100 provide user access to a catalog of items represented in a database 108 or collection of databases.
  • the items preferably include or consist of items that may be purchased via the web site (e.g., book, music and video titles in physical or downloadable form; consumer electronics products; household appliances; magazine and other subscriptions, etc.).
  • the database 108 also stores data regarding how the items are arranged within a hierarchical browse structure. Data regarding the catalog items and the browse structure is accessible via a catalog service 106, which may be implemented as a web service.
  • the system also includes a data repository 116 (e.g., one or more databases) that stores various types of user data, including identifiers of the items in each user's collection.
  • the data repository 116 may store users' purchase histories, item viewing histones. item ratings, and item tags. The purchase histories and item viewing histories may be stored as lists of item identifiers together with associated event timestamps.
  • the various types of user data maybe accessible to other components of the system via a data service (not shown), which may be implemented as a web service.
  • the system also includes a recommendation service/system 1 10 that generates recommendations in real time in response to requests from users.
  • the recommendation service 1 10 includes a recommendation engine 62, and includes a cluster generation and analysis component/system 1 12 that implements some or all of the clustering- related features described herein, including the calculation of item distances.
  • a cluster generation and analysis component/system 1 12 that implements some or all of the clustering- related features described herein, including the calculation of item distances.
  • Separate executable components may optionally be provided for performing, e.g., distance calculations, item clustering, and cluster assessment; these components may all run on a single computer or on separate computers.
  • the clustering-related features may alternatively be implemented as a separate service.
  • a separate cluster service can be provided that receives an input list of items (together with appropriate input parameters) and outputs data describing the resulting set of clusters.
  • This same service could, e.g., be configured to select source items from the input list; to assign names to the clusters, to score the clusters, and to perform various other cluster- related tasks.
  • the recommendation engine 62 may operate as described in U.S. Patent 6,912,505, referenced above. More specifically, the recommendations service may use one or more "similar items" tables/datasets 120 to look up items that are similar or related to the source items, together with associated data values indicating the strengths of such relationships.
  • the similar items table(s) 120 may be generated off-line by an item association mining component 1 18 that analyzes users ' purchase histories, item viewing histories, or some other type of user activity data, and detects and quantifies behavior-based associations between specific items. For instance, if purchase histories are used, item A may be mapped to item B in a purchase-based similar items table 120 if a relatively large number of the users who purchased item A also purchased item B.
  • Other types of recommendation engines including recommendation engines that do not use item-to-item mappings, may also be used.
  • the electronic catalog system may also include services for handling various other types of tasks, such as user authentication, transaction processing, search query processing, storing user-assigned tags and ratings, processing of user-submitted sales listings, etc.
  • the web servers 100 use a data repository of web page templates 104 to dynamically generate web pages in response to browser requests.
  • the templates directly or indirectly specify the service calls that are made to the services to, e.g., request data needed to generate the requested page.
  • an appropriate template may be provided for generating collection management pages of the type shown in Figure 3, and for generating item detail pages, browse node pages, recommendation/browse cloud pages of the type shown in Figure 7, and various other pages of the site.
  • a web server 100 requests recommendations for the user from the recommendations service 1 10.
  • the recommendation service 1 30 uses all or a portion of the user ' s purchase history, item ratings, and/or item viewing history (typically depending upon the context of the user's request) to generate the recommendations.
  • the recommendations service 1 10 may use the cluster-based process described in Section I to select the particular items to use as recommendation sources.
  • the recommendation service 100 may additionally or alternatively use the cluster-based filtering process described in Section 111 to filter the set of items generated by the recommendation engine 62. Additionally or alternatively, recommendation service 100 may use the process described in Section IV to organize the recommendations into a set of cluster-based categories for display to the user.
  • the resulting list of recommended items (or a portion of this list), and/or the names of the cluster-based categories in which these items fall, is incorporated into the requested web page for transmission to the user ' s browser/computer 302.
  • the services and other application components 62, 106, 110, 1 12, and 118 shown in Figure 9 may be implemented in software code modules executed by any number of general purpose computers or processors, with different services optionally but not necessarily implemented on different machines interconnected by a network.
  • the code modules may be stored in any type or types of computer storage, such as hard disk drives and solid state memory devices.
  • the various data repositories 104, 108, 120 may similarly be implemented using any type of computer storage, and may be implemented using databases, flat files, or any other type of computer storage architecture.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Finance (AREA)
  • Databases & Information Systems (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne des procédés permettant de regrouper des articles et d'améliorer l'utilité de recommandations d'articles. Un de ces procédés consiste à appliquer un algorithme d'agrégation (36) à une collection d'articles d'utilisateur, et à utiliser ensuite des informations concernant les agrégats obtenus afin de sélectionner des articles à utiliser comme sources de recommandations (38). Un autre procédé consiste à présenter les agrégats d'articles à l'utilisateur par l'intermédiaire d'une interface de gestion de collections qui permet à l'utilisateur de relier des métadonnées par niveau d'agrégation, par exemple, en classant ou en étiquetant des agrégats entiers d'articles. Les métadonnées obtenues peuvent être utilisées pour améliorer les recommandations générées par un moteur de recommandations (62). Un autre procédé consiste à utiliser un moteur d'agrégation (60) pour former des agrégats d'articles dans lesquels un utilisateur a indiqué un manque d'intérêt et utiliser ces agrégats pour filtrer (64) la sortie d'un moteur de recommandations (62).
EP08744674A 2007-03-30 2008-03-28 Évaluation sur la base d'une agrégation d'intérêts utilisateurs Ceased EP2145265A4 (fr)

