Li/=SrSSKrSSKrSSKJrJrJrJrJ r SSK r SSK J r SSK Jr \ R"\5r"SS5rg)XUser preference learner using average vector method with positive and negative feedback.N)ListOptionalDictAnyTuple) MemoryStore)PatternEmbedderc \rSrSrSrSS\S\4SjjrS\\ \ R\ \ R44Sjr S\ \ R4Sjr S \S\ \ R4S jrSS \S \S\ \ R4S jjrSS\\\\4S \ \S\S\ \\4SjjrS\\\\4S\\S \ \S\\4SjrS\\\\4S\ RS\\4SjrS\ RS\ RS\S\ R4SjrSS \S\S\ \ R4SjjrSS\\\\4S \ \S\S\ \\4SjjrSrg ) PreferenceLearnerr decay_lambdause_preference_weightedcX[5Ul[5UlXlX lg)uInitialize preference learner. Args: decay_lambda: Decay parameter for intent vector weight (α_t = e^(-λt)) use_preference_weighted: If True, multiply similarity by preference weight (adaptive fusion, 2025-style) N)r memoryr embedderrr)selfrrs C/home/ubuntu/codebase/yexijia/保研/iCoLoc/src/learning/learner.py__init__PreferenceLearner.__init__s#"m ') ('>$returnc\URR5n/n/nURS/5GHnURS05nURS/5nUH_n[U[5(aURS/5nO[U[ 5(aUnOMEU(dMNUR U5 Ma URS/5n U H_n[U[5(aURS/5nO[U[ 5(aUnOMEU(dMNUR U5 Ma GM Sn [U5S:adURRU5n U bF[U 5S:a7[R"U SS9n [RS [U5S 35 Sn [U5S:adURRU5n U bF[U 5S:a7[R"U SS9n [RS [U5S 35 U cU cg X4$)aBuild user preference vectors from historical feedback. Returns: Tuple of (positive_vector, negative_vector), or (None, None) if no feedback - positive_vector: Average vector of liked patterns - negative_vector: Average vector of disliked patterns sessionsfeedbacklikepatterndislikeNraxiszBuilt positive vector from z liked patternszBuilt negative vector from z disliked patterns)NN)rloadget isinstancedictlistappendlenrencode_patternsnpmeanloggerdebug) rdataliked_patternsdisliked_patternssessionrliked pattern_datardislikedpositive_vectorvectorsnegative_vectors rbuild_user_vectors$PreferenceLearner.build_user_vectorss{{!xx B/G{{:r2HLL,E % lD11*..y"=G d33*G7"))'2!& ||Ir2H ( lD11*..y"=G d33*G7%,,W5!)%0< ~  "mm33NCG"s7|a'7"$'''": :3~;N:O_`  !A %mm334EFG"s7|a'7"$'''": :3?P;Q:RRdef  "'>//rc*UR5upU$)zBuild user preference vector from historical feedback (backward compatibility). Returns: User preference vector (average of liked patterns), or None if no feedback )r7)r positive_vec_s rbuild_user_vector#PreferenceLearner.build_user_vector]s 113 ruser_idcURRU5nU(dgURS/5nURS/5n[R"UR R 55nU(a:[R"UVs/sHn[R"U5PM snSS9n[R"UR R 55nU(a:[R"UVs/sHn[R"U5PM snSS9nXW- n[RRU5S:XagU$s snfs snf)uCompute user preference vector from feedback (μ⁺ - μ⁻). Args: user_id: User identifier Returns: Feedback-based user vector (positive_mean - negative_mean), or None if no feedback Npositivenegativerr) rget_user_profiler"r)zerosr get_sentence_embedding_dimensionr*arraylinalgnorm) rr> user_profilepositive_vectorsnegative_vectorsmean_posvmean_neg user_vectors rcompute_feedback_vector)PreferenceLearner.compute_feedback_vectorfs {{33G< '++J;'++J;88DMMJJLM ww5EF5E 5EFQOH88DMMJJLM ww5EF5E 5EFQOH) 99>>+ &! + G Gs  E EtcURRU5nSnU(a/URS5(a[R"US5nUR U5nUcUcgUcU$UcU$[ R"UR*U-5nXd-SU- U--n[RSUSSU35 U$)uNGet fused user preference vector combining LLM intent and feedback. Formula: u_t = α_t * u_llm + (1-α_t) * u_feedback where α_t = e^(-λt) Args: user_id: User identifier t: Interaction round number (0 for first interaction) Returns: Fused user preference vector, or None if neither intent nor feedback available Nu_llmzFused user vector: alpha_t=z.4fz, t=) r load_intentr"r)rErOmathexprr+r,)rr>rQ intent_datarS u_feedbackalpha_tu_ts rget_user_vector!PreferenceLearner.get_user_vectorskk--g6  ;??733HH[12E11':  =Z/ =   L((D---12oW :: 273-tA3GH rNpatternsinteraction_roundcU(a7URX#5nUb#URX5nURXU5$UR5upgUcUcgUVs/sHoR S/5PM n nU V s/sH o(dM U PM n n U (d$[ R S5 S/[U5-$URRU 5n U c[ RS5 g/n U HnSnUbk[RRU5n[RRU5nUS:a'US:a![R"X5UU-- nUU- nUbk[RRU5n[RRU5nUS:a'US:a![R"X5UU-- nUU-nU RU5 M /nSnU H6n U (aURU U5 US- nM%URS5 M8 