Systems and methods for determining user interests
View Patent ↗Techniques are provided to determine user-interest features and user-interest parameter weights for a user-interest model. The user-interest features are pre-determined and/or determined dynamically. Pre-determined user-interest features are based on user-interest profiles, prior user activities, documents listed in a resume, reading or browsing patterns and the like. Dynamically determined user-interest features include features learned from an archive of user activities using statistical analysis, machine learning and the like. User-interest parameter weights are pre-determined and/or dynamically determined. Pre-determined user-interest parameter weights include parameter weights manually entered by a user indicating the relevant importance of a user-interest feature and parameter weights previously learned from an archive of the user's past activities. Dynamically assigned user-interest parameter weights include dynamically determined updates to user-interest parameter weights based on newly identified documents or topics of interest.
1. In a computer-based search system configured to receive a search query from a user and, in response to the search query, return documents or links to documents in a repository of documents to the user, a computer-implemented method of determining a user-interest model for the user, comprising the steps of:
determining the repository of documents, the repository of documents comprising at least one of audio, video and textual documents associated with the user;
determining interesting predicates within each of the documents based on machine learning;
determining user-interest features for each class of the interesting predicates, the user-interest features comprising at least one of terms, phrases, and concepts related to the terms or phrases of interest to the user;
determining user-interest identifying features for each conceptual abstraction of the documents;
determining parsing features, the parsing features including at least one of parsing functions and parsing rules;
determining transfer features, wherein the transfer features are associated with a count of the number of times a transfer function is applied to each of the documents;
determining parameter weights for the user-interest features;
filtering redundant features from the user-interest features;
determining a probability distribution based on the user-interest features and the parameter weights;
determining a user-interest model based on the user-interest features and the parameter weights:
casting the user-interest features in a log-linear probability model; and
estimating each associated parameter weight using L1-regularized maximum entropy estimation.
2. The method of claim 1 , wherein estimating each associated parameter weight comprises estimating the associated parameter weight by L1-regularized maximum-entropy estimation from training data.
3. The method of claim 2 , wherein estimating the associated parameter weight by L1-regularized maximum-entropy estimation from training data comprises:
stochastic learning of the user-interest features and the parameter weights based on training data comprising pairs of sentences y, and linguistic structures s which
comprise verified representations of the user-interest features; and
minimizing a negative conditional log-likelihood L(λ) of linguistic structures given sentences, modulo a regularization term R(λ) on the parameters for training data {(sj, yj)}, j=1, . . . , m and the set of possible structures S(y) for each sentence y.
4. The method of claim 3 , wherein minimizing a negative conditional log-likelihood L(λ) of linguistic structures comprises:
optimizing
L
(
λ
)
+
R
(
λ
)
=
-
1
m
∑
j
=
1
m
ⅇ
λ
·
f
(
sj
)
∑
s
∈
S
(
yj
)
ⅇ
λ
·
f
(
sj
)
+
γ
λ
1
1
using a conjugate gradient optimization utilizing a form of the γ-weighted L1 norm γ∥λ∥ 1 1 to perform automatic feature selection that discards noisy and redundant features, wherein the term ƒ comprises feature functions which include count functions based on at least one of transfer functions, parse functions, and keep indicators, wherein the stochastic learning comprises estimating the parameter weights λ associated with the feature functions ƒ according to the log-linear probability model.