Applications Claiming Priority (5)

Application Number Priority Date Filing Date Title
US11/694,821 US7966225B2 (en) 2007-03-30 2007-03-30 Method, system, and medium for cluster-based categorization and presentation of item recommendations
US11/694,815 US7743059B2 (en) 2007-03-30 2007-03-30 Cluster-based management of collections of items
US11/694,720 US8095521B2 (en) 2007-03-30 2007-03-30 Recommendation system with cluster-based filtering of recommendations
US11/694,707 US7689457B2 (en) 2007-03-30 2007-03-30 Cluster-based assessment of user interests
PCT/US2008/058745 WO2008121872A1 (fr) 2007-03-30 2008-03-28 Évaluation sur la base d'une agrégation d'intérêts utilisateurs

Publications (2)

Publication Number Publication Date
EP2145265A1 true EP2145265A1 (fr) 2010-01-20
EP2145265A4 EP2145265A4 (fr) 2011-09-14

Family

ID=39808691

Family Applications (1)

Application Number Title Priority Date Filing Date
EP08744674A Ceased EP2145265A4 (fr) 2007-03-30 2008-03-28 Évaluation sur la base d'une agrégation d'intérêts utilisateurs

Country Status (2)

Country Link
EP (1) EP2145265A4 (fr)
WO (1) WO2008121872A1 (fr)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9361640B1 (en) 2007-10-01 2016-06-07 Amazon Technologies, Inc. Method and system for efficient order placement
US10097664B2 (en) 2013-04-26 2018-10-09 Apple Inc. Recommending media items based on purchase history
CN104636354B (zh) 2013-11-07 2018-02-06 华为技术有限公司 一种位置兴趣点聚类方法和相关装置
US10902442B2 (en) 2016-08-30 2021-01-26 International Business Machines Corporation Managing adoption and compliance of series purchases
CN107463683B (zh) * 2017-08-09 2018-07-24 深圳壹账通智能科技有限公司 代码元素的命名方法及终端设备
US11893543B2 (en) * 2018-05-15 2024-02-06 International Business Machines Corporation Optimized automatic consensus determination for events
CN109145222A (zh) * 2018-09-11 2019-01-04 合肥汇众知识产权管理有限公司 专利交易信息的推送方法及装置
CN112487276B (zh) * 2019-09-11 2023-10-17 腾讯科技（深圳）有限公司 一种对象获取方法、装置、设备及存储介质
CN111046300A (zh) * 2019-12-17 2020-04-21 智者四海（北京）技术有限公司 确定用户的人群属性的方法及装置
CN112381598B (zh) * 2020-10-26 2023-12-05 泰康保险集团股份有限公司 产品服务信息推送方法及装置
CN112669128B (zh) * 2021-01-28 2022-03-22 天元大数据信用管理有限公司 一种实时聚类推荐的方法、设备及介质
CN113761379B (zh) * 2021-09-17 2024-04-16 北京百度网讯科技有限公司 商品推荐方法及装置、电子设备和介质
CN114896491B (zh) * 2022-04-18 2023-06-06 上海二三四五网络科技有限公司 基于聚类的用户个性化热门计算方法、装置、设备及介质
CN114780861B (zh) * 2022-06-20 2022-10-21 上海二三四五网络科技有限公司 基于聚类技术的用户多兴趣推荐方法、装置、设备及介质
CN117194804B (zh) * 2023-11-08 2024-01-26 上海银行股份有限公司 一种适用于经营管理系统的引导性推荐方法及系统