URUUU5$s snfs sn f)uLScore patterns based on user preference vectors (fused LLM intent + feedback). If user_id is provided and Stage0 is enabled, uses fused vector: u_t = α_t * u_llm + (1-α_t) * u_feedback Otherwise, falls back to original method: score = similarity_to_positive - similarity_to_negative Args: patterns: List of pattern dictionaries, each containing 'pattern' field user_id: Optional user identifier for Stage0 fusion interaction_round: Current interaction round number (for decay calculation) Returns: List of similarity scores, or None if no user vectors available NrzNo valid patterns to score.z&Could not encode patterns for scoring.rrT)r\_score_with_vector_apply_preference_weightsr7r"r+infor'rr(warningr)rFrGdotr&)rr^r>r_ fused_vectorrawr: negative_vecp pattern_listsp_listvalid_pattern_listsp_vecsscoresrLscore positive_normv_normpositive_similarity negative_normnegative_similarity final_scores valid_idxs rscore_patterns PreferenceLearner.score_patternss/& //KL'--hE55hWMM&*%<%<%>"   L$88@@x!y"-x @5BLM&VvML" KK5 653x=( (../BC > NNC DAE' " | < *A:-!"3*,&&*AVmE[*\'00E' " | < *A:-!"3*,&&*AVmE[*\'00E MM% +0  #F##F9$56Q ##C( $--h gNNgAMsH?= I Iroc^^UR(aU(a[U5[U5:waU$URRU5nU(dU$[ UR S5=(d /5m[ UR S5=(d /5m/n[ U5GHupgUR S/5n[U[5(aFURS5V s/sH)oR5(dMU R5PM+ nn O0U(a)UV s/sHn [U 5R5PM nn [[U5S5n [U4SjU55n [U4SjU55n SX- U - S --n [S [S U 55n URX&U -5 GM U$s sn fs sn f) zApply preference-weighted fusion (adaptive fusion of similarity and feature-level preference). Aligns with 2025 work on adaptive fusion of multi-dimensional user preferences. weight = 1 + (n_liked - n_disliked) / |pattern| * 0.5, clamped to [0.5, 1.5]. rrr,rTc36># UHoT;dM Sv M g7frTN).0f liked_sets r >PreferenceLearner._apply_preference_weights..s=U9n!!U  c36># UHoT;dM Sv M g7fr}r~)rr disliked_sets rrrsC1l1BQQrg?g?g?)rr'rget_preference_featuressetr" enumerater#strsplitstripmaxsumminr&)rr^ror>pfoutirjplistxnn_liked n_dislikedwrrs @@rrc+PreferenceLearner._apply_preference_weightssn ++7c&kSQY]>ZM [[ 0 0 9Mv,"- 266),23 h'DAEE)R(E%%%,1KK,<J,53x=( (IINN;/ AYY^^A&Fzi!mVVA3v7IJ  j) c"   #F##F$56Q ##C( $EAMsE EEllm_vecfb_vecalphacX1-SU- U--$)uFuse LLM intent vector and feedback vector. Formula: u_t = α * u_llm + (1-α) * u_feedback Args: llm_vec: LLM intent vector fb_vec: Feedback-based vector alpha: Fusion weight (0-1) Returns: Fused user preference vector rTr~)rrrrs r fuse_vectorPreferenceLearner.fuse_vectorSs!e)v!555rcURU5nURRU5nSnU(a/URS5(a[R "US5nUbUbUR XSU5$UbU$UbU$g)zUpdate user vector iteratively (for Stage4). Gets latest feedback and fuses with LLM intent vector. Args: user_id: User identifier alpha: Fusion weight for LLM intent vs feedback Returns: Updated fused user preference vector NrS)rOrrUr"r)rEr)rr>rrYrXrSs rupdate_user_vector_iterative.PreferenceLearner.update_user_vector_iterativebs11': kk--g6  ;??733HH[12E  !7##Eu= =  #   Lrc"URXUS9$)aGScore all patterns for evaluation. This is a wrapper around score_patterns for evaluation purposes. Args: patterns: List of pattern dictionaries user_id: Optional user identifier interaction_round: Current interaction round number Returns: List of similarity scores, or None if no user vectors available )r>r_)rx)rr^r>r_s rscore_all_patterns$PreferenceLearner.score_all_patternss""8Pa"bbr)rrrr)g?F)r)Nr)g333333?)__name__ __module__ __qualname____firstlineno____doc__floatboolrrrr)ndarrayr7r<rrOintr\rrrrxrcrbrrr__static_attributes__r~rrr r sb ?U ?D ?>0E(2::*>@T*T$U>0@8BJJ#7 s x 7K D(s(s(8BJJ;O(TW[/0TOtDcN';TOhsmTO),TO5=d5k5JTOl$tCH~2FPTUZP[,4SM>B5k8-4S#X+?-bjj-]abg]h-^ 62:: 6rzz 6% 6TVT^T^ 6CPXY[YcYcPd>[_45c4S#X+?c(SV-c.1c:B4;:Occrr )rnumpyr)rVtypingrrrrrlogging memory.storer learning.embedderr getLoggerrr+r r~rrrs;^ 33$-   8 $AcAcr