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6064980A (en) * 1998-03-17 2000-05-16 Amazon.Com, Inc. System and methods for collaborative recommendations
US6321221B1 (en) * 1998-07-17 2001-11-20 Net Perceptions, Inc. System, method and article of manufacture for increasing the user value of recommendations
US6334127B1 (en) * 1998-07-17 2001-12-25 Net Perceptions, Inc. System, method and article of manufacture for making serendipity-weighted recommendations to a user
US7720723B2 (en) * 1998-09-18 2010-05-18 Amazon Technologies, Inc. User interface and methods for recommending items to users
US6266649B1 (en) * 1998-09-18 2001-07-24 Amazon.Com, Inc. Collaborative recommendations using item-to-item similarity mappings
US6925460B2 (en) * 2001-03-23 2005-08-02 International Business Machines Corporation Clustering data including those with asymmetric relationships
JP2003242157A (ja) * 2002-02-20 2003-08-29 Honda Motor Co Ltd 顧客の特性に応じた商品を検索するコンピュータシステム
EP1540550A4 (fr) * 2002-08-19 2006-09-27 Choicestream Systeme de recommandation statistique personnalise
US7756879B2 (en) * 2004-07-23 2010-07-13 Jeffrey Parsons System and method for estimating user ratings from user behavior and providing recommendations

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
No further relevant documents disclosed *
See also references of WO2008121872A1 *

Also Published As

Publication number Publication date
EP2145265A4 (fr) 2011-09-14
WO2008121872A1 (fr) 2008-10-09

Similar Documents

Publication Publication Date Title
US7689457B2 (en) Cluster-based assessment of user interests
US7743059B2 (en) Cluster-based management of collections of items
US7966225B2 (en) Method, system, and medium for cluster-based categorization and presentation of item recommendations
US8560545B2 (en) Item recommendation system which considers user ratings of item clusters
US8019766B2 (en) Processes for calculating item distances and performing item clustering
WO2008121872A1 (fr) Évaluation sur la base d'une agrégation d'intérêts utilisateurs
US8175989B1 (en) Music recommendation system using a personalized choice set
US8301623B2 (en) Probabilistic recommendation system
JP4378646B2 (ja) 情報処理装置、情報処理方法、およびプログラム
US8285602B1 (en) System for recommending item bundles
US7668821B1 (en) Recommendations based on item tagging activities of users
US20060173753A1 (en) Method and system for online shopping
US20090089273A1 (en) System for detecting associations between items
US20090089314A1 (en) Duplicate item detection system and method
US20200265491A1 (en) Dynamic determination of data facets
CN102163211A (zh) 信息处理设备、重要度计算方法和程序
US20070192317A1 (en) Method of assessing consumer preference tendencies based on correlated communal information
Badriyah et al. Recommendation system for property search using content based filtering method
US8738459B2 (en) Product recommendation
Gupta et al. A Nobel Approach to Recommend Online Courses by using Hybrid Recommendation System
Koneru Deep learning-based automated recommendation systems: a systematic review and trends
Paul et al. User-item-based hybrid recommendation system by employing mahout framework
Lu et al. Clustering e-commerce search engines based on their search interface pages using WISE-Cluster
Mohan et al. Recommendation system in business intelligence solutions for grocery shops: Challenges and perspective
Schwartz et al. Style similarity as feedback for product design

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20091030

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MT NL NO PL PT RO SE SI SK TR

AX Request for extension of the european patent

Extension state: AL BA MK RS

DAX Request for extension of the european patent (deleted)
A4 Supplementary search report drawn up and despatched

Effective date: 20110811

RIC1 Information provided on ipc code assigned before grant

Ipc: G06F 17/30 20060101ALI20110805BHEP

Ipc: G06Q 30/00 20060101AFI20110805BHEP

17Q First examination report despatched

Effective date: 20160218

REG Reference to a national code

Ref country code: DE

Ref legal event code: R003

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION HAS BEEN REFUSED

18R Application refused

Effective date: 20180210